# Build Your Own Finance Agent Inside ChatGPT - Stefano Fabbri & Oliver Steffmann, OpenAI

Stefano Fabbri, OpenAI · Oliver Steffmann, OpenAI

[Talk page](https://ai.engineer/talks/ZJjRgvjhq3M-build-your-own-finance-agent-inside-chatgpt) · [Video](https://www.youtube.com/watch?v=ZJjRgvjhq3M)

## From a Portfolio Move to a Client Meeting: Building a Finance Agent Inside ChatGPT

Stefano Fabbri and Oliver Steffmann use a wealth advisor’s meeting preparation to show how ChatGPT retrieves financial evidence, where reusable skills shape the analysis, and when a firm needs to connect its own tools through Codex.

![Three columns pair ChatGPT with “Explore a question,” ChatGPT Work with “Complete a task,” and Codex with “Build the method,” each with an example prompt.](https://ai.engineer/api/talk-essay-media/ad33463ff5b22e361e47c14db98c88686aa788b2c7229fdbaffe16d21c1739d2)

Choose the surface by what the finance team needs to do. Use ChatGPT to explore what Rachel should ask about Jordan’s renovation plans, ChatGPT Work to prepare the review using client records and cited company research, and Codex to create and test a reusable advisor-review skill.

### Start with the advisor’s work and provision a financial workspace

[0:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=12s)

A wealth advisor has to explain why a client’s portfolio changed and what it would take to bring it back into balance. OpenAI’s Stefano Fabbri and Oliver Steffmann chose that assignment for a two-hour workshop with people from asset managers and investment banks. It gave them a way to test several capabilities together. Could an advisor get the relevant financial data, use a model to understand what happened, and carry the evidence through to a presentation they could use with the client?

Fabbri introduced the pair as members of OpenAI’s customer-facing applied AI team. Much of his work was with financial institutions in New York, helping them put models into production and get value from what they spent on them. Steffmann brought experience in asset management and investment banking, along with a developer’s interest in what a firm could build on the platform. Their colleague Dom was in the room to help participants individually. The split in experience informed the agenda. First, use the product as an advisor would. Then, take the internal developer’s seat and decide what to add.

The first action was to scan a QR code and provide an email address. OpenAI used it to grant access to a dedicated workspace with financial data and unlimited tokens for the session. The organizers also enabled an ultrafast mode for GPT-6.1 Sol and GPT-6 Astra. Fabbri put average generation speed in that mode at about 300 tokens per second. The immediate benefit for participants was shorter waits as they tried prompts and generated material. It was a figure for model output speed, separate from the age of the market data they would retrieve.

Fabbri budgeted about ten minutes after the introduction to get everyone into the workspace. Steffmann would lead two practical sessions, beginning in ChatGPT Work and then moving to Codex. The organizers reserved the final twenty minutes for questions about both the engineering and its use in financial institutions. Participants could ask for help as they went, which mattered because getting onto the workshop site and getting access to the financial product involved separate steps.

### Three interfaces and three ingredients of the financial product

[4:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=297s)

Fabbri distinguished three interfaces. ChatGPT was the original general-purpose product. ChatGPT Work was the environment aimed at knowledge workers, with agents and plugins available in the course of their work. Codex was oriented toward engineering tasks. A business user could work in Codex, as Steffmann would later point out, but the sequence here was deliberate. An engineer could first see what an advisor could accomplish in ChatGPT Work before deciding that the firm needed more software.

Fabbri described ChatGPT for financial services as an enterprise product built on the same underlying model as ChatGPT. He distinguished three parts of the offering. The first was data access. Some financial sources were built in, and users with access to other providers could query those sources through ChatGPT. The second was the model’s proficiency at financial tasks. Fabbri said OpenAI had trained the model on work such as mergers and acquisitions and equity analysis. Retrieving a filing supplies the evidence. Interpreting it for a valuation draws on that financial training.

The third part was producing artifacts. An analyst may need a spreadsheet that another person can inspect or a presentation suitable for a meeting. The model could use a historical example as a template, carrying its format forward into a new analysis. That makes the old presentation useful even when the underlying numbers have changed. It supplies an example of how the team communicates its work.

Steffmann identified what to look for when using the finance plugin in the web interface. Beyond fetching a figure, it could guide the model in choosing how to chart the information, validating the data, and adding notes about its origin. Those instructions matter at the point where a result moves from a conversation into a deck. An advisor explaining a chart needs to know what its series represents and where the values came from.

### What the finance workspace supplies before a firm connects its own data

[7:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=472s)

The dedicated workspace gave participants access to a product Fabbri said was available to enterprise customers. He named Daloopa, LSEG, PitchBook, and Crunchbase among the providers OpenAI worked with. The available material spanned SEC filings, market data, news, private-company profiles, and earnings transcripts. These sources answer different parts of an analyst’s question. A market series can establish that a price moved. A filing or an earnings transcript gives the analyst material to investigate an explanation for the move.

![An enabled financial-data menu lists nine categories. A table beside it pairs eight sources with the filings, transcripts, company metrics, news, and market data they supply.](https://ai.engineer/api/talk-essay-media/f6a56ed41c88d281f058ad23bf8654fce13bfc652ad666a0ec8775d935f8d9e9)

A deep dive on financial data

Fabbri estimated that the market data was delayed by about fifteen minutes. For the meeting-preparation case, that qualifies what an advisor can say about a recent observation. Even a model that produces an answer in seconds is working with a price from the time represented by the source. Fabbri gave the delay as an approximation and did not break it out by provider or instrument.

The integration also changed what the model received. Fabbri said the team had repackaged the datasets into forms the model could consume more readily and had seen a significant improvement in how it used them. Steffmann recognized the engineering work from an earlier effort in asset management. Downloading SEC files was one job. Handling the collection at scale and continuing to bring in new information were recurring jobs. He pointed to agents continually collecting data and to a layer of context that made the resulting material usable at speed. Later questions would get much more specific about what belonged in that context.

### Use an investment-banking or equity-analysis plugin as a starting procedure

[10:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=608s)

Fabbri described a finance plugin as a library of skills packaged with Model Context Protocol (MCP) connections. The skills give the model procedures it can reuse. The connections give it tools with which to carry out parts of those procedures. OpenAI had worked with investment banks, asset managers, and equity analysts on the contents of the skills. The goal was to give a practitioner a recognizable starting method for an assignment that would otherwise require considerable instruction.

Consider an equity analyst with a recurring assignment. Fabbri suggested starting with a supplied skill, using it as the basis of a scheduled task, and seeing how closely the resulting process matches what the analyst wants to repeat. At that point, the analyst can identify the missing instructions and adapt the procedure to their own requirements and views. Fabbri used an informal 80/20 split to describe the division of effort. Industry knowledge might get much of the base in place, leaving the practitioner to supply the last part. He offered that as a way to think about customization, without claiming a measured completion rate.

The ready-made examples leaned toward investment banking. They included preparing presentations, analyzing earnings, performing valuations, and building leveraged buyout (LBO) models. The day’s exercise would instead follow wealth management, where the immediate question concerned a particular client’s portfolio. One of the presenters recalled having discussed asset valuation with bankers despite already knowing capital markets. Trying the plugin on a few local examples helped him understand more quickly the problem those customers had been describing. Running a recognizable case gave the practitioner something concrete to compare with their own method.

![Five labeled screenshots show an LBO model with adjustable assumptions, a buyer screening table, a valuation spreadsheet, a pitchbook page, and an earnings forecast chart.](https://ai.engineer/api/talk-essay-media/5d10cff1a5f4ead0be862bb750e903ee816b05c7780faff01d46ad8f3c7aff28)

Use cases your team can get started with

### Follow the change from the client’s portfolio into the meeting deck

[13:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=787s)

As participants confirmed that they could sign in, the presenters directed them to the workshop site. A QR code led to the development material, with an access code available if the site requested one. They allowed time for people to get in before proceeding. Meanwhile, Steffmann explained what the advisor was supposed to accomplish with the financial data and the supplied guidance.

The starting event was a change in a client’s portfolio. The team preparing the meeting had to examine the portfolio, work out what happened, explain it, and prepare material for the advisor. Steffmann specified a profit-and-loss statement or a portfolio attribution as the analytical output. These give the advisor different ways into the conversation. A profit-and-loss statement establishes the gain or loss. Attribution addresses which parts of the portfolio contributed to the result. A market move takes on meaning for this client through what the client held.

The next output was a presentation explaining the change and what the advisor could discuss with the client to restore balance. The analysis and the deck therefore had to agree. If the advisor wanted to discuss a particular holding, the preceding work had to establish why that holding mattered to the portfolio change. The exercise supplied a scenario and prompts, leaving the advisor’s team room to use its own working method. Steffmann had seen considerable variation in how advisors prepared for meetings.

The business objective was more time for the advisor and client to spend together. That put a useful boundary around the engineering exercise. Portfolio research, explanation, and meeting preparation were all candidates for assistance because they consumed time before the conversation. The deck was intended to equip the advisor to conduct that conversation. At this point the presenters had specified the assignment and its intended outputs. They had yet to show a computed attribution, a particular rebalancing recommendation, or a finished client presentation.

### Choose between a separate API application and an extension of the existing workspace

[16:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1003s)

Steffmann asked how many participants were developers, engineers, or otherwise involved in technical development. He estimated about half the room. For them, he recast the case as the next request an internal engineer might receive. Suppose ChatGPT has already landed successfully in the finance department. Business users now want it to work with the firm’s data and to fit the way their own teams operate. What should the engineer actually build?

One option was a conventional application. Obtain an OpenAI API key, build a web interface, use Codex to help implement it, and deploy it within the company. The result would have the familiar software development lifecycle of a company application. The other option started with an observation about the users. Many already had enterprise ChatGPT Work, Codex, or both. The engineer could put the new workflow inside an interface those people were already using. For a team trying to shorten application deployment, the interesting question was how much of its requirement that route could satisfy.

Steffmann separated three kinds of customization. A skill is reusable Markdown that tells the model how to perform a task. It fits a difference in working method, such as the procedure an advisor wants followed during analysis. Connecting the firm’s own data introduces a tool requirement. The engineer must define what the agent can call and configure the MCP connection through which it can make those calls. A plugin packages those capabilities with the relevant skills and can be used on supported web and desktop surfaces.

A plugin extension, recently announced at DevDay, addressed a third requirement. A firm might want part of its workflow to appear differently inside the application. An extension let the developer customize parts of the interface for the use case. Steffmann’s distinction gives the engineer a concrete design question at each step. Does the advisor need better instructions? Does the analysis need an internal tool? Does the person working through it need a particular view? Each answer changes the scope of what has to be built.

### Get into the right workspace and open the sample client package

[20:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1253s)

An attendee asked what accounted for the finance product’s performance with transcripts, reports, and market data. How much came from the tools and skills, and how much from changes to the model? The question remained open while the presenters helped people get access. They supplied the workshop access code for anyone missing the full link and asked participants to confirm that it worked. The instructions were on a public website hosted with the Sites feature. Invitations to the financial workspace were arriving separately by email.

The site included a small incentive to work through the exercises. Participants who got in received 100,000 in virtual funds and could earn more by completing steps. The organizers planned to recognize the top three with bottles. Seeing people appear on the site helped the presenters tell that registration was progressing. For the finance exercise itself, the meaningful confirmation was being in the ChatGPT workspace where the participant could see the financial data and the ChatGPT for financial services plugin.

Once there, participants could follow the basic-capabilities instructions and do the setup once. A ZIP file supplied the example client data for the wealth-management case. Steffmann invited people to choose among portfolio work, research, and graphics. Each section could be tried independently. Someone interested in research could explore the financial evidence, while someone interested in charts could concentrate on how a result would be presented. The common client case kept those choices tied to the same advisor’s assignment.

![The workshop page highlights “Set up once” in the sidebar. The main column shows the model and reasoning settings, how to select ChatGPT for Financial Services, and where to download the starter ZIP.](https://ai.engineer/api/talk-essay-media/da320ff9d3fef3a847c09b17764779119bd06944720a208ecbed67136f299123)

ChatGPT for Financial Services workshop setup

Participants interested chiefly in the financial product could stay in that part of the exercise. Those interested in development could install or open Codex on the desktop and begin its setup. The developer material would cover a skill and plugin, an MCP connection to supplement the workflow, and a plugin extension. Keeping the basic exercise available mattered. An engineer deciding what to add could first try the existing product against the sample client data and see which part of the job already worked.

### Make a desktop view respond to the conversation or a file

[26:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1563s)

Steffmann returned to the structure of a plugin to explain what the extension added. A plugin could contain skills, an MCP server, configuration for the tools, and interface elements for presenting the result. The extension gave the developer control over designated areas of the desktop interface. What appeared there could depend on the conversation a wealth-management user was having with ChatGPT or Codex.

![An OpenAI Developers page shows a ChatGPT interface diagram with purple highlights on a sidebar item, settings, message content, the composer, and a right-hand panel.](https://ai.engineer/api/talk-essay-media/fae41ab66da04db3e5226aad7a74f59ccd0d164859bcaa0f83c6606bb1b84ec7)

Plugin Extensions in the ChatGPT interface

His concrete example was a particular file type. The developer could arrange for an extension-controlled view to appear when that type of file came up in the interaction. The user could expand it and work in an interface tailored to that material. That is a different deliverable from asking the model to produce a chart for a presentation. In one case the analyst needs an artifact to take to a meeting. In the other the analyst needs a way to interact with something while the work is underway. Steffmann described this as a design possibility. He did not identify a specific file format or show a completed client-specific extension in this passage.

### Check the web, desktop, and CLI boundaries before designing a pipeline

[27:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1658s)

An access question exposed a consequential constraint. At the time of the workshop, ChatGPT for financial services and its integrated financial data were accessible in the ChatGPT web interface. A company might provide both web ChatGPT and Codex, and a single employee might use both, but the workflow still had to account for which environment could reach the data. An advisor’s successful query on the web established access on that surface.

Plugins were available in ChatGPT Work and on desktop. The interface customization provided by plugin extensions was desktop-only at that point. Codex also had a command-line interface, which made it a natural candidate for a developer building a pipeline. Steffmann specifically cautioned that the CLI could not then retrieve the finance product’s integrated data. This qualified the earlier example of putting a skill on a schedule. Before reproducing a successful interactive analysis in a pipeline, the engineer needs to identify where it will execute and whether the required source is available there.

For the exercise in ChatGPT Work, participants were advised to select Astra or 6.1 Sol, open the speed setting, and choose ultrafast. They could then follow the supplied prompts for historical market data and research. The prompts gave participants a common starting point from which to inspect the product’s behavior and customize it for an advisor’s use case. By the time the presenters checked again, almost everyone had received the email and entered the intended workspace.

### Give filings and market series enough context to be interpreted correctly

[30:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1847s)

An attendee returned to the unanswered technical question. Beyond the model itself, had OpenAI built a financial ontology or another layer that explained why the product handled these sources better? Fabbri located much of the work in the data ontology, meaning how each source was represented for retrieval and what the model learned about the information it contained. The team changed the representation to make relevant information easier to obtain and supplied context specific to the kind of source being used.

Fabbri used a 10-K or 10-Q filing as the example. Delivering the file gets its contents into reach. Explaining what is in that kind of filing helps the model decide when to use it and what information to seek there. Fabbri said the same principle applied to integrated sources and to plugins from data-provider partners where the user had access. The useful unit was the data together with its context and a format the model could work with.

The plugin’s instructions supplied another kind of knowledge. OpenAI had worked with financial institutions over time and selected partners, including investment banks, to check product quality. Their contribution informed what counted as a realistic expert procedure. Steffmann added the scale consideration from his experience assembling financial data. Building the context layer independently means processing a large, changing collection. He described agents working across public sources and providers so the product could continue finding the right material.

A second question asked whether developers could get at the underlying structures, perhaps as a graph. Fabbri said those internals were unavailable through the product. What he could explain was the work with each supplier to understand the dataset and decide how to present it. His market-data example was the distinction between prices and returns. A price series and a return series can refer to the same security while representing different quantities. Choosing the right security still leaves room for a basic analytical error if the model mistakes one series for the other. The context should help it choose the appropriate information for the intended operation.

When an attendee suggested that special embeddings for 10-Ks and 10-Qs might explain the improvement, Fabbri rejected that explanation. He said the team had preprocessed the dataset so the model could retrieve information from it easily. He suggested thinking of the retrieval step as a tool call specialized for that dataset.

### Run a supplied prompt and inspect which plugin answered it

[35:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2131s)

Steffmann made the retrieval path visible in the product. Participants could copy a prompt from the workshop site and paste it into ChatGPT. In his example, the prompt indicated that ChatGPT for financial services should be used. Once it ran, icons in the interface showed that ChatGPT had picked up the finance plugin. The user could see the financial data being used and inspect the tool calls made within the plugin.

That gave participants two things to distinguish in their own runs. The prompt expressed the analyst’s intent to use the financial capability. The interface then supplied evidence of which capability had actually handled the request. Participants could inspect that behavior even though the provider-specific representations and backend remained internal. The demonstration established plugin selection and tool use. It did not supply a numerical portfolio result to reproduce.

Steffmann ended this portion at the boundary of the firm. The integrated sources could serve analysts across organizations, but an individual firm would also have datasets and tools of its own. The next engineering task would be to package those capabilities in a custom plugin. The wealth-advisor case explains why that step matters. The market source can describe a security’s movement. To connect that movement to a client outside the sample exercise, the analysis also needs the firm’s information about what the client owns.

### Give the finance plugin the context to present an answer

[36:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2212s)

For a wealth advisor using ChatGPT for Work, the finance product already supplies a place to interact with data, perform an analysis, and turn the result into an Excel workbook or a PowerPoint presentation. The engineering opportunity begins with what is specific to the advisor’s firm. It has its own tools, data, and logic. An engineer can make those available through a plugin that the advisor uses in the same workspace as the financial analysis.

Some of the value of the existing finance plugin is easy to miss if you look only at the number it retrieves. Consider a question about income. A table might be the appropriate answer, or the figures might lend themselves to a graph. The plugin can supply context about how to present that kind of data. With a standard model, the person building the experience would have to supply that guidance. The adapted interface is visible to the user even though the full context behind it is internal to the finance product. The speakers attributed that behavior to context supplied to the model without claiming that a separate model had been trained for the display.

![Four blue bars rise from $65.59B in September 2024 to $76.44B in June 2025. Beneath the chart, a table identifies Microsoft’s FY2025 Form 10-K and Q4 earnings commentary.](https://ai.engineer/api/talk-essay-media/74c4d4be1333af90646791f68e07624a628ce4db465c59cbc0cc56d7c48d2f22)

Microsoft FY2025 quarterly revenue

The distinction matters when adding an internal tool. Giving the model access to a figure solves the retrieval problem. Telling it what an advisor normally wants to see when asking for that figure solves a different part of the job. The answer can arrive in a form the advisor can inspect immediately, before anyone asks for a finished workbook or deck.

### Historical prices expose the boundary around the data service

[38:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2334s)

While checking participants’ progress, the presenters mentioned that participants could also use a mobile app. Steffmann then raised a separate question about historical market prices. An advisor could request a price series within the financial service. Could an engineer retrieve the same series through an API for use in another program?

The speakers treated such an API as a desirable possibility. Market data comes with provider licenses, and the way prices can be made available in a service bears on whether they can be distributed through another interface. The ability to retrieve a series in a conversation therefore gives an engineer only part of the information needed to design an automated job. The engineer also needs to know where that job will run and what data connection it can use there.

Steffmann suggested a route for firms that already have market data in their internal systems. Build a plugin around the database the firm can access, expose the retrieval operation where the advisor works, and reuse the visualizations that make sense for those prices. He corrected his terminology as he described it. The connection to the internal data is a plugin. A further customization of how the user sees or interacts with the result can be an extension. Those are separate decisions even when the advisor experiences them as one answer.

### Choose a model for the work that actually takes time

[40:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2435s)

Even small requests in the workshop could take minutes. While one was running, the presenters looked for the lightning control that enabled an ultra-fast mode. They had to select the appropriate model to get to it. Steffmann described what the speed difference meant in daily use. With a slower agent, he would start a task and find several other things to do while it ran. With Ultra, it was easier to stay with the task. He also acknowledged the advantage of being able to use that mode extensively at OpenAI.

Fabbri put the demonstration in the context of customer discussions. Efficiency and cost had been prominent concerns over the previous three months. His first distinction was between unfamiliar work and a recurring assignment. For a new codebase or an open research problem, he favored Astra, the most capable model he discussed. There is value in having a model that can work out how to approach a problem whose solution the engineer cannot yet describe in detail.

A recurring report changes that calculation. Once the team knows what a good result looks like, it can put considerably more of that knowledge in a skill. Fabbri had seen Luna and 6.1 Sol produce high-quality reports and analyses with sufficiently specific instructions. The relevant comparison is how well each model performs the actual assignment with that skill. Running the strongest model on a single advisor’s report may be an easy choice. Running the report for every client makes both the per-run cost and the time it takes consequential.

He described a related division of labor in multi-agent systems. Give the more capable model the orchestrator’s job of deciding how to divide and direct the work. Some subordinate agents can use Terra or Luna. Fabbri said they had seen configurations retain roughly 95 percent of the overall result. He gave no common benchmark, price, or workload behind that figure. His condition for adopting the substitution was to build evaluations for the use case and measure whether the resulting analysis still met expectations.

Steffmann had made a similar choice while building the workshop website. He gave Astra the initial site and the more involved live room view, started the work on a Friday, and came back to it. For subsequent small changes, he used Sol because it was faster and more efficient for those edits. His advice was to revisit capability, effort level, and price as models change. The model that earns its keep on the first implementation may have little advantage on the next afternoon’s revisions.

### Choose the surface, then prototype the connection to holdings

[46:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2784s)

While participants continued the exercises, Steffmann invited observations about how they used these products and where they ran into limits. He then took up the developer’s side of the problem. A knowledge worker sees a finance product that can answer questions and make artifacts. A developer looking at the same product can inspect the structure of a plugin and ask which pieces to supply for another workflow.

The plugin structure brought together three things. Skills describe procedures the model can reuse. An MCP server exposes tools that connect it to something outside the conversation. Interface components determine what can appear as the plugin is used. ChatGPT for financial services was itself the example of a plugin. An engineer could use the same arrangement to bring a different set of procedures, tools, and views together.

Steffmann started his architecture diagram with the surface, meaning the place where the user will do the work. ChatGPT on the web, the desktop environment, and Codex running locally offer different ways to configure and reach tools. Local Codex is particularly convenient for a prototype because the developer can install a local Model Context Protocol (MCP) server and exercise it while building it.

For the wealth-management case, the sequence is concrete. Identify the source the advisor needs, such as an internal securities portfolio database or a customer database. Salesforce was another possible source. Use code and prompts in Codex to build an MCP server that can query that system. Connect locally, run Codex against the server, and try the tool. At that point the developer has a prototype with a path from a request to the firm’s data. Getting the same capability onto an advisor’s machine, under the firm’s controls, is the next piece of work.

### Distribute the plugin and decide whose identity it uses

[49:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2968s)

A developer can install tooling for creating a plugin and configure plugins through ChatGPT. The desktop environment also supports plugin installation. Steffmann’s enterprise example was an asset-management firm where employees receive both Codex and ChatGPT for Work. The firm could define a managed folder structure containing its plugins and distribute it through its existing software deployment mechanism. A working connection on the developer’s laptop would then have a route to the people doing the analysis.

That changes the amount of application infrastructure the team has to own. A standalone internal application requires decisions about login and access control and can take considerable time to get approved. Building in an environment the company already manages gives administrators a place to decide which employees can receive a plugin. Fabbri emphasized how often customers asked for precisely that control. If ChatGPT is where the employees will use the result, that existing administration can shorten the path from prototype to deployment.

There were still two different data paths in the diagram. The integrated financial context supplied with ChatGPT for financial services was available through the web service at the time of the workshop. It could not be called from the CLI. A developer’s plugin could separately connect to a tool that worked with the firm’s databases, using another MCP connection or another authenticated integration. In a design that needs both sources, it matters which part of the analysis will run in the service and which tools will be reachable from the chosen surface.

![A three-tier diagram separates the local Codex project, MCP server and browser prototype from OpenAI-hosted ChatGPT and the firm’s systems. Arrows distinguish the tested local connection, optional integration and existing financial-source path.](https://ai.engineer/api/talk-essay-media/e084637a83c9986e015aea89dc187dbfac91533a6b8b660e1e6cb7b68ed3951c)

What runs where

An audience exchange clarified how much has to go into the package. A plugin can be as small as a collection of skills. If the work requires an internal system, someone has to build the connection and manage authentication. A system-to-system connection authenticates one service to another. A connection acting on behalf of a user has to establish the identity of the person signed in to Codex. That choice determines what the integration needs in order to make a request. The employee entitled to install the plugin and the identity entitled to retrieve a particular record are both relevant.

Steffmann also pointed to the Codex sandbox and the policies that can constrain what plugin code does. For the developer, the attraction was being able to deliver code through an approved corporate mechanism and run it within an environment with controls. The firm still has to choose how to deploy that code and how its connection to the data will authenticate.

### Use a skill for a method and an extension for a client view

[54:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3278s)

Steffmann described writing a skill as putting reusable knowledge into a small number of Markdown files. Describe in natural language what to do, use Codex to help create the files, and share the resulting procedure with someone else doing the same work. Developers were already using skills this way in Codex. He distinguished that authoring and reuse experience from what was then directly available on the website.

A plugin gives the developer a way to deliver a larger unit. It can contain those same skills alongside executable logic, code to obtain authentication credentials, MCP connections, and user-interface components built to the plugin’s interface conventions. The folder described earlier can be packaged into an archive. A team distributing a common research method might need only the instructions. A team distributing a client lookup needs the instructions and the machinery that can actually perform the lookup.

To make the extension more tangible, Steffmann returned to a familiar ChatGPT interface. In the example, the lower left offered a place for a custom icon. Selecting it opened an experience controlled by the plugin. The icon was an entry point: the developer could also customize settings and navigation. That gives a firm somewhere to put a view that its staff would otherwise have to visit another application to see.

His example was an advisor preparing for a client meeting. The conversation already establishes what the advisor is doing. Given that context, the extension could surface a view for meeting preparation and display a particular indicator from the firm’s systems for that client profile. The advisor could inspect it and continue the conversation in the same workspace. Steffmann asked the audience to think about the employee who ordinarily has to visit five sites to get the job done. The reason to build this view is quite specific: the employee needs that internal indicator at this point in the work.

### A faster model still leaves the laptop to run the code

[59:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3540s)

Steffmann switched to his Codex setup and asked who already had it running locally. Quite a few participants did. He pointed out that Codex was accessible through the ChatGPT app interface and started experimenting with creating a plugin by prompt. The work took time even with Ultra. Code still had to be written and run locally, with compilation where needed. The example became a discussion of where the agent’s work executes while the room waited.

![The workshop’s “Build a plugin” page shows a copyable Codex prompt with the reference skill, plugin manifest path, version and required advisor brief format.](https://ai.engineer/api/talk-essay-media/41cdaa86c61487740647e75acc2e7c9513f373a2bda188ac212b850f4ad22e11)

Package Advisor Review

On a desktop, Steffmann could open other chats, work on another project, or start building a second plugin. All those tasks still had a machine on which to execute. His example from financial services was an application-security check against a large repository. A sufficiently large scan could tie up his computer. Increasing the model’s output speed does little for the portion of the job that spends its time running code on the laptop.

For that workload, he wanted a company environment to which he could hand off the calculation and get the result when it finished. The point applies to several concurrent tasks as well as to one large scan. The interface can remain the place where the employee initiates and follows the job while the compute runs remotely. The passage describes that deployment choice and the local bottleneck that motivated it. The plugin experiment itself was still running when the discussion moved on.

### Put examples of finished work beside the instructions

[1:02:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3778s)

Asked about reuse, Fabbri described an internal marketplace where teams share plugins and skills. A shared package can contain files as well as instructions. An artifact template might be a website, a PowerPoint presentation, or an Excel workbook. For research delivered to private wealth-management clients, the team can supply several historical examples and instruct the skill to use them as references. An advisor asking for the next report then has access to the same examples as the colleague who developed the procedure.

Fabbri proposed a simple exercise. Take a spreadsheet, presentation, or PDF whose format matters to the team. Supply a fresh analysis and ask the model to put it into that format. Inspect the result against the original. This tests two requirements at once: whether the model can produce the requested artifact and whether it understands the conventions embodied in a real piece of the firm’s work. For a spreadsheet delivered to an investment-banking audience, the quality of the formatting is part of whether the artifact is usable.

![A chat request for a five-slide Project Atlas overview sits beside the resulting presentation. Its financial profile page charts revenue, operating margin, and free cash flow.](https://ai.engineer/api/talk-essay-media/6645f06cd89fbf6de119285aeeb9bbc033c6e6b38b79724d4acaab4b8ba6dca3)

Turn the analysis into client materials in firm style

Fabbri said the team had seen considerable improvement in spreadsheet creation and formatting over the previous six months. He mentioned internal benchmarks but gave no scores. The test he offered participants was one they could perform with their own material. If the result contains an inaccuracy, identify it precisely and revise the skill so the next run has that instruction. Editing the delivered spreadsheet helps one recipient. Putting the correction into the shared skill gives the rest of the team a chance to benefit from it.

Sharing also raises the question of what happens at the next model update. Within a model family, Fabbri had generally found skills reusable across incremental updates with relatively little refinement. He expected a larger transition, such as moving from a fifth generation to a sixth roughly a year later, to require adjustments. The examples do double duty here. They communicate the desired format to the model and give the team something concrete against which to inspect a new model’s output.

### Keep portfolio analysis small enough for a team to share

[1:06:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3970s)

Steffmann had seen customers create skills so enthusiastically that teammates struggled to find and maintain them. Two front-office colleagues or quants may perform much the same task with a few differences. If each records the whole workflow as a personal skill, the team acquires several descriptions of its common method. Someone trying to improve the method first has to discover which descriptions exist and decide where the change belongs.

His advice to one customer was to remove some skills and define the capabilities the team actually shared. Portfolio analysis and trading analysis were his examples of a useful scope. Each names a recognizable piece of work that can be reused in several workflows. A single skill covering everything an advisor does makes even a small variation harder to maintain. For an application developer who needs control over deployment, Steffmann leaned toward packaging the solution as a plugin. For someone who wants to specify how a well-defined task is performed, the skill is still the unit worth getting right.

Steffmann added a qualification to the discussion of model upgrades. Teams still need to retest the models on the work that matters to their business. He described customers with early access to pre-production models using that access to identify problems before release. His concern was as much organizational as technical. Someone needs to decide what a skill encapsulates, which version colleagues should use, and whether it still produces the expected work after a change.

### Let a disagreement in the research trigger the next agent

[1:08:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4113s)

Fabbri’s team had focused primarily on applications built against the API. In the months he had been at OpenAI, he had seen work that once required a separate codebase and API keys become feasible as a plugin. His example came from an internal case study in advanced financial research. It went beyond the advisor’s recurring report to a question where the investigation itself determines what needs to happen next.

In the original API application, a planner received the financial question and broke it into subquestions. For each one, it created a separate subagent to conduct independent research. A critic then gathered the findings, combined them, and checked whether the conclusions conflicted. This is a more demanding job than concatenating several answers into one report. Each investigation may appear satisfactory on its own while leaving the combined account internally inconsistent.

When the critic found a contradiction, it dispatched a new subagent to investigate that particular discrepancy. A further pass therefore had a defined purpose. The process continued toward a report that met the required quality and addressed all the original questions. Coverage and consistency are separate tests here. Several agents can agree while overlooking one of the planner’s questions, or they can cover every question and still produce incompatible findings.

Fabbri said Dom had recreated the research workflow as a Codex plugin. The analyst could stay in Codex and see which subquestions the agents were investigating as the work progressed. A developer could also build a dedicated view to help the analyst follow those investigations through to the assembled report.

### Rebuild equity research around the tools Codex can already call

[1:12:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4323s)

The author of the equity-research application then explained the move from the implementation side. The original standalone application handled professional stock research through parallel investigations and checks. It needed considerable orchestration because its workflow mixed deterministic operations with decisions made by language models. Some steps had a prescribed next action. At other points, what the system had just learned determined which tool or research task it needed next.

Users were already consolidating this work in Codex or ChatGPT. That made a separate application harder to justify and suggested a specific engineering question: how much of the old orchestration could Codex take on? It already had subagents, could direct a changing set of agent tasks, and could connect plugins and data sources. A research plugin could draw on those capabilities without requiring the application author to build a custom connection for each part.

The author used the term “distributed tool calling” for the way the research could call the tools appropriate to the investigation as it developed. He credited it with faster, more accurate tool use, without offering a timed comparison or accuracy score. The resulting design could still keep parts of the workflow rigid where particular filters were required and leave others flexible enough for the model to choose its next action. At the end of this passage, he was tracing those choices back to the manually configured application, which had used language models at some nodes and traditional deterministic processing at others.

### Let the researcher change the sources and the quality checks

[1:13:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4422s)

A researcher who uses a separate research site has to leave their working environment and accept the choices its builders made. Bringing that process into Codex gives the researcher two concrete things to change. They can specify the data sources they want the research to draw on, and they can specify the checks a result should pass. Someone who knows the subject can adapt the next run while working on the analysis. They can also see enough of the procedure to decide which part needs revising.

How does a process developed this way become something a whole team can run? The presenters distinguish the desktop application from the Codex harness, the engine that runs the agent and its tools. The desktop provides a managed product in which a business technologist, quant or researcher can put together a workflow quickly. Plugins give that workflow a way to connect to systems the organization already uses. A developer can then use the Codex SDK to run the harness programmatically on a server. App Engine is one hosting example the presenters mention.

Those choices let the team defer a consequential interface decision. The researcher can work out which sources and checks produce a useful answer before an engineer builds a dedicated front end. If the team eventually wants a server-side program behind its own interface, the engineer can reuse the work that went into the research process. For a deployment that calls for closer inspection or customization, the application is proprietary, but the underlying runner is open source. An enterprise can examine the code that executes the workflow and choose runner settings appropriate to its own interface.

### How the harness handles delegation and a long research context

[1:17:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4631s)

Fabbri gave a reason to use the Codex harness even when a developer could assemble the tools independently. OpenAI uses its own tools during model post-training, and he attributed some of the models’ performance in Codex to familiarity with that environment. Delegation was one example. The agent could decide during a task whether to start a subagent. Fabbri said the Ultra setting encouraged it to delegate more often, which mattered for research with several independent lines of inquiry. The earlier ultrafast control governed how quickly the model generated tokens.

A second mechanism deals with the material those investigations accumulate. A secondary model compacts the current conversation in the background, removing redundancy to keep the working context within bounds. This matters in research because an early instruction can still govern the work hours later, after the agent has written code and produced intermediate results. The team says compaction let it undertake research tasks that would otherwise have needed much more context. Compaction is also available through the API. With the Codex harness, the developer gets that behavior as part of the runner.

A presenter who had worked as a quantitative researcher describes a test using a quantitative finance paper on Brownian motion and options pricing. He asked a model he identifies as version 5.5 to reconstruct the codebase from the paper, generate a numerical result and see whether it agreed with the published one. That last instruction matters. Producing plausible simulation code was only one step. The agent also had to run the calculation and carry the paper’s original claim far enough through the process to make a comparison.

About two hours later, he says, the agent had recreated a substantial quantitative codebase, including the stochastic simulation and model calibration. He credits compaction with making a run of that length practical. The evidence he gives is the scope of the reconstruction and the elapsed time. He supplies neither the paper’s numerical value nor the model’s value, so the degree of agreement between them remains open. For someone reproducing the exercise, the comparison with the paper is a distinct deliverable beyond the working implementation.

### A plugin extension brings a custom interface into the desktop

[1:20:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4843s)

While the discussion continues, a presenter has Codex working in parallel on a plugin extension. He points to the right side of the interface to show where a custom view can connect to the existing product. The example is deliberately small. What the audience can observe is the place an extension occupies in the workspace and that a prompt has set its construction in motion. The presenter refers to Figma and Adobe as partners that have built more complex applications. His broader prediction is that software providers will increasingly distribute applications through AI platforms, and enterprises will develop their own internal equivalents.

![The filter lists Jordan, Alex and Priya. Jordan is selected, slow responses are enabled, and the expanded local JSON preview lists six unresolved items. Codex reports that all 21 tests pass.](https://ai.engineer/api/talk-essay-media/d7f4bf999bb946dada2ddff62db75a177167c7bdd1bc379c054c2ef897d0e360)

Needs client information filter and Jordan’s bounded context preview

Why would a technical user want that interface? One presenter recalls joining OpenAI when many engineers around him still worked in the Codex CLI. He had since seen more of those users spend their time in the native application. His own example comes from quantitative experiments. He used to keep a Jupyter notebook open alongside his other tools. He can now do that notebook work within Codex. The desktop has value for him before he distributes anything to a colleague, because the place where he conducts an experiment can also hold its working materials.

The prospective audience for a plugin is wider. A developer can give a colleague who understands finance, but has little reason to use a CLI, access to a procedure and the tools it needs. Enterprise distribution controls provide a route to those users. The presenters suggest that teams can deliver useful tools in weeks rather than months. Finance makes an attractive test because an analysis often combines changing data sources, technical work and a finished artifact. They invoke claims of tenfold or hundredfold productivity gains in engineering as an ambition for knowledge work. They offer no corresponding measurement for the finance workflow.

### What carries over to cybersecurity, and what structured data leaves unsolved

[1:24:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5063s)

An audience member asks whether finance is especially amenable to this approach because its data is structured. Are the presenters building an ontology over that data, and is there a parallel effort in healthcare? The first response identifies healthcare as a priority, including for research, then develops a more concrete comparison in cybersecurity. There, the team wants a security specialist to be able to start using a model on a familiar problem without first having to become an AI specialist.

The example combines a specialized cybersecurity offering called Daybreak, tools the practitioner can use, and a plugin for static asset scanning tuned to the model. The team also has a library of skills covering stages from scanning through final review and correction. Each piece supplies something different. The model works on the problem, the tools let it perform operations, and the plugin and skills give the security team an initial procedure. The team can start from a scan it knows how to assess and follow the findings through review. Healthcare is named as a field of interest, but the presenters give this implementation detail for cybersecurity.

The other response returns to the audience member’s point about an ontology. Structured data is attractive, but an enterprise may have many structured stores that evolved in isolation. Each holds usable records, yet the context needed to use them together can be missing. One strategy is to design a context layer across every store before making the data available to an agent. The presenter’s objection is the time that takes in a company with a substantial legacy estate.

He proposes starting with teams and the work they already do. Give a team a working environment in ChatGPT and Codex, connect the sources its workflow requires, and build up the necessary context as those connections are made. Other teams and sources can follow as the organization modernizes its data architecture. He draws a distinction from his experience inside OpenAI, where he says much of that context already exists and the company has less legacy infrastructure. An established firm has to budget for discovering how its stores fit together, even when each one is well structured.

### Set the extension’s filesystem, network and approval boundaries

[1:28:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5313s)

Installing an extension can cause code to execute on someone else’s desktop. That changes the deployment question considerably. A developer may know exactly why a prototype calls a particular service from their own machine. Once colleagues install it, the security team needs to know which destinations it can contact, what it can read locally and who can use it. A tool call that needs one URL gives the presenters a concrete example. The organization can restrict where that call goes while still allowing the operation that made the extension useful.

They describe several layers of control. Corporate network security and endpoint inspection still apply. Inside Codex, an enterprise administrator can deploy configuration policies in TOML files. Those policies can constrain actions to a workspace, determine what the client can access locally, restrict networks or individual URLs, and control which people can use particular plugins. The administrator can set different policies for different teams and distribute the settings during rollout. Code executing inside the client is then subject to the configured sandbox.

Consider a team developing a plugin in its project workspace. It needs room to run tools and operate on the project’s files. The same workstation may contain sensitive folders belonging to another project. A workspace boundary lets the administrator express which files are in scope. A network rule addresses a different question, such as whether that plugin may contact a particular site. Plugin policy addresses a third, namely who may use the integration at all. The presenters also put a limit on what those controls accomplish. Whoever operates the underlying tool still has to protect it.

Approval prompts introduce a tradeoff the end user feels directly. A security team wants scrutiny for sensitive operations. A business user who has to click through a prompt for each ordinary visit to an expected site experiences the tool as slow. The presenters describe configuring which actions can proceed and which require a decision, taking account of the data, destination and working folder. Builders need to explain that behavior to the people receiving a plugin. In the large institutions the team works with, approval and permission settings are usually among the first deployment discussions.

### From a workshop extension to an internal or public plugin

[1:32:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5555s)

The presenters pause for a workshop interlude and invite three participants to introduce themselves, with OpenAI water bottles for the recipients. Cohen is a data engineer at the Dutch branch of Xebia Data, a consultancy. He had attended AI Engineer in London and wanted to see what people in New York were doing in finance. Matt describes building an AI-oriented family office at X1 Wealth and using Codex extensively. Zach, founder of The Clockworks, describes making applications that use AI to investigate things he previously could not, with much of his attention on development.

The suggested homework is to build a plugin extension and distribute it. The presenters distinguish two audiences for that exercise. A public marketplace offers a software provider a channel to external customers. Applying takes some work because the extension can affect people outside its maker’s organization. The discussion mentions checking that it works reliably and safely and belongs in the right category. A company can also distribute an extension internally using the same general idea. A developer whose intended users are colleagues can serve that audience through the corporate system.

The analogy one presenter reaches for is the App Store. An emerging marketplace might change how a software provider reaches customers, much as app stores did. He presents that as a reason for developers to pay attention to distribution. The exercise itself is smaller and more testable. The builder has to turn something that ran in their workspace into something another person can receive and use.

### Use evaluations to find where a production agent fails

[1:36:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5804s)

The applied AI team also works with customers on custom API applications, where an engineer takes on more of the system design. A recurring part of that work is developing an evaluation for a particular problem. A person can read an answer in an individual chat and decide whether it helped. A team deploying an agent repeatedly needs a way to establish how well the product performs across cases. Without that, a claim that the agent increases productivity gives the team little help deciding whether it is ready for production.

The team starts from an evaluation dataset and investigates cases where performance falls short. Several explanations are possible. The prompt may communicate the task poorly. The system may communicate poorly with a tool. The model may also reach a genuine limit and behave unexpectedly even under favorable conditions. Those explanations lead to different work. Engineers can revise instructions or the tool interaction and rerun the case. A reproducible failure that persists in a well-specified environment gives the applied team something concrete to take to model researchers for a future iteration.

An audience member asks how mature companies are at this, from understanding evaluation fundamentals to evaluating automated research. The former quantitative researcher answering considers evaluation one of the largest obstacles, chiefly because two kinds of expertise have to meet. Subject-matter experts know what counts as a consequential mistake in the business. Engineers build the dataset and the tests. Getting both groups to define and implement a useful check can be harder than getting the agent to produce an initial answer.

He gives 70 percent as an example of a score that, on its own, offers little direction. A team needs to inspect the underlying metrics to see which behaviors are strong and which are failing. Applied to his earlier paper-reproduction example, reconstructing the simulation, running the calibration and matching the published quantity are distinguishable outcomes. An implementation could accomplish the first two and miss the third. Keeping those results visible would tell the investigator more than a single rating for the finished response.

### Start with simple checks and give a model judge a binary question

[1:40:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6037s)

The evaluation advice has an order. Start with simple checks whose behavior the team understands. Add complexity once it can interpret the results. Beginning with an LLM as judge introduces two uncertain systems at once. When a run gets a poor score, the engineer has to work out whether the agent performed badly or the judge failed to recognize a good result. A simple initial test gives the team a reference point before it makes the evaluation itself depend on model judgment.

When a model judge is warranted, the presenter recommends decomposing a one-to-five rating into individual true-or-false questions. Each question should ask for a yes or a no. In the paper example, an evaluator could ask whether the result includes a numerical comparison with the paper. That is a narrower judgment than rating the entire reconstruction, and a reviewer can inspect the answer that led to it. Separate questions can capture other requirements without asking the judge to decide how much each one contributes to an overall grade.

The presenter mentions the Decision API as one possible implementation of this pattern. He describes it as returning a fast binary decision together with a probability or indication of confidence. The attraction for evaluation is the fit between the interface and the question. An engineer can ask about a specific condition and retain the decision for that condition. Choosing conditions that actually matter to the business still takes the subject-matter work he has just described.

### Give automated research a target it can iterate against

[1:41:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6105s)

The second half of the audience question concerns automated research. The presenter connects recent gains in mathematics with earlier gains in programming. Quantitative work often needs both. Reproducing a paper requires understanding the mathematical procedure well enough to implement it, then using the implementation to test a scientific claim. He suggests reproduction as an entry point because papers offer tasks at different levels of difficulty and, in cases such as the pricing example, an external result against which to compare the output.

For an iterative problem, he describes Codex goal mode. The user specifies an objective, preferably a numerical one when the problem admits it. His example uses an existing agent, an evaluation dataset and a benchmark. Ask Codex to adjust the agent’s prompt to maximize measured performance. Codex can consider a change, try it, evaluate the result against the target, and keep exploring new ideas and hypotheses. This is what he means by hill climbing. The benchmark gives successive attempts a common measure and supplies a reason to keep or reconsider a change.

That example also explains why the preceding evaluation discussion matters. If a benchmark compresses several requirements into a score the team barely understands, the system can pursue that number while the team learns little about which capability improved. With individual checks available, the researcher has a way to examine the effect of a prompt change. The presenter calls prompt calibration a simple instance of automated research. The same general instruction to pursue a goal can support a much less prescribed investigation.

He points to mathematical results for which the team used goal mode and says that, in one case, it published the prompt. In describing the task he corrects himself from a theorem to a conjecture. The prompt is of interest because it shows how the research objective was put to Codex. For a research run in ultra mode, he describes the agent dispatching subagents to investigate independent avenues in parallel, combining their findings and iterating. Goal mode specifies what to pursue. Ultra influences how readily the agent recruits help pursuing it.

A colleague endorses the shift in emphasis. In his experience, six months earlier a builder had to spend more time engineering prompts and organizing the agents’ work by hand. A more capable system can take on some of those choices, which leaves the builder with the demanding job of saying what outcome matters. The finance team invites customers to bring specific cases and evaluation problems back through their account contacts. A goal that succeeds in a demonstration still has to hold up against the work the firm actually needs to do.

### Write back through a tool to the system that owns the record

[1:46:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6401s)

The final audience question picks up a conspicuous feature of the examples. Many skills and plugins read information. Can an agent also write the result back to the system a team works in, such as its accounting system? Dom, the colleague invited to answer, starts by separating orchestration from storage. Codex can coordinate a plugin extension while its tools interact with other systems and store data remotely. The first design decision is where the result needs to exist after the user finishes working on it.

The simplest destination is the project’s local filesystem. A tool can return output in a specified form, and skill instructions can tell Codex how to process it and where in the project to write it. During extended work, the filesystem gives the agent a place to retain intermediate material and return to it. This complements conversation compaction. The harness can manage the size of the conversation while a file holds an output the project still needs. For an individual working locally, that file may itself be the desired result.

For a workflow shared with another application, an extension can use a tool that contacts a central server or connects to a particular database. Dom describes a proxy-voting implementation concerned with what the voting systems had received and where material had been sent. The user worked in Codex. An application server connected through the plugin extension owned the record. The write operation had concrete contents: the report, its findings, the date and the sources. The tool wrote those back to the database that powered a production application outside Codex.

That placement served two audiences. The operator could conduct the work in Codex, while administrators and people responsible for legal verification could use the firm’s existing system of record. Dom describes connecting particular tools through MCP, the Model Context Protocol, and assigning the permissions appropriate to those connections. A builder can keep a project local or expose a path to the service that needs the result. In the proxy-voting case, saving the finished report only in the operator’s workspace would have left the production application without the record it needed.

A second presenter adds one more output for audit-sensitive work. Alongside the artifact, the team often retains standardized logs that let a technical reviewer trace actions and decisions to the data sources and logic behind them. A report gives the reviewer the findings. The log lets that person examine how the workflow reached a particular finding and which action followed. For the financial processes the team describes, retaining that history is part of the plugin’s work when it creates the artifact.

## Key takeaways

- Try the supplied finance workflow against one client case first. Follow the evidence from portfolio change through P&L or attribution into the explanation the advisor needs for the meeting.
- Match the customization to the missing capability. Use a skill for a repeatable procedure, an MCP-backed plugin for access to firm tools or data, and a desktop plugin extension for an interactive view.
- Carry the meaning of a financial value with the value itself. A price, a return, and a delayed observation support different claims even when they refer to the same security.
- Verify both the tool call and the execution environment. The workshop’s web interface exposed which finance plugin ran, while the integrated financial data was unavailable to a Codex CLI pipeline.
- Prototype an internal portfolio or customer lookup in local Codex, then verify which data sources the advisor’s actual ChatGPT or Codex surface can reach and whose identity the tool will use.
- Package a recurring method such as portfolio analysis with examples of the finished workbook or presentation. Use discrepancies in real outputs to refine the shared skill and retest it when changing models.
- Measure delegated tasks separately from planning. A cheaper model may handle a well-specified report, while a faster model does little for a repository scan that consumes the laptop’s compute.
- In a research workflow, give the critic two jobs: check coverage of the original questions and dispatch a targeted investigation when independent findings conflict.
- Prototype the research procedure where the researcher can change its sources and checks. Use the Codex SDK when the same work needs to run on a server or behind a dedicated interface.
- Before distributing a desktop extension, specify the project files, network destinations and tools it needs. Use enterprise policies to set those boundaries and decide which actions require approval.

## Slides

### A Quick Introduction to our surfaces

[4:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=297s) [Three columns pair ChatGPT with “Explore a question,” ChatGPT Work with “Complete a task,” and Codex with “Build the method,” each with an example prompt.](https://assets.ai.engineer/sites/aiecode2025/sha256/ad33463ff5b22e361e47c14db98c88686aa788b2c7229fdbaffe16d21c1739d2.webp)

A Quick Introduction to our surfaces

ChatGPT
Explore a question
“What should Rachel ask about Jordan’s renovation plans?”

ChatGPT Work
Complete a task
“Prepare the review using client records and cited company research.”

Codex
Build the method
“Create and test a reusable advisor-review skill.”

OpenAI
4

### A deep dive on financial data

[8:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=496s) [An enabled financial-data menu lists nine categories. A table beside it pairs eight sources with the filings, transcripts, company metrics, news, and market data they supply.](https://assets.ai.engineer/sites/aiecode2025/sha256/f6a56ed41c88d281f058ad23bf8654fce13bfc652ad666a0ec8775d935f8d9e9.webp)

A deep dive on financial data

GPT-6 Astra Extra High
Financial data
Open desktop app
Built-in financial data
Company Filings
Financials & KPIs
Earnings Calls & Events
Investor Presentations
Market Prices & Trading
Analyst Estimates & Ratings
Company News
Ownership & Holdings
Fund Letters

Source	Available data
Public filings	SEC and other public-company filings
Quartr	Earnings calls, transcripts, slides and disclosures
Daloopa	Statements; operating, segment and geographic KPIs
Fiscal.ai	Financials, ratios, segments, KPIs and ownership
PitchBook Essentials	Private-company profiles and recent financing
Crunchbase	Profiles, funding, investors and acquisitions
LSEG News / Reuters	Reuters news and selected historical archives
FMP / Nasdaq	US equity prices and market data

OpenAI
7

### Put GPT-6 Astra to work on financial research & analysis

[10:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=604s) [An analyst’s revised revenue and earnings estimates appear beside a line chart. At Q2 2027E, the highlighted upside revenue forecast is $1.6B, compared with the prior $1.4B.](https://assets.ai.engineer/sites/aiecode2025/sha256/3e19de9070bb049e1c7b43b1bd18d22f9430a24d8eac80e61bd2d1ad612c5e06.webp)

Put GPT-6 Astra to work on financial research & analysis

→ Research, model, chart, analyze, and check its work like a banker

Compare Avenquill estimates
Adjusted diluted EPS	−$0.60	−$0.60	$0.62	−$0.02

The earnings constraint includes tax: management now assumes 22%, approximately seven percentage points above its original assumption. At the original tax rate, EPS guidance would have been $0.65.

Our proposed revisions are:
Period	Revenue: prior → revised	Revenue revision	EPS: prior → revised	EPS revision
FY2026	$4,700m → $4,800m	+$100m / +2.1%	$0.62 → $0.60	−$0.02
FY2027	$5,750m → $6,000m	+$250m / +4.3%	$0.86 → $0.83	−$0.03

The accompanying chart compares Avenquill’s prior and revised revenue forecasts.

Financial Charts
Avenquill: prior vs upside-case revenue forecasts
Sources 6
Line chart
Date range: Q3 ’26 – Q4 ’27
Adjusted net revenue (USD)
$2B
$1.5B
$1B
$500M
$0.00
Q3 2026E	Q4 2026E	Q1 2027E	Q2 2027E	Q3 2027E

Q2 2027E
Upside case	$1.6B
Prior	$1.4B
Source: R&D Bank model

Data
Source	Quarter	Forecast	Adjusted net revenue
R&D Bank Model	Q3 2026E	Prior — R&D Bank	$1.2B
R&D Bank Model	Q3 2026E	Upside case — R&D Bank	$1.3B
R&D Bank Model	Q4 2026E	Prior — R&D Bank	$1.3B

OpenAI
8

### Turn the analysis into client materials in firm style

[12:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=722s) [A chat request for a five-slide Project Atlas overview sits beside the resulting presentation. Its financial profile page charts revenue, operating margin, and free cash flow.](https://assets.ai.engineer/sites/aiecode2025/sha256/6645f06cd89fbf6de119285aeeb9bbc033c6e6b38b79724d4acaab4b8ba6dca3.webp)

Turn the analysis into client materials in firm style

→ Centrally managed firm templates let your teams build client materials in your firm’s style
→ Format with sources, citations, and assumptions included

Create a five-slide company overview for Project Atlas.

Cover the business, industry landscape, competitive positioning against five peers, financial performance, and growth opportunities and risks. Make the slides rich with analysis, charts and a competitor comparison table, with a clear takeaway on each page and supporting sources

Worked for 9m 35s
Created the five-slide overview with editable charts, a five-peer comparison table and supporting sources.
Project Atlas - Company Overview.pptx
Slides · PPTX
Open in

Project Atlas financial profile
Margin expansion and cash generation now matter more than acquisition-driven revenue growth
Reported earnings growth is converting to cash
Revenue ($m)
Operating margin (%)
Free cash flow ($m)
Growth composition
Margin development
Cash generation and earnings quality
Balance sheet capacity

Latest LTM revenue	Operating margin	FCF / adjusted EBITDA
$702.1m	14.8%	73.1%
R&D Bank

OpenAI
9

### Use cases your team can get started with

[12:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=725s) [Five labeled screenshots show an LBO model with adjustable assumptions, a buyer screening table, a valuation spreadsheet, a pitchbook page, and an earnings forecast chart.](https://assets.ai.engineer/sites/aiecode2025/sha256/5d10cff1a5f4ead0be862bb750e903ee816b05c7780faff01d46ad8f3c7aff28.webp)

Use cases your team can get started with

LBO modeling
Project Atlas
Transaction assumptions
Operating assumptions

Buyer screening

Valuation analysis

Pitchbook preparation
Project Atlas financial profile

Earnings analysis

OpenAI
10

### ChatGPT for Financial Services workshop setup

[24:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1460s) [The workshop page highlights “Set up once” in the sidebar. The main column shows the model and reasoning settings, how to select ChatGPT for Financial Services, and where to download the starter ZIP.](https://assets.ai.engineer/sites/aiecode2025/sha256/da320ff9d3fef3a847c09b17764779119bd06944720a208ecbed67136f299123.webp)

OpenAI | ChatGPT for Financial Services
FINANCIAL SERVICES / WORKSHOP
Get the lab files
Build together
Instructions | Room progress
Learn at your own pace.

ChatGPT for Financial Services
Overview
Join the workshop
Room progress
1 · Core capabilities
Set up once
Portfolio
Firm guidance
Research & charts
Recommended
Advisor brief
Client deck
Optional labs
2 · Extend with Codex
Set up once
Build a skill
Build a plugin
Run a local MCP
Recommended
Build a plugin extension

Use the ChatGPT workspace named by your host. The website may be hosted in a different workspace; its address does not determine where you should run the lab.

Select your workshop workspace, then start a Work chat.

Workshop default: GPT-6.1 Sol · Medium
Open the model picker in the message composer, choose GPT-6.1 Sol, and set reasoning to Medium. Use these settings for ChatGPT Work chats, including new test chats. For Codex labs, follow the model guidance on the lab’s page.

Type @ and select ChatGPT for Financial Services from the picker. Pasting its name as plain text does not select it.

Financial Services and financial data are preconnected here. No provider account, API key or connector setup is needed for the core exercises.

2. Attach the single starter ZIP.
Download this ZIP and attach it directly to your Work chat. Keep it zipped. This is the only client-data attachment you need for the core workflow.

Complete workshop starter kit    Download ZIP
The archive contains the client inputs, guides and examples. The prompt below reads only 01-inputs. The optional standalone skill reference has its own small ZIP. Part 2 uses the same starter kit in independent Codex labs.
See what is in the kit.

3. Check your files.

On this page
1. Set the model and select Financial Services.
2. Attach the single starter ZIP.
3. Check your files.
4. Choose your first lab.

### Financial research & charts

[24:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1485s) [The recommended finance lab opens with instructions for running it independently. A six-step outline on the right follows the research from retrieving evidence and building a chart to carrying reviewed evidence into a brief.](https://assets.ai.engineer/sites/aiecode2025/sha256/defacda1d826736a74bda1ab111070e2828023195712287cdc8ef04b84f7938c.webp)

OpenAI | ChatGPT for Financial Services
FINANCIAL SERVICES / WORKSHOP
Get the lab files
Build together
Instructions | Room progress
Learn at your own pace.

ChatGPT for Financial Services
Overview
Join the workshop
Room progress
1 · Core capabilities
Set up once
Portfolio
Firm guidance
Research & charts
Recommended
Advisor brief
Client deck
Optional labs
2 · Extend with Codex
Set up once
Build a skill
Build a plugin
Run a local MCP
Recommended
Build a plugin extension
Recommended

01 · FINANCE LAB
Financial research & charts
30–40 minutes · Build evidence, inspect historical prices, deepen the filing and earnings review, and explore chart controls.
Recommended lab
← 1 · Core capabilities

Run this lab independently
Complete Set up once, then start here. Use the Work chat containing your starter ZIP, or attach the same ZIP in a new chat. Select @ ChatGPT for Financial Services.
No client files or earlier lab outputs are required for this company-research lab. Each exercise below retrieves missing evidence and reuses verified evidence when available.
Output galleries show recorded results from the 9 October test. Read each caption for gaps and unverified steps. A standalone run may retrieve evidence at a different step and produce a different layout.

1. Retrieve evidence and build a chart.
This is the clearest Financial Services-specific step: retrieve filings, earnings material and structured financials through included data sources, then inspect citations and chart data. Ordinary ChatGPT can analyze files you provide; this lab exercises the plan’s built-in financial-data access. Included sources and access.

Use only pre-meeting client inputs and verified company evidence at this stage. Do not read the later meeting transcript,

On this page
Start this lab
1. Retrieve evidence and build a chart.
Inspect the first source and chart.
2. Inspect historical market prices.
3. Read the filing behind the takeaway.
4. Compare management commentary with the filing.
5. Inspect the chart and its underlying values.
6. Carry reviewed evidence into the brief.

### Prepare the advisor brief

[25:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1502s) [The advisor-brief lab lists three input files for Rachel’s brief. Below the file list, the page explains how to combine client-file analysis with company evidence and retain the original financial citations.](https://assets.ai.engineer/sites/aiecode2025/sha256/33b989c2779e5d5f0e6c1a4fb48e9b603c4ed975dc9bd4959a5d009b5ba76e2a.webp)

OpenAI | ChatGPT for Financial Services
FINANCIAL SERVICES / WORKSHOP
Get the lab files
Build together
Instructions | Room progress
Learn at your own pace.

ChatGPT for Financial Services
Overview
Join the workshop
Room progress
1 · Core capabilities
Set up once
Portfolio
Firm guidance
Research & charts
Recommended
Advisor brief
Client deck
Optional labs
2 · Extend with Codex
Set up once
Build a skill
Build a plugin
Run a local MCP
Recommended
Build a plugin extension

01 · CLIENT REVIEW JOURNEY
Prepare the advisor brief
5–7 minutes · Bring verified facts, research and open questions together.
← 1 · Core capabilities

Run this lab independently
Complete Set up once, then start here. Use the Work chat containing your starter ZIP, or attach the same ZIP in a new chat. Select @ ChatGPT for Financial Services.
Inputs from 01-inputs: client-profile.md, holdings.csv, workshop-policy.md. The prompt reads these files directly; earlier lab outputs are not required.
The page estimate assumes the files are readable. Allow extra retrieval time when starting without company evidence.
Output galleries show recorded results from the 9 October test. Read each caption for gaps and unverified steps. A standalone run may retrieve evidence at a different step and produce a different layout.

Prepare Rachel’s brief.
The brief combines client-file analysis with company evidence retrieved through Financial Services. Writing the brief is a general ChatGPT Work task; the added value here is carrying financial evidence into a concise advisor view. The bracketed source IDs are this workshop’s output format, alongside the original financial citations.

Use only pre-meeting client inputs and verified company evidence at this stage. Do not read the later meeting transcript, meeting notes or action register until meeting follow-through.

On this page
Start this lab
Prepare Rachel’s brief.
Check before turning it into slides.

### Plugin Extensions in the ChatGPT interface

[26:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1600s) [An OpenAI Developers page shows a ChatGPT interface diagram with purple highlights on a sidebar item, settings, message content, the composer, and a right-hand panel.](https://assets.ai.engineer/sites/aiecode2025/sha256/fae41ab66da04db3e5226aad7a74f59ccd0d164859bcaa0f83c6606bb1b84ec7.webp)

developers.openai.com/plugins/build/extensions
OpenAI Developers
Home    API    ChatGPT    Resources
Start searching    API dashboard
Overview    Plugins    Sign in with ChatGPT    Commerce    Ads    Workspace Agents    User Docs

Home
Quickstart
Core concepts
Plugin architecture
Skills
MCP server
Plan
Brainstorm use cases
Define tools
Build
Build an MCP server
Add UI to your MCP server (optional)
Add events to your MCP server (optional)
Extensions
Authenticate users
Build skills
Package your plugin
Examples
Test and publish
Connect and test your plugin
Submit and publish

Plugin Extensions
Build plugins that feel like native, first-class features of ChatGPT.

OpenAI MCP Extensions enables developers to hook their plugins into key surfaces of the ChatGPT user experience, including the sidebar, composer, and file viewers.

Plugin extensions on the web are coming soon to ChatGPT Free and Go users. Existing plugin functionality is unaffected.
Composer mentions are available only in the ChatGPT desktop app.

Message...

Try it out
Get a feel for extensions with Bits & Bolts, our kitchen sink plugin. We’ll keep it up to date as new extensions roll out.
Install Bits & Bolts Remote

Try it out
Start building
Extension reference
Extensions in the wild
Copy Page
Ask AI

### Microsoft FY2025 quarterly revenue

[37:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2269s) [Four blue bars rise from $65.59B in September 2024 to $76.44B in June 2025. Beneath the chart, a table identifies Microsoft’s FY2025 Form 10-K and Q4 earnings commentary.](https://assets.ai.engineer/sites/aiecode2025/sha256/74c4d4be1333af90646791f68e07624a628ce4db465c59cbc0cc56d7c48d2f22.webp)

ChatGPT
New chat
Your dot
Projects
No projects
Recents
Review Financial Website

Review Financial Website
Share
Opened Financial Charts

Microsoft FY2025 quarterly revenue
USD billions · Quarter-end dates shown; Fiscal flags all four restated and Q4 calculated

Value
$80.00B
$40.00B
$20.00B
$0.00
Financial Charts

Sep 30, 2024    $65.59B
Dec 31, 2024    $69.63B
Mar 31, 2025    $70.07B
Jun 30, 2025    $76.44B
Date
Source: Fiscal

Retrieved both through Financial Data for Microsoft Corporation (MSFT; CIK 789019):
Document | Reporting period | Publication/filing date
FY2025 Form 10-K | Year ended June 30, 2025 | July 30, 2025
FY2025 Q4 earnings commentary | Q4 ended June 30, 2025 | July 30, 2025

Outputs
Create a file or site
Sources
FinData
Financial Charts
wealth-advisor-site-export.zip

ChatGPT can make mistakes. Workspace data isn’t used to train models.
Work with ChatGPT
GPT-6.1 Sol Medium

### Microsoft FY2025 revenue data and provider flags

[38:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2323s) [A four-row table gives Microsoft’s quarterly revenue in USD millions. Fiscal marks all four quarters as restated and Q4 as calculated. The text below compares those flags with Daloopa’s records.](https://assets.ai.engineer/sites/aiecode2025/sha256/bcb1856721dac9de72eca104333e24c6b718605bad179e8a3fb278ecc999acaf.webp)

ChatGPT
New chat
Your dot
Projects
No projects
Recents
Review Financial Website

Review Financial Website
Share

expansion addresses capacity constraints, but its returns depend on sufficient usage and successful execution.

Reported results: Azure and other cloud services revenue grew 39% in Q4; management said additional datacenter capacity came online, while demand remained above supply.

The interactive Financial Charts bar chart was created. Its Fiscal-backed data retains quarter-end dates, USD currency, provider, source references and flags:

FY2025 quarter | Quarter end / source | Revenue, USD millions | Provider | Restated flag | Calcula… flag
Q1 | 2024-09-30 | 65,585 | Fiscal | True | False
Q2 | 2024-12-31 | 69,632 | Fiscal | True | False
Q3 | 2025-03-31 | 70,066 | Fiscal | True | False
Q4 | 2025-06-30 | 76,441 | Fiscal | True | True

These are current provider records supported by later filings, not a July 2025 information snapshot. Daloopa-backed Financial Data returned identical revenue values but marked them unrestated; its inspected schema lacked an explicit calculated-quarter flag. Fiscal labels Q4 derived_lineitem; I preserved that designation rather than asserting a calculation method.

I opened and read the 10-K risk passage and inspected the chart’s returned data: all four dates, values and flags matched the retrieved rows. No requested document or revenue quarter was missing. Source inspection/highlighting tools and a separate chart-inspection action were unavailable; source opening and inspection of the returned chart data succeeded.

Outputs
Create a file or site
Sources
Web search
Financial Charts
wealth-advisor-site-export.zip

ChatGPT can make mistakes. Workspace data isn’t used to train models.
Work with ChatGPT
GPT-6.1 Sol Medium

### Microsoft quarterly revenue — FY2025

[43:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2637s) [A Financial Charts line chart in ChatGPT shows Microsoft revenue rising from about $65 billion to about $76 billion across four fiscal quarters, with little change between the middle two quarters.](https://assets.ai.engineer/sites/aiecode2025/sha256/230cbc4b0af761615713c959b5b9f6d2771f339de7472120016a154312bbce04.webp)

Review Financial Website
Working for 1m 21s
I’ll reuse the verified revenue data and chart, show the source table, and request a line view in USD billions. I’ll check the returned data and configuration without claiming that proves the visible controls changed.
Used Financial Charts integration
Review chart evidence finished
Opened Financial Charts

Microsoft quarterly revenue — FY2025
Fiscal Q1–Q4 actual-period revenue · All restated; Q4 provider-calculated · Fiscal
Value
$80.00B
$60.00B
$40.00B
$20.00B
$0.00
Sep 30, 2024    Dec 31, 2024    Mar 31, 2025    Jun 30, 2025
Date
Source: Fiscal

Activity
Subagents    1 used
Outputs
Create a file or site
Sources
Web search
Financial Charts
wealth-advisor-site-export.zip

### What runs where

[48:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2895s) [A three-tier diagram separates the local Codex project, MCP server and browser prototype from OpenAI-hosted ChatGPT and the firm’s systems. Arrows distinguish the tested local connection, optional integration and existing financial-source path.](https://assets.ai.engineer/sites/aiecode2025/sha256/e084637a83c9986e015aea89dc187dbfac91533a6b8b660e1e6cb7b68ed3951c.webp)

OpenAI / Codex workshop
What runs where
Codex builds the method. ChatGPT hosts the interaction. MCP supplies scoped client context.

YOUR COMPUTER
Local workshop components
Codex + local project
Skill: reusable advisor method
Plugin: method + metadata/config
stdio
Local stdio MCP
Read-only status + client triage
Node process · fictional fixtures
Browser prototype
Separate HTTP server
127.0.0.1:8788
Not an MCP endpoint

Install reviewed package
Installation is a separate step
Later: Secure MCP Tunnel
Optional Work connection

OPENAI-HOSTED CHATGPT
Separate financial-source path
Conversation + installed plugin
Reusable instructions and tool config
Uses authorized MCP tool results
Native review panel
Host sandbox · MCP backend
Research
Financial Services
Connected financial sources
Research, charts and artifacts

HTTPS · auth + scope
FIRM ENVIRONMENT
Production pattern · deploy and authorize separately
External HTTPS MCP
Firm / provider hosted
Authentication + client scope
scope
Firm internal systems
CRM · holdings · policy
Authorized back-end access
Not preconfigured
No firm endpoint or data included in the starter kit.

Codex local lab · tested
Native / production path
Existing Financial Services path
Native SDK panel and tunnel connections require separate verification. Backend tools run in the local or external MCP.

Solid connections show the local workshop build. Dashed connections show optional Work integration and the production pattern. An external MCP and real internal systems are not provisioned by the starter kit. Open for full size.

Where it runs    What lives there
Your computer    Codex works on the extracted repository. Node runs the local MCP and browser preview using synthetic records.

### Skill, plugin, MCP or extension?

[58:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3529s) [A three-column table compares four components of an advisor review desk. Each row pairs a component with its purpose and a financial-services example.](https://assets.ai.engineer/sites/aiecode2025/sha256/df0e6731cee5623be4a38d7324941eaf5bec62ac2dc7c5d2620554ed6e7aac01.webp)

Skill, plugin, MCP or extension?
Start with the smallest component that solves the problem. A skill can be part of a plugin; these pieces work together.

Choose    When you need    In this use case

Skill
A reusable method and output rules.
Write a sourced advisor brief, preserve gaps and keep clients separate.

Plugin
An installable package of capabilities, which can include skills, MCP connections and UI.
Package Advisor Review so its method has a name, version and installation step.

MCP
Tools that retrieve or compute from a system.
Read scoped client context and triage a company event against dated holdings.

Plugin extension
An interactive interface inside ChatGPT.
Select a household, inspect evidence gaps and bring bounded context into the conversation.

A skill does not create a data connection. A plugin package does not prove installation. An extension’s interface does not replace its backend.
See the skills, plugin packaging and extension documentation.

Build an advisor review desk.

### Package Advisor Review

[59:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3584s) [The workshop’s “Build a plugin” page shows a copyable Codex prompt with the reference skill, plugin manifest path, version and required advisor brief format.](https://assets.ai.engineer/sites/aiecode2025/sha256/41cdaa86c61487740647e75acc2e7c9513f373a2bda188ac212b850f4ad22e11.webp)

Package review, installation and an actual invocation are separate results.

1. Ask Codex to build the package.
Send this in a fresh task opened at the kit folder. Codex writes the files locally; you decide when to install.

Package Advisor Review    Copy prompt
Send in Codex with the extracted kit open.
Use    Codex

Reference and scope
• Create an Advisor Review plugin inside this kit only.
• Use 06-skills/wealth-concierge/SKILL.md as a reference.

Plugin files
• Create plugins/advisor-review/plugin.json with the Agent Plugins 1.0 schema, name advisor-review, version 0.1.0 and a short description.
• Add skills/advisor-review/SKILL.md with valid name and description frontmatter.

Advisor method
• The method should return a brief under 200 words, exactly three Observation / Evidence / Advisor question rows and a source register.

On this page
Start this lab independently.
1. Ask Codex to build the package.
2. Inspect the local files.
3. Install the local test entry yourself.
4. Verify native reuse.

### Build the advisor method

[1:02:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3767s) [The “Build a skill” workshop page shows where to save the advisor-review skill and how its brief must cite sources, flag evidence gaps and stay within the current case.](https://assets.ai.engineer/sites/aiecode2025/sha256/4e9e948ca85966ace5960e0e7ef9630a8262ff68dda6143761f570218e1fcad6.webp)

Repository skill discovery and ChatGPT plugin installation are separate checks.
Repository skill discovery and ChatGPT plugin installation are separate checks.

1. Build the skill.
Build the advisor method    Copy prompt
Send in Codex.
Use    Codex

Reference and destination
• Read 06-skills/wealth-concierge/SKILL.md as a reference.
• Create a separate skill at .agents/skills/advisor-review/SKILL.md.

Skill output and rules
• Return a brief under 200 words, exactly three Observation / Evidence / Advisor question rows, and a source register.
• Put source IDs after each factual or missing-input sentence and in every evidence row; use individual IDs, not ranges.
• Keep dates, fictional fixtures, missing holdings, stale evidence and forecasts explicit.
• Use only the current case’s facts.
• No outreach, trades or invented research.

On this page
Start this lab independently.
1. Build the skill.
2. Review and try it.

### Advisor Review package validation

[1:05:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3924s) [Codex displays its edit summary beside a validation report for advisor-review 0.1.0. The report lists the added files and passing static checks. The summary marks installation and native invocation as pending.](https://assets.ai.engineer/sites/aiecode2025/sha256/7adb986d860e57082849cabb322e224305115cc67d66efd8582ce149a681b708.webp)

Add Advisor Review plugin
Worked for 3m 59s
Download the updated kit.
SKILL.md
marketplace.json
Validation results

Passed: official manifest schema, JSON parsing, skill frontmatter, catalog values and paths, and ZIP integrity. All 134 original kit files remain unchanged. The package contains only the manifest and generic skill, following the official format.

The method requires fewer than 200 words, exactly three rows, inline source IDs, explicit evidence limitations, and case isolation.

Installation and native invocation: pending. Nothing installed or published; no home settings changed or MCP, hook or creator dependency added.

Edited 3 files
+71 -0
marketplace.json +20 -0
plugin.json +6 -0
SKILL.md +45 -0

Advisor Review package validation
Checked: 2026-10-12. Result: PASS for static package checks.

Created inside the workshop kit extracted from wealth-advisor-site-export.zip → wealth-advisor-site-export/dist/downloads/wealth-concierge-workshop-kit.zip. The Desktop export was not modified.

Files added
- plugins/advisor-review/plugin.json
- plugins/advisor-review/skills/advisor-review/SKILL.md
- .agents/plugins/marketplace.json
- 02-guide/advisor-review-validation.md (this record)

The plugin package contains only its manifest and skill. Existing client inputs, case examples, engineering labs and the reference skill remain outside plugins/advisor-review. All 134 original kit files were compared by SHA-256 and remain byte-for-byte unchanged.

Results
Check    Result
JSON syntax and duplicate keys    PASS — both new JSON files parsed with duplicate-key rejection
Agent Plugins 1.0.0 manifest    PASS — validated against the downloaded official JSON Schema using Draft 2020-12
Name and version    PASS — advisor-review, 0.1.0
Catalog    PASS — workshop-local, Workshop Plugins, AVAILABLE, ON_INSTALL, Productivity
Catalog source path    PASS — ./plugins/advisor-review resolves inside the kit root to the actual plugin directory
Skill frontmatter    PASS — valid YAML, no duplicate keys, matching advisor-review name, 250-character

### Advisor Review plugin validation and SKILL.md drafting rules

[1:16:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4565s) [Codex shows the plugin’s validation results beside its skill file. The file specifies a 199-word maximum, inline source IDs, explicit evidence gaps and a final check before returning the draft.](https://assets.ai.engineer/sites/aiecode2025/sha256/1e48f4ef8c09dc013b6d89d508a3738abdf3ac3babe8dc85604e62adbbef780b.webp)

Add Advisor Review plugin

Created inside the extracted workshop kit. Download the updated kit.
Files:
• plugin.json
• SKILL.md
• marketplace.json
• Validation results

Passed: official manifest schema, JSON parsing, skill frontmatter, catalog values and paths, and ZIP integrity. All 134 original kit files remain unchanged. The package contains only the manifest and generic skill, following the official format.

The method requires fewer than 200 words, exactly three rows, inline source IDs, explicit evidence limitations and case isolation.

Installation and native invocation: pending. Nothing installed or published; no home settings changed or MCP, hook or creator dependency added.

Edited 3 files
+71 -0
marketplace.json +20 -0
plugin.json +6 -0
SKILL.md +45 -0

SKILL.md
Metadata
name    advisor-review
description    Prepare a brief with three evidence-linked observations and advisor questions from the current case's supplied records, while identifying missing, fictional and historical evidence. Use for advisor review preparation, not execution of client actions.

Keep the entire response under 200 words (199 maximum), including the prose brief, table text, headings and source register. Count whitespace-separated words in the rendered text; omit Markdown table delimiters from the count. Shorten wording, not evidence or uncertainty.

Every factual, calculated, fictional, historical, forecast or missing-input sentence in the brief and table must end with supporting inline IDs; for example: “Current holdings were not supplied [M1].” Cite factual fragments too. A question with a factual premise also needs its supporting IDs; prefer a neutral question when support is absent. Table-only or register-only citations do not cover uncited prose. Source-register entries carry their own inline ID labels and define those references; any additional factual or gap sentence there must also carry its supporting ID. Every ID must resolve within this response. Do not use an M ID as evidence for a positive factual claim.

Keep fictional status explicit wherever fictional facts are used, and historical dates explicit wherever time affects an observation. A citation alone does not turn fictional, historical or unverified material into current verified evidence.

Check and stop at the draft

Before returning, count words and table rows, audit every sentence and factual fragment for citations, match all IDs to the current register, verify calculations and dates, and check case isolation and uncertainty labels. If evidence is insufficient, replace the unsupported observation with a cited gap and advisor question; never fill the space with invented research.

Drafting is not permission to send, reply, forward, schedule, publish, trade, transfer or update records. This method stops at the reviewable draft. Never claim outreach, research or trades occurred without actual evidence. Do not install dependencies or change host settings. A source package or catalog entry does not establish installation or native invocation; those remain pending until separately authorized and verified.

### Northstar Wealth prototype with a research trigger and household selector

[1:18:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4715s) [The local workshop panel shows a synthetic AI investment research trigger above Rachel’s client list. The filter shows all authorized clients, with Jordan selected. Codex reports the baseline test results alongside the panel.](https://assets.ai.engineer/sites/aiecode2025/sha256/573f824ab2a50e3ef48b5425783467011a33c57119452691022e26410da59cb1.webp)

Before editing, add tests: Jordan, Alex and Priya pass the client-input filter; Sam does not. Exclude the shared company-research gap from that filter. When the selected client disappears, clear their details and bounded context immediately. Test a late response after Jordan → Priya → Jordan switching.

Implement the smallest change, rerun the tests and show the diff. Preserve source labels, authorization boundaries and the human-review requirement.

I’ll read the review-panel guide, inspect the running panel, and add the requested tests before changing the filter or selection handling.
Used the browser, read files, ran commands
Verify format finished

The baseline’s 15 tests pass. No panel was running, so I started the kit’s loopback server and inspected it. Jordan is a confirmed match with missing client inputs; Sam has only the shared research gap. The filter will use those gap types, independently of relevance, and preserve the existing request-sequence guard.
Running node 08-review-panel/server.mjs
5 files changed +127 -9

127.0.0.1:8788
NORTHSTAR WEALTH · WORKSHOP
Make more time for the conversation.
Connect an issuer event, client context and the evidence behind a draft deck.
LOCAL PROTOTYPE

Fictional clients and a synthetic review trigger. This page is not connected to ChatGPT or Financial Services, and it does not install a plugin.

01 / The research trigger
Source verification required
AI investment: prepare relevant client conversations
A training trigger to connect company research with a client review. This is not a news alert, a current market development or an investment recommendation.
Inspect issuer and source details

02 / Client review
Rachel’s synthetic book
Client filter
All authorized clients
Household
Jordan
Jordan    Confirmed relevance
Alex    Needs information
Priya    Needs information

### Needs client information filter and Jordan’s bounded context preview

[1:20:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4804s) [The filter lists Jordan, Alex and Priya. Jordan is selected, slow responses are enabled, and the expanded local JSON preview lists six unresolved items. Codex reports that all 21 tests pass.](https://assets.ai.engineer/sites/aiecode2025/sha256/d7f4bf999bb946dada2ddff62db75a177167c7bdd1bc379c054c2ef897d0e360.webp)

Implement the smallest change, rerun the tests and show the diff. Preserve source labels, authorization boundaries and the human-review requirement.

I’ll read the review-panel guide, inspect the running panel, and add the requested tests before changing the filter or selection handling.
Used the browser, read files, ran commands
Verify format finished

The baseline’s 15 tests pass. No panel was running, so I started the kit’s loopback server and inspected it. Jordan is a confirmed match with missing client inputs; Sam has only the shared research gap. The filter will use those gap types, independently of relevance, and preserve the existing request-sequence guard.

Used the browser, edited files, ran commands
Verify format finished
The new tests failed before implementation, and all 21 now pass. In the running panel, “Needs client information” shows Jordan, Alex and Priya. Filtering Sam out immediately empties his details, claims and bounded context while the replacement loads. Source labels, authorization checks and the human-review requirement are unchanged.
Locating missing screenshot file
5 files changed +127 -9

Client filter
Needs client information
Household
Jordan
Jordan    Confirmed relevance
Alex    Needs information
Priya    Needs information

Teaching ledger
Seeded errors
☑ Simulate slow responses
Switch Jordan → Priya quickly. The last selection must win, and prior household details must clear immediately.

Bounded context preview
Local preview only. Nothing is sent to ChatGPT.
{
  "selected_client_id": "jordan-ellis",
  "event_id": "training-ai-investment-review",
  "selected_claim": null,
  "unresolved": [
    "tax_lots",
    "cost_basis",
    "payment_schedule",
    "other_cash_obligations",
    "jurisdiction",
    "event_source_unverified"
  ],
  "local_only": true
}

### Review-panel filter implementation and test results

[1:21:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4896s) [Codex lists the filter behavior and 21 passing tests beside a JavaScript diff. The diff shows how the panel rebuilds the household selector when the client filter changes.](https://assets.ai.engineer/sites/aiecode2025/sha256/f8df6e653d144b5a86823331f4a8aa4db19e787773d480347c55d0e5bde3e3ea.webp)

Implemented in the running panel.
• Filter uses actual client-input and stale-snapshot gaps: Jordan, Alex and Priya included; Sam excluded.
• Shared company-research gaps stay visible but don’t affect filtering.
• Hidden selections clear details and bounded context immediately.
• Late Jordan → Priya → Jordan responses cannot overwrite the latest selection.

Tests were added first: five expected failures before implementation; 21/21 pass afterward. Browser checks also passed. Source labels, authorization boundaries and human review remain unchanged.
Full diff
5 files changed +127 -9

create-a-new-site-with-sites-3 › outputs › review-panel.diff
--- a/08-review-panel/app.mjs
+++ b/08-review-panel/app.mjs
@@ -115,20 +115,32 @@
}, render);
function loadSelection() {
  const id = byId('household').value;
+ if (!id) { controller.clear(); return; }
  const delay = byId('latency').checked ? id === 'jordan-ellis' ? 1000 : 100 : 0;
  controller.select(id, byId('ledger').value, {delay});
}
+let clients = [];
+function applyClientFilter() {
+  const selected = controller.state.clientId;
+  const visible = controller.filterClients(clients, byId('client-filter').value === 'needs-input');
+  byId('household').replaceChildren();
+  byId('client-groups').replaceChildren();
+  for (const client of visible) {
+    const option = el('option', client.name); option.value = client.client_id; byId('household').append(option);
+    const row = el('div', undefined, 'group-row'); row.append(el('span', client.name), el('span', groupLabels[client.group])); byId('client-groups').append(row);
+  }
+  byId('household').disabled = visible.length === 0;
+  if (visible.some(client => client.client_id === selected)) byId('household').value = selected;
+  else if (visible.length) loadSelection();
+}
byId('household').addEventListener('change', loadSelection);
byId('ledger').addEventListener('change', loadSelection);
+byId('client-filter').addEventListener('change', applyClientFilter);
try {
  const boot = await getJSON('/api/bootstrap');
  renderEvent(boot.event);
  byId('fixture-date').textContent = boot.status.as_of;
- byId('household').replaceChildren();
- for (const client of boot.clients) {

### Workshop room progress for 29 simulated advisor portfolios

[1:37:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5859s) [A room-progress dashboard arranges 29 practice portfolios in three rows. Each circle shows an advisor’s name, portfolio value, percentage change and funding. A connection-timeout banner covers part of the summary strip.](https://assets.ai.engineer/sites/aiecode2025/sha256/a10303c6e99119acdc9255e3520cceb99477b54ab408e292d6ae59a55db63261.webp)

OpenAI    ChatGPT for Financial Services
Get the lab files
Build together
Instructions    Room progress
One room. Every advisor’s progress.

SIMULATED
The advisor floor
Practice portfolios · USD
Funding unlocked
$1.7M
From all completed labs
The connection took too long. Your unsaved entries are still here. Please try again.
$4.6M contributed
+4.15% index change
Gentle upward simulation bias

Top 29 by portfolio value · select for full details
Koen    $273K    +1.1% chg    Funding +$170K
Mathieu    $273K    +1.1% chg    Funding +$170K
Zach    $272.8K    +1.1% chg    Funding +$170K
John    $272.7K    +1.0% chg    Funding +$170K
Pranit    $272.4K    +0.9% chg    Funding +$170K
Sapna    $272.3K    +0.9% chg    Funding +$170K
Cooldeep    $212.5K    +1.2% chg    Funding +$110K
Mark    $211.8K    +0.9% chg    Funding +$110K
Ajay    $211.6K    +0.7% chg    Funding +$110K
Carozzi    $201.9K    +0.9% chg    Funding +$100K

Baba    $191.5K    +0.8% chg    Funding +$90K
Mike    $152.2K    +1.4% chg    Funding +$50K
Durian    $151.8K    +1.2% chg    Funding +$50K
Karthik    $142.9K    +2.1% chg    Funding +$40K
Felix    $142.1K    +1.5% chg    Funding +$40K
Oliver Steffmann    $111.9K    +1.7% chg    Funding +$10K
PA    $111.6K    +1.5% chg    Funding +$10K
Mat    $102.3K    +2.3% chg    Funding +$0
Isaac    $101.8K    +1.8% chg    Funding +$0
AviC    $101.8K    +1.8% chg    Funding +$0

jae lee    $101.8K    +1.8% chg    Funding +$0
Iboelle    $101.8K    +1.8% chg    Funding +$0
Karthik K    $101.7K    +1.7% chg    Funding +$0
A_Amely    $101.7K    +1.7% chg    Funding +$0
Nick    $101.6K    +1.6% chg    Funding +$0
Harsh    $101.6K    +1.6% chg    Funding +$0
Somya Rai (EXL)    $101.6K    +1.6% chg    Funding +$0
cfu    $101.5K    +1.5% chg    Funding +$0
MathieuS    $101.5K    +1.5% chg    Funding +$0

First 3
1 John    2 Cooldeep    3 Mathieu
29 of 29 advisors
Pause motion

## Full transcript

[0:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=12s) So, congratulations to everyone.

[0:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=15s) Glad to see so many people,

[0:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=17s) even despite the holidays, Monday morning,

[0:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=20s) 9:00.

[0:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=21s) I am Stefano Fabri.

[0:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=23s) I'm from OpenAI, the applied engineering team.

[0:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=25s) AI.

[0:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=29s) Oliver is with me, and there's another one in the room.

[0:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=31s) our colleague, Dom.

[0:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=34s) So don't be shy during the session if you don't

[0:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=37s) want to ask questions, do you have any more

[0:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=39s) a third person to whom one can turn.

[0:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=43s) Today's session will be a two-hour session.

[0:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=45s) workshop.

[0:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=46s) It will be interactive.

[0:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=49s) Later you will work in more detail with

[0:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=52s) ChatGPT and Codex.

[0:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=56s) We want to demonstrate the possibilities

[0:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=58s) our product in the field of financial

[1:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=60s) services.

[1:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=65s) How can financial data be combined,

[1:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=67s) LLM AI intelligence and correct results

[1:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=70s) to help your daily business

[1:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=72s) .

[1:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=75s) Today we will focus on a case study with

[1:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=77s) capital management, because we know,

[1:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=79s) that there are many company representatives in the hall

[1:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=82s) asset management and investment

[1:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=84s) banks.

[1:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=85s) We decided this would be an interesting example

[1:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=88s) .

[1:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=90s) To give some context, our team

[1:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=92s) mainly deals with interaction with

[1:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=94s) customers.

[1:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=99s) We work with some of you every day,

[1:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=101s) helping to understand how to implement these

[1:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=104s) models into production in your company and

[1:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=106s) get maximum benefit.

[1:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=112s) So for every dollar you

[1:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=114s) invest in us, our task is—

[1:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=115s) help you make the most of

[1:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=118s) its value.

[2:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=120s) Regardless of whether it is surgery,

[2:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=122s) agents or coding productivity

[2:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=124s) —this is our daily work. Our team

[2:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=127s) mainly serves financial institutions

[2:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=129s) here in New York. This is all about me.

[2:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=132s) Oliver.

[2:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=133s) Yes, I work on the same team. Also

[2:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=136s) I have a long history in the financial sector.

[2:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=138s) services, asset management and

[2:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=140s) investment banking. But I'm a developer.

[2:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=143s) in the soul, so today we want

[2:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=145s) talk about how to build on

[2:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=147s) platform, and show you this case. Well, what?

[2:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=149s) , happy to help. I work with big

[2:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=152s) customers who are likely to be in this

[2:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=154s) room. I already recognized a few people.

[2:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=156s) So if you contact us to

[2:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=159s) to resolve an API or Codex issue, we—

[2:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=161s) exactly the ones you need.

[2:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=166s) Before I continue, can I

[2:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=168s) ask you to scan this QR code and

[2:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=171s) enter your information?

[2:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=174s) This is basically an email address.

[2:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=176s) .

[2:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=178s) The reason is that we want to provide you with

[3:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=180s) access to a dedicated workspace

[3:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=182s) in OpenAI.

[3:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=184s) So, you will get access to financial

[3:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=186s) data and unlimited tokens on

[3:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=189s) this session.

[3:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=190s) In addition, we also activate the ultra-fast

[3:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=193s) regime.

[3:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=194s) This is a special mode for our

[3:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=196s) the latest model.

[3:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=200s) 6.1 Soul and 6 Astra, which reach in

[3:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=202s) average of 300 tokens per second.

[3:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=207s) I think that for a workshop it allows

[3:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=209s) It's easier and faster to create content.

[3:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=213s) I'll show the QR code later, but please

[3:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=216s) , this is very important.

[3:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=218s) Otherwise, you will not be able to access it.

[3:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=221s) Perhaps you can contact our colleague at

[3:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=223s) hall. If you need it now, you can

[3:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=225s) come later. People are sitting there. We need this.

[3:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=228s) required to access

[3:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=230s) ChatGPT website for work. So, this is it.

[3:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=232s) what we need.

[3:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=237s) A quick overview of today's session.

[4:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=244s) This is a tentative schedule, we can adjust it a little

[4:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=247s) change it, but we'll do a quick

[4:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=249s) introduction to ChatGPT for working in finance

[4:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=251s) in the field and in the ecosystem in general.

[4:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=258s) Then we will spend 10 minutes on

[4:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=260s) settings to make sure that all

[4:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=263s) you have access to ChatGPT to work in

[4:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=265s) financial sector, and then we will move on to

[4:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=267s) two main sessions that Oliver will lead

[4:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=270s) .

[4:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=272s) The first one will be dedicated to ChatGPT for work

[4:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=275s) , and then we'll move on to the Codex.

[4:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=280s) Finally, we left 20 minutes for

[4:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=282s) questions and general feedback about what

[4:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=285s) is happening in this area, and any

[4:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=287s) questions you may have,

[4:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=289s) both in terms of programming and

[4:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=291s) industry in general.

[4:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=297s) Now let's move on to a brief introduction about

[5:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=300s) our services.

[5:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=301s) As you probably know, we have three

[5:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=303s) main interfaces, three main

[5:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=306s) the platforms we use.

[5:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=308s) ChatGPT—this is, you know, our original

[5:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=311s) product. We have ChatGPT for work,

[5:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=313s) which is essentially more focused on

[5:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=316s) intellectual workers.

[5:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=320s) It can run agents, work with

[5:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=322s) plugins, but it's not quite Codex, which

[5:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=325s) more intended for engineering

[5:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=327s) tasks.

[5:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=329s) So these are the three platforms that we are currently

[5:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=332s) we offer.

[5:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=333s) As already mentioned, we will focus

[5:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=336s) first on ChatGPT for work, and then

[5:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=338s) on Codex.

[5:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=342s) So, within the framework of work on ChatGPT and

[5:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=344s) ChatGPT for financial services we

[5:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=346s) tried to create a product,

[5:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=348s) focused on the corporate sector, where

[5:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=350s) the model remains the same, but the

[5:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=352s) the product is customized specifically for

[5:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=355s) users in the financial services sector.

[6:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=360s) So, this product consists of three

[6:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=362s) main components. First, it is data

[6:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=365s) .

[6:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=368s) We have some built-in data quite

[6:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=371s) high quality, but you can also

[6:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=373s) access other people's data

[6:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=375s) third-party suppliers directly

[6:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=377s) via ChatGPT.

[6:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=382s) Thus, the agent can interact with

[6:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=384s) data sources directly, without

[6:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=386s) no unnecessary waste of resources.

[6:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=391s) Secondly, we have significantly improved our

[6:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=393s) model and taught it to perform

[6:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=395s) specific financial tasks.

[6:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=399s) For example, these could be mergers and

[6:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=401s) takeovers, stock analysis, etc.

[6:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=405s) The third component is artifacts.

[6:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=410s) In recent months, we have significantly expanded

[6:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=413s) the capabilities of our model to create

[6:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=415s) presentations, spreadsheets, and

[6:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=417s) also using templates based on

[6:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=419s) historical examples to reproduce them

[7:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=422s) in the current analysis.

[7:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=426s) Therefore,

[7:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=427s) Yes, maybe it's worth adding.

[7:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=430s) In the labs, we will first focus on

[7:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=432s) on this plugin—ChatGPT for

[7:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=434s) financial services.

[7:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=437s) It is located in the web interface, in fact

[7:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=440s) on the website.

[7:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=441s) Next we will show all the functions, because

[7:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=443s) ability to use financial data

[7:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=445s) —this is just one aspect.

[7:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=448s) This is one advantage, but there are also

[7:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=450s) certain ways of providing model guidance

[7:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=452s) regarding the presentation of information in the form of

[7:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=455s) charts, data validation, and addition

[7:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=457s) notes about their origin.

[7:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=460s) All of this will be very important to see if

[7:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=462s) you will stay for the lab part. It

[7:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=465s) and there is that specialized industry

[7:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=467s) the product we are demonstrating.

[7:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=472s) So let's start with the embedded data, and that's it.

[7:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=474s) so we want to include you in this

[7:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=477s) version, because ChatGPT is for financial

[7:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=479s) service is available only for corporate

[8:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=481s) customers.

[8:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=484s) These are some of the providers integrated

[8:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=486s) in the product.

[8:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=488s) We cooperate with companies such as

[8:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=491s) Lupa, LSEG, PitchBook, Crunchbase and

[8:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=493s) other suppliers.

[8:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=495s) AND

[8:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=498s) actually, so you have an idea of ​​what

[8:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=500s) which is available immediately after installation: in

[8:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=503s) We have SEC reports and market data. Market

[8:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=506s) the data, of course, has a delay, it seems,

[8:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=508s) about 15 minutes, but you still

[8:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=510s) you get data that is as close as possible to

[8:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=512s) to real time. You have news.

[8:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=516s) And also, you know, profiles of private

[8:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=518s) companies and transcripts of reports on

[8:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=520s) profits.

[8:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=526s) The main difference here is that these

[8:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=528s) data sources are directly integrated

[8:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=530s) in the product; it's not just "raw" data that

[8:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=532s) get inside, we spent

[8:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=534s) special database setup to

[8:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=537s) ensure that models can easily

[8:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=539s) consume.

[9:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=547s) So we saw a significant improvement in that,

[9:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=549s) how the model is able to assimilate this data,

[9:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=551s) because we repackaged them into a format,

[9:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=554s) which is not just a "raw" data set,

[9:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=556s) and is better suited for LLM.

[9:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=561s) Yes. So maybe it's worth adding a little more

[9:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=564s) .

[9:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=564s) I did something similar inside myself.

[9:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=567s) companies, and it's quite difficult:

[9:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=568s) download all SEC files,

[9:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=570s) scale the data and ensure that

[9:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=573s) receive new information.

[9:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=576s) I did this for the asset management industry.

[9:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=578s) , but, in my opinion, the uniqueness of this

[9:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=581s) the solution lies in agents who are constantly

[9:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=583s) collect this data.

[9:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=585s) We have financial data providers,

[9:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=587s) who collaborate with us, but also

[9:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=589s) a contextual level is needed, which

[9:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=591s) efficient enough to handle them, and

[9:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=594s) this is often a technical challenge. This is what

[9:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=596s) we can show you later. That

[9:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=599s) contains this product. There are

[10:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=601s) certain technical solutions to ensure

[10:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=603s) this at high speed.

[10:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=608s) Moving on to how to use

[10:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=610s) this data to create analytics: our

[10:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=613s) the product team created several

[10:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=615s) plugins specifically for financial

[10:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=617s) tasks.

[10:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=620s) In particular, we can talk about

[10:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=622s) plugins for investment banking and

[10:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=625s) stock analysis.

[10:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=628s) Within these plugins—and what we are in

[10:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=630s) We call OpenAI a plugin, it's our own

[10:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=633s) kind of a skills library plus MCP.

[10:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=639s) We cooperate with leading

[10:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=641s) investment banks, companies with

[10:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=643s) asset management and market analysts

[10:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=645s) shares to create the most flexible and

[10:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=648s) realistic skill sets that can be

[10:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=650s) use as a base for your

[10:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=652s) analysis.

[10:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=657s) So what we see not only in the field of

[10:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=659s) financial services, but also in general, when we

[11:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=661s) we provide certain skills, they serve

[11:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=664s) a good base.

[11:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=668s) Let's say you're a stock market analyst,

[11:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=670s) who wants to optimize their work; you

[11:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=672s) you can literally take a ready-made skill "

[11:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=675s) out of the box."

[11:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=678s) There is a lot of work behind this.

[11:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=682s) And then, using this skill, you

[11:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=684s) create a scheduled task, your

[11:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=686s) a kind of conveyor belt that reproduces your goals,

[11:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=689s) using this as a basis.

[11:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=692s) And on top of that, you're trying to improve

[11:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=695s) it, customize it and add everything

[11:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=697s) what is currently missing.

[11:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=700s) We have noticed that such plugins

[11:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=702s) provide a solid base, so you don't

[11:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=704s) need to start from scratch.

[11:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=709s) You start with something built on

[11:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=712s) industry knowledge, where 80% is already ready, and

[11:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=714s) you add your 20% settings

[11:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=717s) according to your needs, views

[11:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=719s) etc.

[12:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=726s) So finally, these are just some examples

[12:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=728s) usage, available "out of the box", although

[12:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=730s) they are more investment-oriented

[12:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=733s) banking.

[12:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=739s) Today we will work more with

[12:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=741s) asset management, but if you take

[12:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=743s) preparation of presentations, analysis of profits

[12:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=746s) , valuation or LBO modeling—all of this

[12:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=748s) available immediately.

[12:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=757s) It's interesting because even I, with experience

[12:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=759s) working in capital markets, previously

[12:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=761s) talked to bankers about the valuation

[12:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=763s) assets, and using this plugin on

[12:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=766s) several local examples helped

[12:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=768s) I quickly get to the heart of the problem,

[12:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=770s) which customers talked about.

[12:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=779s) This was a small introductory presentation about

[13:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=782s) our efforts in financial services.

[13:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=787s) Now we will move on to the practical part.

[13:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=789s) construction.

[13:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=791s) Finally, please add your details

[13:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=794s) here.

[13:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=795s) This is very, very important.

[13:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=798s) I'll stop here for a minute to

[13:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=800s) make sure everyone has time.

[13:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=805s) Perhaps it would be right if we

[13:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=807s) Let's switch, I'll show you a little bit.

[13:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=809s) the practical part, and in the meantime you

[13:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=811s) confirm that you have access to the site and

[13:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=814s) you can log in to ChatGPT.

[13:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=817s) Yes, is anyone having trouble entering?

[13:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=820s) data, such as email addresses

[13:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=822s) mail?

[13:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=824s) I think this is a good sign.

[13:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=827s) So now we can move on to

[13:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=829s) the most interesting.

[13:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=831s) Um, actually, this is the site we're going to

[13:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=833s) use for development. So, be

[13:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=836s) affection

[14:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=841s) Um, feel free to scan, and if it

[14:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=844s) will ask for the code, it is there, but theoretically

[14:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=846s) this must be the correct QR.

[14:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=850s) Yes, maybe we'll wait a little before

[14:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=852s) than to move on, because today we have a lot

[14:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=855s) time.

[14:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=856s) Um, maybe just make sure

[14:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=858s) everyone has access, and then I can do a little

[14:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=860s) tell you about the content and show you

[14:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=863s) technical details.

[14:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=866s) So, we saw some great ready-made

[14:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=869s) ChatGPT features for financial services.

[14:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=873s) So you have the data, and also some

[14:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=874s) specialized use cases.

[14:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=878s) So, in the first part we considered, in

[14:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=880s) Why ChatGPT will be beneficial for financial

[14:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=882s) services for business investment

[14:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=885s) advisors.

[14:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=888s) I think we want to show how

[14:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=890s) actually do this process

[14:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=892s) more effective because ultimately you want

[14:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=894s) communicate more with your advisor.

[15:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=900s) You don't want your team to

[15:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=902s) worked on preparation, analysis

[15:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=904s) briefcases, finding out what happened,

[15:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=906s) explaining this, preparing the meeting,

[15:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=909s) and also, perhaps, preparation for

[15:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=911s) client presentation that explains why

[15:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=913s) there have been changes in the portfolio and what is needed

[15:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=916s) do to balance it.

[15:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=920s) So, the first part is really a script.

[15:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=922s) using. You are implementing this

[15:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=925s) product, ChatGPT for financial services,

[15:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=927s) into your teams of investment advisors, and

[15:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=929s) they can significantly speed up the work

[15:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=932s) processes. This is a tool for the ultimate

[15:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=934s) user in the team.

[15:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=936s) Often, in my experience, they have very

[15:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=939s) different workflows or different methods

[15:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=941s) work.

[15:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=942s) So this can be customized, and we

[15:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=944s) Let us introduce you to this scenario.

[15:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=947s) The client has had changes in their portfolio, and

[15:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=950s) you want to go all the way: prepare

[15:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=952s) profit and loss statement or

[15:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=954s) portfolio attribution, and then show

[15:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=956s) to the customer what to do to get it back

[15:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=959s) to the balance.

[16:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=960s) So, this is scenario number one. And we

[16:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=963s) We provide a few tips for this.

[16:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=966s) And I still don't know who's in the room...Oh, yes, there is

[16:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=968s) question?

[16:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=979s) Interesting.

[16:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=980s) So, yes, we will fix it. This is normal.

[16:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=983s) . Yes.

[16:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=990s) So I'll work on it. You can

[16:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=993s) continue. So this site is like

[16:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=995s) last slide, right?

[16:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=996s) ChatGPT for financial services, it seems

[16:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=998s) , yes. So my colleague was working on this

[16:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1001s) , and I'll explain in a little more detail. Perhaps,

[16:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1003s) for those who are in the hall now. I don't know, in

[16:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1005s) There are several of us. What is your experience?

[16:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1008s) work? I mean, you're developers,

[16:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1010s) builders, engineers, specialists in

[16:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1012s) artificial intelligence, or any of you

[16:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1014s) engaged in technical development?

[16:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1017s) We have a lot of them, I can't see exactly, but

[17:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1020s) about half.

[17:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1023s) Therefore, I think for those who are more

[17:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1025s) practical and technical, I want to explain

[17:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1027s) the second part of the seminar: as soon as we

[17:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1030s) we will get access, this will be our service on

[17:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1032s) Codex side.

[17:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1035s) Desktop solution. Codex can also

[17:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1037s) be used as final

[17:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1039s) users, business users,

[17:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1041s) intellectual workers, but in

[17:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1043s) in this case we want to make it

[17:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1045s) more personalized because

[17:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1047s) let's assume you have successfully implemented

[17:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1049s) ChatGPT in the finance department of your

[17:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1051s) company, and it meets the needs

[17:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1054s) business users.

[17:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1055s) But the question arises: how can I do it?

[17:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1058s) expand? How do I make sure my data

[17:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1060s) worked with him? How can I do it?

[17:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1062s) set up?

[17:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1065s) So, for the second part of the seminar we

[17:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1067s) identified two main ways to do this

[17:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1069s) .

[17:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1070s) You can say: let's create

[17:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1072s) own program. I will get an API key from

[17:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1076s) OpenAI.

[17:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1076s) Let's make a web interface. This is easy to do.

[17:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1079s) also using Codex and web-

[18:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1081s) interface. I'm expanding on this in my

[18:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1083s) companies, and then I have a typical

[18:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1085s) software development life cycle

[18:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1087s) software.

[18:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1089s) Is it path A or path B—you just

[18:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1091s) Think about it.

[18:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1095s) Oh, my users already have

[18:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1097s) corporate ChatGPT for work.

[18:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1100s) Most of them already have a Codex.

[18:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1105s) Why not leave the needs of my business

[18:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1107s) within this ecosystem?

[18:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1113s) So, in the second part we will explore what

[18:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1115s) would you do as an internal developer,

[18:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1117s) using Codex to create

[18:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1119s) skills, plugins and extensions for

[18:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1121s) plugin.

[18:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1124s) Now I will explain what it is. That

[18:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1126s) actually implement a full

[18:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1128s) program. No? She could work.

[18:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1131s) inside the interface.

[18:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1133s) Yes, actually, I'll tell you what it is.

[18:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1136s) things.

[18:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1138s) You can customize workflows.

[19:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1141s) No? Probably everyone already knows what "

[19:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1144s) skill.

[19:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1146s) This is a markup that provides a little more

[19:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1148s) settings for you and is reusable.

[19:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1152s) But as soon as you want to connect your

[19:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1154s) data, you need an MCP, certain

[19:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1156s) tools, you need all of this

[19:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1158s) to determine. This is where the plugin comes in.

[19:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1161s) . A plugin is a bit more complicated, and you

[19:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1163s) you can customize it around your

[19:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1165s) needs. So, we create a plugin there. AND

[19:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1168s) then, with this plugin, he can

[19:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1169s) work on the desktop, as well as on

[19:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1172s) website.

[19:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1174s) And what we recently announced on

[19:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1176s) Dev Day is an extension for plugins.

[19:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1181s) So in the end we also do a little tweaking

[19:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1183s) user interface.

[19:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1186s) You can customize the interface to your liking

[19:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1189s) specific case. We have a consultant

[19:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1191s) , which would use ChatGPT for

[19:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1193s) financial services online. But the worker

[19:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1196s) the process you are implementing may also

[19:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1199s) be present in the desktop application, or

[20:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1201s) wrong?

[20:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1202s) And that's why we're investigating it now.

[20:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1206s) Isn't that right?

[20:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1208s) OK.

[20:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1209s) So, we'll get to that. So, probably,

[20:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1211s) I will have to say more. I can

[20:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1213s) tell you about it.

[20:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1215s) Maybe you can do it on...I can too

[20:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1218s) show content.

[20:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1219s) Yes, let's do this.

[20:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1220s) Yes. Yes. But I think the most interesting thing is,

[20:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1223s) what we did in this workshop, how

[20:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1225s) we will only gain access to it. And not

[20:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1228s) worry if you don't succeed, web-

[20:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1230s) the site will be available later.

[20:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1235s) In fact, what I, as a developer at the company

[20:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1238s) or in a company that already has these

[20:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1240s) platforms, I can do to make it easier

[20:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1242s) your job or reduce your time

[20:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1244s) application deployment?

[20:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1248s) You know, I think this is a great idea. It

[20:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1251s) becomes possible.

[20:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1253s) So, the question about built-ins

[20:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1255s) tools you mentioned, such as

[20:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1258s) transcripts, reports and all other data that

[21:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1260s) you have, and I assume that

[21:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1262s) the toolkit has some awareness.

[21:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1265s) It's simply because of the tools available.

[21:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1267s) and skills, have you done anything with the

[21:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1270s) model? Just curious what exactly it does

[21:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1272s) its much better.

[21:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1278s) The microphone on the podium is also turned on.

[21:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1280s) Good.

[21:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1281s) So, for those who don't have a full

[21:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1284s) link, here is the access code to the workshop.

[21:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1287s) Did you manage to get in?

[21:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1297s) Someone, please confirm that this works.

[21:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1300s) Okay, I see. Fix this.

[21:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1309s) Perhaps it is also worth explaining that we

[21:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1311s) we use under the hood for this

[21:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1314s) workshop.

[21:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1316s) This is the "sites" feature, hosting capability

[21:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1318s) website.

[22:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1321s) It's pretty simple, and in this case it's

[22:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1323s) public website.

[22:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1326s) This is very convenient for creating beautiful and

[22:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1328s) functional interfaces.

[22:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1333s) I'll just wait a little while so everyone can

[22:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1336s) set up a workshop.

[22:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1341s) We also send emails

[22:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1343s) regarding the workspace, as mentioned

[22:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1346s) earlier.

[22:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1349s) They have to come.

[22:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1352s) Yes, while we're at it, I also think,

[22:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1354s) that we have small bottles or something

[22:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1356s) similar for everyone here, not for everyone, but

[22:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1358s) for those participating today.

[22:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1365s) And to make it a little more interesting, as soon as you

[22:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1367s) you can register in the system, and I

[22:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1369s) I see that there are already a lot of people there.

[22:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1373s) So this is also created using

[22:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1375s) website application.

[22:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1378s) So everyone who was able to get to the site,

[23:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1380s) already received 100,000 virtual money.

[23:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1383s) We want to make it a little more fun. AND

[23:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1386s) , of course, you all know that the market always

[23:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1388s) fluctuates, it goes up, and now,

[23:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1390s) maybe a little down.

[23:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1393s) But if you follow all the steps, you

[23:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1395s) receive additional funding, and

[23:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1397s) at the end of the day, at the end of the seminar, we

[23:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1400s) we want to see the top three

[23:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1402s) participants who will definitely receive bottles.

[23:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1405s) So this is to make the process

[23:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1407s) a little more interesting and maintain interest.

[23:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1411s) But with regard to the website, I think we see that

[23:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1414s) people are already there.

[23:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1417s) So what do I actually want to point out?

[23:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1424s) The first thing we need to do is get

[23:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1426s) access to ChatGPT for financial services

[23:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1428s) , it's actually this site where you see

[23:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1431s) financial data and ChatGPT plugin for

[23:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1433s) financial services.

[23:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1437s) This is what my colleague is working on, who

[24:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1440s) will register you.

[24:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1443s) When you access this site

[24:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1445s) with ChatGPT for financial services and

[24:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1447s) financial data that we showed in

[24:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1449s) presentations, you can freely move

[24:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1452s) to follow the instructions in the section

[24:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1454s) basic capabilities.

[24:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1456s) Isn't that right?

[24:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1459s) You can do the setup once

[24:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1461s) , we have a zip file with data for

[24:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1463s) client example, etc., this is a real case

[24:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1466s) on capital management.

[24:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1470s) How can I help my consultant?

[24:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1472s) work faster?

[24:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1475s) I would definitely recommend each of them.

[24:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1477s) work on the subtopic independently.

[24:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1481s) So if you are interested in portfolio issues

[24:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1484s) , research or graphics, boldly

[24:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1486s) start any of these sections

[24:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1488s) independently.

[24:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1490s) And when you do this and it arises

[24:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1492s) question, please raise your hand

[24:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1494s) hall. We want to help you.

[24:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1497s) This really demonstrates the main points and

[25:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1500s) key product features.

[25:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1504s) And for those who are more interested in development,

[25:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1507s) I think the question arises whether it is necessary

[25:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1509s) have Codex on your desktop now.

[25:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1514s) So for those who are connected, you can

[25:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1516s) install it. Hmm, I think this is the second one.

[25:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1519s) part, and I, I think, do not hesitate.

[25:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1522s) That is, if you want to better imagine this

[25:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1524s) the product, what it makes, what it has

[25:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1526s) basic capabilities, stay within limits

[25:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1528s) basic functions when you get access

[25:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1531s) to the site.

[25:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1533s) But with Codex, um, I think that's the idea,

[25:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1535s) which I spoke about earlier.

[25:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1540s) How can we develop, perhaps, AI-

[25:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1542s) tool for corporate applications

[25:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1545s) , work processes, etc.

[25:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1550s) So, here, um, we'll walk you through

[25:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1552s) skills plugin.

[25:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1554s) Um, we're also considering the scenario

[25:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1556s) using MCP here to complement this

[25:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1558s) the workflow we perform with

[26:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1560s) first side. And then, um, I definitely want to

[26:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1563s) emphasize plugin extension attempt

[26:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1565s) . And I think maybe what is the extension

[26:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1568s) plugin? I just want to explain this a little bit.

[26:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1571s) .

[26:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1571s) Um, I think we, we, I think people now

[26:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1574s) know that there are ways to expand,

[26:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1577s) um, AI capabilities through skills

[26:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1580s) .

[26:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1580s) And a plugin is something a little more substantial, and you

[26:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1583s) you see the structure. He has skills. He

[26:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1586s) has an MCP server. It has a configuration

[26:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1588s) tools. And you probably have

[26:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1591s) user interface to display

[26:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1593s) it's as you want it. This is cool. This is it.

[26:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1596s) exactly what I mentioned.

[26:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1598s) But I want to say that it is also possible

[26:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1601s) now.

[26:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1603s) This essentially means you have ChatGPT or

[26:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1606s) the Codex desktop app, and you see

[26:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1608s) these highlighted areas here.

[26:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1611s) You can actually customize it

[26:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1614s) depending on the conversation that yours might be having

[26:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1616s) the wealth management business is conducted with

[26:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1619s) Codex or ChatGPT.

[27:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1620s) Isn't that right?

[27:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1623s) So you can say: if I enter

[27:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1625s) a certain thing and it's a certain type of file, it

[27:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1627s) suddenly appears in the interface and

[27:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1629s) extension that you control, you can

[27:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1631s) expand and do more

[27:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1633s) individual.

[27:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1635s) Isn't that right? And there are actually several

[27:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1637s) examples. So, yes. I want to highlight these

[27:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1640s) things, um, we could, um, don't be shy

[27:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1642s) to proceed further.

[27:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1644s) So, people who already have Codex,

[27:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1646s) start with the setup.

[27:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1650s) And the people who still access the site

[27:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1653s) , um, please start, um, with

[27:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1655s) settings there.

[27:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1658s) Yes, there was actually a question that is

[27:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1660s) important.

[27:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1663s) So, therefore, ChatGPT for financial

[27:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1665s) services exist exclusively on our website

[27:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1668s) ChatGPT interface.

[27:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1672s) Isn't that right?

[27:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1673s) So, this may change, but right now it is

[27:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1676s) the only way to access it.

[27:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1678s) Isn't that right? So you, in essence, I have

[28:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1681s) I mean, you have a kind of, I saw

[28:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1683s) businesses that have a web version

[28:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1685s) ChatGPT, they have Codex, they have

[28:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1687s) both, and I even know users who have

[28:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1690s) there are both. Um, so you have to do a little

[28:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1692s) to stitch it together. In the future this may

[28:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1695s) change, but right now it's right there.

[28:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1697s) Uh, but that, that doesn't matter.

[28:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1700s) You can extend plugins to

[28:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1702s) desktop, as well as in the working version

[28:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1705s) ChatGPT. Now the plugins are there too

[28:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1707s) also. So, you have these options.

[28:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1711s) And now the plugin expansion,

[28:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1712s) interface settings only exist in

[28:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1715s) desktop versions.

[28:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1718s) So, you can't do it on the web yet,

[28:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1720s) but it probably will be.

[28:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1722s) Our website explains that this is coming soon

[28:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1725s) will appear.

[28:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1727s) Uh, so just keep that in mind. Uh,

[28:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1729s) Everything here is quite new.

[28:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1732s) And, uh, interfaces provide different

[28:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1734s) opportunities.

[28:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1736s) Just be aware of it. And now, if you

[28:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1739s) a real developer and you say, "Hey, I

[29:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1742s) I want to receive financial data through

[29:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1744s) command line interface (CLI)”.

[29:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1746s) Since Codex also has a CLI, you want

[29:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1748s) add this to the pipeline, but you can't

[29:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1751s) access the data now. Yes?

[29:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1754s) Not yet, or it just hasn't been provided yet.

[29:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1759s) So, as mentioned earlier, during

[29:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1761s) you can choose from ChatGPT jobs

[29:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1763s) model.

[29:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1766s) We recommend you choose Astra or 6.1 Soul

[29:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1769s) .

[29:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1771s) And if you press the speed, you

[29:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1774s) you can choose "ultrafast".

[29:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1777s) This will make your life much easier.

[29:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1780s) .

[29:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1780s) Therefore,

[29:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1781s) And yes, these are just hints. You

[29:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1783s) Follow the prompts.

[29:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1785s) And I mean, you learn by reading.

[29:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1788s) a little about what the product offers

[29:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1790s) you.

[29:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1792s) So, and I mean ChatGPT for

[29:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1793s) financial services, I think when you

[29:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1796s) enter this market, historical market

[29:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1798s) data, as well as research, you see

[30:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1801s) some key features.

[30:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1803s) Now, as Stefano said, we

[30:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1805s) we focus on setting up

[30:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1807s) This is better in the product.

[30:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1809s) Not fine-tuned, but

[30:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1811s) customize for your cases

[30:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1812s) using. Now,

[30:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1818s) So, everyone has email, everyone

[30:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1820s) entered the correct workspace and

[30:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1823s) can use ChatGPT for

[30:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1825s) Robots are fantastic.

[30:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1829s) Almost everyone managed.

[30:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1832s) So, I mean, before that they

[30:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1834s) they just want to listen.

[30:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1837s) I can do a few other things at the same time.

[30:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1839s) . Uh, but if there's a question, I think we

[30:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1841s) We would rather help you with

[30:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1843s) questions, so don't be shy, you know?

[30:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1845s) Yes.

[30:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1847s) You mentioned the specifics of our

[30:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1849s) financial assistance.

[30:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1851s) You mentioned that special setting

[30:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1853s) for financial services has...Is there more?

[30:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1856s) one level of ontology? What else are you

[30:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1858s) built? It seems that it was before

[31:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1860s) question. I didn't quite get an answer.

[31:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1863s) . What specific steps have you taken, besides

[31:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1865s) the model itself?

[31:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1867s) I think, as I mentioned before, this

[31:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1870s) primarily data ontology.

[31:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1873s) That is, it is not just connecting sources

[31:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1875s) data as if it were some kind of API call.

[31:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1879s) We actually changed the data ontology,

[31:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1882s) to make the model extremely easy

[31:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1884s) get the necessary information.

[31:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1886s) There is also some context regarding

[31:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1889s) what is relevant for each type

[31:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1891s) data sources and what not. For example, if

[31:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1893s) take 10K or 10Q reports, then it is not

[31:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1896s) only in providing the file, but also in explaining

[31:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1898s) of what exactly is in it.

[31:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1902s) I think that in terms of data, this is key.

[31:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1904s) difference.

[31:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1908s) This applies not only to integrated

[31:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1910s) sources, but also plugins from our partners

[31:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1912s) -data providers that can be accessed

[31:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1915s) enter if you have access—concept and

[31:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1917s) itself.

[31:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1919s) So, it is important not just to get the data, but

[32:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1921s) get them in the right context and

[32:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1924s) in the proper format. Yes.

[32:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1929s) And when we talk about plugins, this,

[32:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1931s) of course, not just any markdown files

[32:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1933s) or templates, but real expert knowledge

[32:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1936s) .

[32:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1940s) We have been cooperating with for quite a long time.

[32:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1942s) financial institutions.

[32:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1944s) When we created this product, we

[32:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1946s) selected a number of partners, including leading

[32:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1949s) investment banks they worked with

[32:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1951s) to check product quality.

[32:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1956s) So this is literally the baggage of knowledge that

[32:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1959s) we used to transform the model

[32:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1961s) on a product for financial experts.

[32:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1965s) So.

[32:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1966s) Yes.

[32:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1970s) I think if you try to do this

[32:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1972s) on your own, build a level of context

[32:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1975s) and ontology yourself at scale, you

[32:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1977s) you will encounter huge amounts of data

[33:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1980s) for processing.

[33:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1984s) Therefore, you can do this in

[33:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1986s) scalable product using

[33:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1988s) agents working with both public

[33:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1990s) data, as well as with suppliers,

[33:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1992s) ensuring the search for the right information

[33:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1994s) .

[33:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1996s) We have several blog posts about

[33:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1998s) how our data architecture is structured

[33:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2001s) .

[33:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2003s) There is interesting information there about services that

[33:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2006s) we use, in particular, open-source,

[33:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2008s) and this demonstrates our experience in building

[33:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2010s) effective context for billions

[33:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2012s) users.

[33:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2016s) And this is the same thing that is now working "under"

[33:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2018s) "hood" of this product. So I think that

[33:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2021s) this is perhaps the engineering side, now in

[33:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2023s) in the context of engineering. Hmm, it's early now.

[33:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2026s) product and I think you'll see how we

[33:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2028s) we will develop it, but also listen,

[33:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2030s) ChatGPT currently does not work through the interface

[33:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2033s) command line, and so, let's see how

[33:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2035s) development will take place here.

[33:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2037s) Is there a way to access some of the

[34:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2040s) these more primitive data structures,

[34:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2042s) that you have developed, or get an ontology

[34:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2044s) , similar to a graph?

[34:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2048s) Well, now what's "under the hood" isn't really

[34:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2050s) available because it is a product,

[34:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2052s) intended a little more for

[34:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2054s) experienced user, not for

[34:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2056s) pure engineer.

[34:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2060s) But I think

[34:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2064s) without going into too many details,

[34:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2066s) literally every data set was involved

[34:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2068s) together with the supplier to really

[34:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2071s) understand the best way to present it

[34:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2073s) , and, moreover, it is mostly an understanding

[34:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2075s) the data set itself.

[34:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2078s) So it's not easy, you know, which one?

[34:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2080s) Is the information relevant?

[34:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2084s) Like, you know, a simple example, when you

[34:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2086s) load market prices, prices against

[34:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2089s) profitability, explaining all these details,

[34:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2091s) for example, if you want to do this

[34:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2093s) operation if you want to extract this

[34:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2096s) information on how to avoid basic

[34:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2098s) misconceptions and so on.

[35:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2101s) So,

[35:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2102s) You, when you talk about 10K and 10Q,

[35:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2105s) Yes.

[35:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2106s) regarding embeddings, do you use

[35:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2109s) special embeddings, etc.?

[35:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2111s) Is this the backend of this?

[35:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2113s) No, no, no, no.

[35:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2115s) We literally created a dataset and

[35:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2117s) pre-processed it in this way,

[35:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2119s) so that the model can easily access it

[35:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2121s) access.

[35:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2122s) Think of it as a kind of

[35:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2124s) tool challenge, but very

[35:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2126s) specialized call for this set

[35:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2128s) data. So,

[35:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2131s) Don't worry.

[35:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2132s) Yes. I'm just showing you. I have on

[35:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2134s) I mean, I'm doing prompts now. Now,

[35:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2137s) if you are on the site, you just copy

[35:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2139s) prompt, insert it into the product, and,

[35:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2141s) besides, when you see the icons here, you

[35:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2144s) you can see a little bit that he actually picked up

[35:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2146s) plugin.

[35:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2148s) So, consider a plugin to be a way to

[35:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2150s) which the prompt hints at.

[35:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2152s) I hinted at the prompt to use

[35:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2155s) ChatGPT for financial services, so

[35:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2157s) now it follows the plugin concept

[35:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2159s) .

[36:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2160s) And then, you can see the financial

[36:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2162s) data used "under the hood"

[36:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2165s) , as well as internal tool calls

[36:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2167s) in this plugin.

[36:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2170s) So it's really like a plugin concept

[36:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2173s) now, and we provide a scalable backend

[36:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2175s) "under the hood".

[36:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2179s) So I would say I will give you a very

[36:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2180s) general overview of this product.

[36:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2186s) I think this is really cool because it

[36:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2188s) interface that will be used

[36:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2191s) every analyst, but in the second part we

[36:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2193s) Let's think about how to add plugins.

[36:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2199s) We all know that these data sources

[36:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2201s) not enough.

[36:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2202s) You probably have your own internal

[36:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2205s) tools. You will have your own sets

[36:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2207s) data. You can create your own plugin

[36:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2210s) with their capabilities. Especially as we

[36:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2212s) mentioned earlier if you want to add

[36:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2215s) the last layer of settings.

[36:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2218s) As an engineer, you can create

[37:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2220s) a tool that can then be

[37:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2222s) use as a plugin in ChatGPT for

[37:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2224s) Work.

[37:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2226s) So your end user will have this as

[37:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2229s) basis.

[37:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2232s) Let's say this is a very solid foundation for

[37:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2234s) interaction with data, creation

[37:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2236s) financial analysis and reporting

[37:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2239s) in Excel or PowerPoint.

[37:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2242s) But you can add a level

[37:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2244s) personalization that the product will make

[37:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2246s) unique due to specific

[37:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2248s) tools, data and logic of your

[37:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2250s) companies.

[37:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2251s) So

[37:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2253s) Yes. Just a note.

[37:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2255s) You can do it yourself, but I think

[37:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2258s) what does this show internal challenges

[37:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2260s) tools in financial services

[37:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2262s) ChatGPT.

[37:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2264s) You don't have full access to the context

[37:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2266s) there.

[37:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2267s) But the second part is the interface, which

[37:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2270s) is being configured.

[37:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2272s) We recently added an intelligent

[37:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2274s) interface.

[37:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2278s) Now that you get the answer, it

[38:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2281s) has a more adapted interface.

[38:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2285s) And here is one example. This is what also

[38:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2288s) provides a plugin.

[38:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2290s) It has a context: when I ask about

[38:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2292s) income, you want to see it in a certain

[38:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2295s) in the form of a table or perhaps a

[38:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2297s) graphics.

[38:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2298s) So these are additional opportunities.

[38:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2302s) If you were using the standard

[38:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2304s) model, it wouldn't do it automatically

[38:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2306s) .

[38:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2307s) Now you have to do this.

[38:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2309s) independently.

[38:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2311s) I just want to point out these little things that

[38:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2313s) matter because it's already a little, it's

[38:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2316s) not a fully trained model, but it has some

[38:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2318s) context, how to present data.

[38:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2321s) Yes.

[38:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2334s) Let's check the progress here.

[38:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2339s) Okay, I see that Mike, Zach, and Kuldeep

[39:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2342s) are doing well.

[39:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2347s) Oh, and it's also available from mobile, by the way

[39:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2349s) just in case you want

[39:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2351s) use a different interface. This is not

[39:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2354s) for desktop computers only. It

[39:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2356s) also works from the mobile app. It

[39:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2358s) works.

[39:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2359s) Yes.

[39:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2361s) So, I think that's the only thing I want.

[39:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2363s) note: market prices would be great

[39:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2366s) because I know that now you have to

[39:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2368s) to be in this service, you know?

[39:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2372s) So if I ask now: give me

[39:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2374s) historical market prices, and, um, yes, I

[39:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2376s) I think it would be really cool to have an API, but,

[39:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2379s) as always with data providers, you

[39:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2381s) must be careful with how here

[39:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2383s) Licensing is working.

[39:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2387s) But yeah, in this case, it's essentially,

[39:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2389s) a very important thing.

[39:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2392s) You want to know market prices in a simple way

[39:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2394s) way, and that would also be an opportunity,

[39:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2396s) if I created this internally,

[39:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2399s) to take on this because you all have

[40:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2401s) these internal systems, and then have

[40:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2403s) extension, sorry, plugin, which

[40:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2406s) is located, perhaps, in ChatGPT, and

[40:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2408s) also on the desktop side, and

[40:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2410s) does exactly that, you know, because you have

[40:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2412s) already have databases, access to internal

[40:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2415s) data.

[40:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2416s) And then, yes, maybe use

[40:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2419s) the same visualizations, and then provide

[40:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2421s) expansion in addition to the finished one

[40:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2423s) product, you know.

[40:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2435s) Yes, it, um, takes a little time. So

[40:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2438s) even small requests here take a while

[40:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2440s) minutes.

[40:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2441s) So if you're impatient and want some

[40:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2444s) rather, you know how to change the model,

[40:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2446s) or you, in fact,

[40:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2448s) Actually, you know what you can show

[40:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2451s) . For example, if you click on

[40:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2453s) a small zipper on the left.

[40:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2456s) You can go, oh, no, oh, no

[41:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2461s) No, only for 61 souls. If you go

[41:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2465s) on 61 souls

[41:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2467s) I thought ultra-fast mode was here

[41:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2470s) allowed.

[41:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2471s) Okay, fantastic. Use this.

[41:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2474s) This is really beautiful.

[41:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2475s) Yes. Yes.

[41:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2476s) I mean, the advantage of working at OpenAI

[41:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2478s) in that you can use

[41:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2480s) ultra fast mode.

[41:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2485s) No, it has to be, I also work on the web.

[41:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2489s) Yes.

[41:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2490s) Oh, yes. I mean, yes. I chose

[41:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2493s) the right model now, so yeah. Therefore,

[41:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2495s) yes, use ultra and me too

[41:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2497s) I think that, as I said, the advantage

[41:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2499s) working at OpenAI is that you

[41:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2501s) you can actively use ultra in

[41:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2503s) to your daily work.

[41:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2506s) Actually, I think it's like night and day.

[41:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2511s) It seems to me that if this becomes more

[41:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2513s) accessible to everyone, then just think

[41:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2515s) how many people sometimes waited for

[41:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2517s) performing some tasks by an agent, and

[41:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2519s) then they did a few more things in parallel

[42:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2521s) .

[42:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2525s) But now, I think, yes, I really do.

[42:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2527s) I appreciate and want to point out that this is very

[42:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2529s) helps with work.

[42:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2534s) I would say that on this topic, summing up

[42:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2536s) many discussions with customers, I think that

[42:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2539s) a question of efficiency and cost, perhaps

[42:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2541s) , were a top priority during

[42:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2544s) the last three months.

[42:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2547s) Of course, we're talking about Astra now,

[42:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2549s) because it is our most productive

[42:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2551s) model.

[42:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2553s) But we noticed that if you create

[42:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2555s) something completely new, a new codebase or

[42:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2557s) you are doing research, using

[42:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2559s) the most powerful model gives you an advantage.

[42:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2564s) However, when you create something like

[42:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2566s) recurring task, you can

[42:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2568s) get top-level performance

[42:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2571s) if you build the right context.

[42:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2576s) So if you develop the right skill

[42:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2579s) , which explains in more detail what you

[43:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2581s) expect from a model, then even

[43:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2583s) using models such as Luna or

[43:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2585s) even 6.1 solar, actually very

[43:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2587s) effective and can give you the highest

[43:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2589s) results.

[43:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2592s) So you don't always have to choose

[43:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2595s) the most expensive or most modern model.

[43:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2598s) We have noticed that the general approach is

[43:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2601s) in that you can think about

[43:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2603s) building skills properly so that

[43:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2605s) maximize productivity.

[43:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2609s) And this is a common practice: if you work

[43:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2611s) in a multi-agent system, we

[43:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2613s) saw that the orchestrator probably has

[43:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2616s) to be the smartest agent, but some

[43:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2618s) subordinate agents do not necessarily have

[43:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2621s) to be like Astra.

[43:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2623s) You can switch to something on

[43:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2626s) like Terra or Luna and still get

[43:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2628s) approximately 95% of the total result.

[43:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2632s) Of course, to make all these measurements

[43:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2634s) , you need to consider the option

[43:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2636s) use, create estimates and

[43:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2638s) to actually measure the effectiveness of what

[44:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2640s) you do.

[44:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2642s) But you know, Astra is more suitable for

[44:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2644s) something like brand new

[44:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2646s) research or a new code base.

[44:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2652s) But also other models, especially even

[44:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2654s) Luna, which is our cheapest, I don't

[44:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2656s) I remember the price by heart, but it

[44:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2658s) extremely cheap, we saw that with

[44:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2660s) with the right skills it can be

[44:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2662s) use for content creation

[44:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2664s) extremely high quality, from reports

[44:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2667s) to analysis.

[44:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2671s) You just need to use

[44:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2672s) the right skill and make sure that with

[44:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2675s) this skill actually results

[44:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2677s) meets your expectations.

[44:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2682s) But it's a great strategy that, as we

[44:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2684s) seen, works everywhere, you know, for

[44:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2686s) cost optimization, especially if you

[44:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2689s) think about it, let's say you're running this

[44:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2691s) capital management service for

[44:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2693s) your entire customer base, which

[44:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2695s) scales quite quickly as per

[44:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2697s) both in cost and time, use

[45:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2700s) models like Luna, with the right

[45:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2702s) a set of skills, can really help

[45:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2704s) you to mitigate all these problems.

[45:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2710s) Sorry.

[45:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2711s) Yes. No, I think, yes, I think the point is

[45:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2713s) in that models change rapidly, and

[45:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2716s) new models are emerging.

[45:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2718s) But yes, often, I mean that

[45:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2720s) every time we have a model, I

[45:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2722s) I also think you should just

[45:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2724s) find out what your current effort level is

[45:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2727s) because opportunities sometimes combine

[45:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2729s) other than price, compared to

[45:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2731s) opportunities. So I think that now in

[45:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2734s) I have a lot of daily life

[45:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2736s) I use 6.1 Soul, and then, of course,

[45:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2738s) I can afford the luxury.

[45:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2740s) use Ultra.

[45:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2744s) But I think sometimes when I was creating

[45:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2746s) the workshop, the website itself, it was all

[45:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2748s) created, of course, with Astra,

[45:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2751s) and even more advanced live viewing

[45:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2754s) rooms, I think, yes, these are the things that I

[45:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2756s) , essentially, passed on Astra.

[46:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2760s) I let it work. I did it in

[46:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2762s) Friday, so it started and

[46:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2764s) came back after that, and then I

[46:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2766s) just slightly modified the website for

[46:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2768s) using Soul because it was faster and

[46:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2770s) more efficient. And yes, I think very much

[46:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2773s) important every time it appears

[46:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2775s) new model, explore possibilities,

[46:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2777s) level of effort, and yes, it's always

[46:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2779s) changes, I can't say right now how

[46:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2781s) you can.

[46:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2784s) I would say, come on, while you're passing by.

[46:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2787s) different exercises, the question for you will be like this

[46:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2789s) :

[46:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2793s) some initial thoughts on how you

[46:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2795s) currently using this model, this type

[46:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2798s) products and models in your daily

[46:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2800s) work, and which, in your opinion, are

[46:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2802s) limitation or potential?

[46:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2805s) Someone wants

[46:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2808s) express passionate opinions about the use

[46:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2811s) such products?

[46:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2814s) No one?

[46:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2818s) Yes.

[46:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2819s) Good.

[47:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2821s) We'll have a few more questions later.

[47:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2823s) , but I think when I first saw this

[47:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2825s) product, we didn't have it inside before

[47:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2828s) saw.

[47:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2829s) I said, "Okay, great." But to me

[47:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2832s) need a little more, so maybe we

[47:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2834s) we can change the focus, this product exists

[47:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2836s) , and I think it's definitely focused on

[47:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2838s) intellectual workers. But let's

[47:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2841s) Let's think about developers. I have

[47:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2843s) a certain passion for this because my

[47:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2846s) experience is also related to development.

[47:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2851s) And I want to dwell a little more on plugins

[47:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2854s) .

[47:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2857s) I think the plugin architecture,

[47:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2859s) folder structure, what's inside

[47:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2861s) plugin, it is well described here.

[47:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2864s) So you see the skills, MCP server,

[47:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2866s) which connects to something outside

[47:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2868s) also through tools and then you have

[47:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2870s) some interface components that may

[47:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2872s) displayed during use

[47:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2874s) plugin.

[48:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2880s) And ChatGPT for financial services is

[48:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2882s) plugin, and the thing you can do with

[48:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2884s) such architecture, and this is perhaps

[48:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2886s) transition to developers, is to come up with

[48:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2889s) architecture that uses all of this

[48:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2891s) in a way that brings together all these different

[48:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2893s) parts together.

[48:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2900s) So, this is a kind of architectural

[48:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2902s) diagram because you have to put

[48:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2904s) Question: What is your surface? When I

[48:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2907s) I say "surface", we have ChatGPT, which

[48:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2909s) is on the website, we have a Codex, which

[48:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2911s) is located locally, or we have

[48:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2913s) share locally on the desktop

[48:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2916s) table. And as a developer, I mean,

[48:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2918s) that we all now usually have access to

[48:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2920s) Codex, but I see that it is more

[48:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2922s) comes to knowledge workers

[48:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2924s) , and there will probably be more in the future

[48:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2926s) similarities in the interface for both

[48:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2928s) use cases.

[48:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2930s) So, it's on your desktop, and you

[48:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2932s) you can do much more in Codex,

[48:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2935s) because now you can install

[48:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2937s) local MCP.

[48:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2938s) Now you can create using

[49:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2940s) code and hints MCP server, which,

[49:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2942s) maybe referring to your

[49:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2944s) internal database of securities portfolios

[49:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2947s) papers or customer database. Perhaps

[49:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2949s) , our Salesforce could be there too

[49:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2951s) use. But then you create

[49:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2954s) It's there.

[49:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2956s) You have a prototype, usually being

[49:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2958s) developer. You are actually connecting

[49:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2961s) to your local website and have

[49:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2964s) Codex is running, which connects to

[49:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2966s) MCP.

[49:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2968s) But you also want to use

[49:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2970s) plugin infrastructure. Therefore,

[49:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2972s) the plugin infrastructure is hosted in

[49:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2974s) ChatGPT. It can be administered. So I

[49:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2977s) , maybe I'll just point that out.

[49:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2980s) So, if you want your own

[49:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2982s) plugins worked in ChatGPT, you can

[49:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2985s) install a tool to create them.

[49:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2993s) Yes, you can install this plugin.

[49:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2997s) And when you install this plugin, you

[50:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3000s) you can also add plugins to the desktop

[50:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3002s) ChatGPT website.

[50:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3004s) This is another way to set it up. The same

[50:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3007s) You can also do it on the desktop version. Not

[50:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3010s) I don't know if I confused you.

[50:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3012s) This is essentially an interface, and you have different

[50:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3015s) ways to configure it.

[50:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3018s) So just know that the web plugin

[50:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3020s) website—this is one option.

[50:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3023s) And then on the desktop you, of course,

[50:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3025s) you can install plugins. And if in

[50:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3028s) you are users, you can develop

[50:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3030s) corporate mechanism for deployment

[50:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3033s) these plugins along with your Codex.

[50:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3035s) For example, in our management business

[50:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3038s) assets everyone gets Codex, as well as

[50:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3040s) access to ChatGPT for work. You can

[50:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3043s) define the folder structure with plugins,

[50:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3045s) which is managed by the enterprise, and then

[50:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3047s) deploy it through the existing mechanism. AND

[50:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3050s) this is a huge advantage.

[50:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3053s) I mean, if I were creating right now

[50:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3056s) app on my own, I would have to

[50:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3058s) think about login and access control.

[51:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3062s) So, if you create a third-party

[51:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3064s) application within the company, you need

[51:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3066s) a lot of time for its approval.

[51:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3069s) But if you create it in an environment

[51:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3072s) , which is already approved and managed, you

[51:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3074s) you can easily connect.

[51:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3077s) This, in fact, is the main idea.

[51:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3081s) So yes, we would then use

[51:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3083s) OpenAI plugin infrastructure.

[51:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3087s) In this case, ChatGPT for financial

[51:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3089s) services has access to our contextual

[51:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3092s) data, where there is financial information, etc.

[51:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3094s) This still only works in a web service. You

[51:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3096s) cannot use the interface

[51:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3098s) command line interface (CLI). Next step

[51:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3100s) you connect using your

[51:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3103s) a plugin to a tool that works with

[51:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3105s) your databases.

[51:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3109s) It could be another MCP, or you could

[51:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3111s) do it differently using

[51:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3113s) authentication, etc., ensuring

[51:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3116s) corporate way to connect to

[51:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3118s) OpenAI hosting.

[52:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3121s) And all this is quite possible to implement.

[52:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3124s) So, that would be one way to do it.

[52:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3126s) a little more settings, staying in

[52:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3128s) this system.

[52:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3130s) And I think I'm just really excited.

[52:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3133s) from all these options now.

[52:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3135s) I mean, I love to create.

[52:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3138s) agents independently and place them, but

[52:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3140s) get permission for this in corporate

[52:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3143s) environment is often very difficult.

[52:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3147s) And if you can use something and,

[52:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3149s) besides, if it's in Cortex, there are certain

[52:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3151s) safeguards and policies that can

[52:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3153s) actually prevent the plugin from going beyond the limits

[52:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3156s) sandboxes.

[52:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3158s) So, the sandbox is in Cortex. IN

[52:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3161s) There are certain elements of control there.

[52:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3164s) And if you deploy the code, I just

[52:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3166s) need to find a corporate way

[52:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3169s) deploying my plugin.

[52:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3171s) So, I think this is a great idea for

[52:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3173s) study, and in the future it may

[52:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3176s) only become better.

[52:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3179s) Yes.

[53:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3180s) And I think that when we talk about

[53:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3182s) corporate control, for example,

[53:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3184s) while remaining within the ChatGPT environment,

[53:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3187s) and about which plugin can be deployed

[53:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3189s) for which employee, it is really very

[53:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3192s) common request.

[53:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3193s) So all corporate controls

[53:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3195s) available here. You can really control

[53:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3198s) access, control who has access

[53:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3200s) to data, who has access to the plugin, etc.

[53:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3203s) .

[53:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3206s) So, I feel like what I mentioned

[53:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3208s) Oliver, it's really important, especially if

[53:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3211s) ChatGPT is your ultimate interface for

[53:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3213s) users, then turning an idea into

[53:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3215s) prototype into a deployed plugin that can

[53:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3218s) be used by your end

[53:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3220s) user, there is a lot going on

[53:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3222s) faster than we are used to.

[53:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3227s) Do we have any questions?

[53:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3235s) It depends on the plugin.

[53:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3237s) It depends on the plugin, it depends on

[54:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3240s) of what you need to do.

[54:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3242s) For example, a plugin can even be

[54:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3244s) a set of skills that, you know, I...

[54:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3254s) Yes, then you will have to create a plugin

[54:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3257s) , for example, MCP.

[54:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3262s) That's right.

[54:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3264s) And also manage authentication.

[54:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3267s) Yes.

[54:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3269s) Maybe think about this world where this

[54:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3272s) is already a platform, and I think it's already

[54:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3274s) So there is something to do in such conditions.

[54:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3278s) So, first, skill, we have some

[54:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3281s) markup files.

[54:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3284s) So I would create in natural language, maybe

[54:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3286s) , a couple of markup files.

[54:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3289s) What to do, and I also think that now

[54:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3291s) it's quite easy to create a skill with a certain

[54:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3294s) with help from Codex.

[54:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3298s) And this, you could say, is a kind of method

[55:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3300s) reuse of hints, I would

[55:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3302s) said, or knowledge, and I think that this

[55:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3305s) now available on the Codex side.

[55:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3308s) So, if you have a developer, we

[55:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3310s) we usually use skills to

[55:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3312s) work faster. In fact, on the web

[55:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3314s) This site is currently unavailable. But interesting

[55:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3317s) part is that a plugin like you

[55:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3319s) They said this is something I can't handle.

[55:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3321s) fully implement corporate

[55:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3323s) integration. I have a certain opportunity.

[55:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3325s) reuse in the workplace

[55:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3327s) process that I can scale and

[55:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3329s) share it within the company.

[55:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3331s) But the plugin is where "oh, now I

[55:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3333s) you need to authenticate the system with

[55:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3336s) system or maybe I want to do this

[55:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3338s) on behalf of the user.

[55:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3341s) We need to authenticate somehow

[55:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3343s) the user who actually logged in

[55:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3346s) your Codex system.

[55:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3348s) So that's why plugins will be a package where

[55:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3350s) you have opportunities, maybe some too

[55:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3353s) skills you want to re

[55:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3355s) use, but as you noted, I

[55:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3357s) some code is needed.

[55:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3359s) I need some logic. Me

[56:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3361s) you need to have a way to get accounts

[56:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3364s) authentication data, etc. So, this is

[56:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3366s) and there is a plugin. This is a plugin, and you saw it

[56:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3369s) folder structure before.

[56:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3371s) And it also uses MCP connections, and

[56:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3373s) also contains an interface component

[56:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3375s) user. There is an interface standard

[56:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3378s) programs, a kind of way of creating

[56:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3380s) all interface components using

[56:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3382s) plugin.

[56:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3383s) And then all these things can just

[56:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3385s) pack it into an archive, and that's the plugin.

[56:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3389s) So, I would say plugins really

[56:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3391s) the most powerful because they contain

[56:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3393s) a superset of everything.

[56:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3399s) And then, what I mentioned earlier:

[56:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3401s) when you are already in the interface, then if I

[56:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3403s) going back to plugin extensions, I want

[56:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3405s) to give you a good example so that

[56:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3407s) understand, because it was difficult for me to understand,

[56:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3410s) what is an extension, but here it is, maybe

[56:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3412s) , more clearly.

[56:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3416s) So I think that's an example.

[56:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3419s) I mean, you see the interface,

[57:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3422s) familiar to you from ChatGPT, but at the bottom left

[57:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3424s) you see the option to add your own

[57:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3427s) icon.

[57:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3429s) Um, your own app, essentially.

[57:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3432s) And then when you click on it,

[57:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3434s) now it's all controlled by the plugin you

[57:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3437s) you can control. So, the beauty of it is

[57:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3439s) that it's not just a little icon that

[57:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3442s) can be customized.

[57:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3444s) You can customize different parts

[57:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3446s) interface.

[57:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3449s) So, we can adjust the parameters.

[57:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3451s) We can customize the navigation.

[57:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3454s) Also, think about this.

[57:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3457s) I mean, if I come to now

[57:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3459s) company and I will have to go to

[57:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3462s) five different sites to get the job done

[57:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3464s) , I don't really want this. I want

[57:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3466s) to stay in your favorite

[57:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3468s) a tool that works on the basis of AI and,

[57:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3470s) I hope it has a very good front end.

[57:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3472s) model. Hmm, and I could go on.

[57:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3474s) communication for your workflow,

[57:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3476s) and I don't need to leave him. Therefore,

[57:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3479s) What I would like to do: I work as

[58:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3481s) end user, but then it

[58:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3483s) appears, it knows the context.

[58:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3488s) Oh, I have something special now,

[58:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3490s) possibly related to preparation for

[58:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3492s) meeting with a client, right?

[58:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3495s) For example, a meeting with a client, to which

[58:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3497s) you need to prepare.

[58:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3499s) So, it knows about it in conversation and

[58:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3501s) will show my customized interface for

[58:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3503s) this.

[58:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3506s) Now maybe your management business

[58:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3508s) capital asked you about a certain thing that

[58:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3510s) you want to do it faster, or you want

[58:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3513s) to see a certain indicator from the inside

[58:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3515s) systems for this customer profile.

[58:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3518s) This is what you can do now, and you

[58:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3520s) no need to leave the workspace.

[58:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3523s) So, I guess I want to give you an idea

[58:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3526s) about where these concepts are located.

[58:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3530s) And if you take this from this session, it's

[58:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3532s) That would be great.

[58:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3535s) Yes.

[58:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3537s) So, any questions?

[59:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3540s) Yes, maybe I'll switch in mine.

[59:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3542s) visualizations, I set up the Codex side

[59:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3545s) . That would also be a question in the audience.

[59:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3549s) Um, who already has Codex running locally?

[59:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3551s) .

[59:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3554s) Okay, I see.

[59:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3556s) Yes, quite a few people.

[59:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3558s) Yes, Codex now works through the app

[59:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3560s) ChatGPT, where they are in the same interface.

[59:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3567s) Therefore.

[59:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3577s) Yes, I'm just going to experiment a little bit, because

[59:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3579s) now to me it looks more like

[59:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3581s) creating a plugin.

[59:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3599s) Yes, and now everything becomes just a prompt,

[1:00:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3601s) isn't that right?

[1:00:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3631s) Yes, I think it will take some time,

[1:00:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3633s) ago

[1:00:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3635s) even with Ultra, because I think even if

[1:00:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3637s) you are using Ultra, you need

[1:00:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3639s) write code and compile or you

[1:00:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3641s) compile but run it locally

[1:00:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3643s) , no?

[1:00:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3644s) So maybe this is also one moment, because I,

[1:00:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3647s) Of course, I do a lot of this for people like that.

[1:00:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3649s) seminars.

[1:00:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3651s) It's basically like I'm working on something right now.

[1:00:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3654s) your desktop, right?

[1:00:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3656s) So I might have a few more open.

[1:00:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3658s) chats and do another project, maybe

[1:01:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3661s) create another plugin or something like that

[1:01:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3663s) .

[1:01:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3665s) But yes, I think you really want to

[1:01:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3667s) scale what actually works

[1:01:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3669s) agent, so you essentially want to have

[1:01:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3672s) cloud environment, no?

[1:01:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3674s) So, that's another thing that's very useful.

[1:01:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3677s) Um, I'm also doing a lot right now.

[1:01:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3679s) application security scans for

[1:01:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3681s) financial services. You know, this is very

[1:01:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3683s) common usage scenario. And so,

[1:01:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3686s) even on your desktop you take up

[1:01:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3688s) a large repository and do something on

[1:01:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3690s) like an application security check. It

[1:01:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3693s) locks your computer locally. You know,

[1:01:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3695s) so you want to move this to remote

[1:01:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3698s) side. You know, so in this case,

[1:01:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3700s) if you have many tasks running,

[1:01:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3702s) I think so, we have a few ready

[1:01:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3704s) opportunities, but you can also

[1:01:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3706s) configure where exactly they are executed

[1:01:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3708s) code calculation. You know, in this

[1:01:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3710s) case the code is run locally, which

[1:01:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3712s) takes time, even if the model is fast,

[1:01:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3715s) or as much time as mine has

[1:01:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3717s) laptop.

[1:01:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3718s) You know, so it would be nice to have those

[1:02:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3720s) infrastructure components as well

[1:02:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3722s) inside the company where you can

[1:02:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3724s) easy to run calculations.

[1:02:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3728s) You know, and then you hand over the task, and

[1:02:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3730s) this is essentially what is done when the job

[1:02:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3732s) completed.

[1:02:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3756s) Yes.

[1:02:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3762s) Yes.

[1:02:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3776s) Yes.

[1:02:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3778s) So, there are several approaches here.

[1:03:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3784s) Yes, the idea is, especially with

[1:03:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3786s) plugins and skills that not everyone has

[1:03:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3788s) individual skills, but actually create your own

[1:03:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3791s) kind of marketplace and share them between

[1:03:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3793s) teams.

[1:03:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3797s) And it also works with what we're talking about

[1:03:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3799s) mentioned earlier that it's more like

[1:03:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3801s) templates. So artifact templates can

[1:03:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3804s) be a website, a PowerPoint presentation

[1:03:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3806s) or Excel file.

[1:03:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3809s) We noticed that you want to have

[1:03:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3811s) possibility even inside plugins

[1:03:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3813s) add some historical examples like

[1:03:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3816s) files, and then the plugin will be

[1:03:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3817s) use this as a reference.

[1:03:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3823s) Let's assume we are in the context of a study,

[1:03:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3825s) which needs to be created for this case

[1:03:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3828s) , for example, for your private

[1:03:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3829s) wealth management clients.

[1:03:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3835s) You can actually provide several

[1:03:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3837s) historical examples, and then ask

[1:03:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3839s) within the skill to use it as

[1:04:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3841s) reference material.

[1:04:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3843s) Then, of course, you can change them, but

[1:04:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3845s) the idea is to do this

[1:04:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3847s) as reusable as possible.

[1:04:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3850s) Since we know that the format has a large

[1:04:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3853s) meaning, and that's why, even if it's not

[1:04:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3855s) completely related to the financial aspect,

[1:04:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3857s) we are making a lot of efforts to

[1:04:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3859s) creation of artifacts.

[1:04:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3864s) So, if you've probably noticed by now

[1:04:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3866s) the last 6 months, the ability of our

[1:04:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3868s) it's not easy to create the right models

[1:04:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3870s) spreadsheets, but also format them

[1:04:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3872s) at the highest level, corresponding to

[1:04:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3874s) investment banking standards,

[1:04:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3877s) has increased significantly.

[1:04:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3879s) We have several benchmark tests that we

[1:04:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3882s) we can show you.

[1:04:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3885s) But I think the best example is this:

[1:04:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3887s) try it if you have some template

[1:04:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3889s) spreadsheet, presentation, or PDF that you

[1:04:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3892s) want to use, and just say: "

[1:04:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3894s) Okay, take this new analysis and build it in.

[1:04:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3897s) it in this format."

[1:05:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3901s) You will see that the result is likely to be

[1:05:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3903s) will be of quite good quality, and, of course,

[1:05:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3905s) there will only be one last stage left, where you,

[1:05:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3907s) may need to adjust a bit

[1:05:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3910s) skills, saying, "Hey, here are some

[1:05:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3912s) "inaccuracies", but if you mention it in

[1:05:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3914s) skill, then next time everything will be fine

[1:05:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3916s) right.

[1:05:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3919s) So yeah, the best part is the opportunity

[1:05:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3921s) share skills because you too

[1:05:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3923s) you want to constantly improve them. Therefore

[1:05:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3926s) , another interesting point: is it possible

[1:05:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3928s) reuse skills between

[1:05:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3930s) different types of models or different

[1:05:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3932s) model families? The answer is:

[1:05:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3934s) if we move within the same boundaries

[1:05:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3936s) families themselves, then yes.

[1:05:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3939s) So the idea is that, perhaps, with

[1:05:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3942s) transition, say, from the fifth version to

[1:05:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3944s) sixth, that is, when jumping approximately

[1:05:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3946s) year, you will need some adjustments

[1:05:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3949s) .

[1:05:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3951s) But theoretically, between incremental

[1:05:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3954s) updates we see that usually

[1:05:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3956s) less refinement is needed because the models

[1:05:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3958s) are getting smarter, and these skills

[1:06:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3960s) , usually quite easy to reuse

[1:06:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3962s) use, and the costs of them

[1:06:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3964s) The service is quite acceptable, and

[1:06:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3966s) not overpriced.

[1:06:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3969s) So, um.

[1:06:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3970s) Yes. I think so, maybe with skills

[1:06:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3972s) you can go too far, so actually

[1:06:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3975s) Yes. Because I believe that the repetition

[1:06:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3977s) using skills and maintaining them...I

[1:06:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3979s) saw some customers a little

[1:06:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3981s) were excited, creating too much

[1:06:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3983s) skills, and then it becomes difficult for teams

[1:06:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3985s) share them, so they need to be shared somehow

[1:06:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3987s) standardize. So, it's rather

[1:06:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3989s) organizational problem.

[1:06:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3991s) But I think if you're an app developer

[1:06:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3994s) , I would lean more towards plugins,

[1:06:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3996s) because you are also in control

[1:06:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=3997s) deployment and you can pack better

[1:06:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4000s) decision.

[1:06:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4002s) If we are talking about the front office or

[1:06:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4005s) quantum, now almost everyone has

[1:06:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4007s) your skills in your code.

[1:06:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4010s) So, I think the question is, if two

[1:06:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4012s) people on your team do similar things

[1:06:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4014s) , creating a skill of how to make sure that

[1:06:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4017s) Can it be reused?

[1:07:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4021s) I would say, um, if you have a few

[1:07:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4023s) team members, and they find out that

[1:07:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4025s) perform similar tasks, perhaps with

[1:07:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4028s) small differences.

[1:07:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4031s) I strongly recommend removing some

[1:07:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4034s) skills in your team because they are hard to find

[1:07:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4036s) support, and I recently advised

[1:07:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4038s) to one client, which I think they have

[1:07:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4041s) too many skills.

[1:07:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4043s) Maybe try to find a good one

[1:07:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4045s) definitions for your keywords

[1:07:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4047s) opportunities.

[1:07:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4050s) If you are conducting a portfolio analysis or

[1:07:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4053s) trading analysis, leave it as a skill,

[1:07:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4055s) but not for the entire workflow, because

[1:07:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4057s) it becomes difficult to maintain.

[1:07:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4062s) And as for your comment, you don't care.

[1:07:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4064s) will have to retest the models because

[1:07:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4067s) one of the advantages of working with us is

[1:07:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4069s) access to early pre-production models

[1:07:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4071s) viewing.

[1:07:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4075s) So when you have a relationship with us, you

[1:07:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4078s) see what affects your business, you can

[1:08:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4080s) identify problems, test them

[1:08:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4082s) in advance and be sure that after

[1:08:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4085s) startup they work as expected.

[1:08:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4089s) But yes, my advice: if you want

[1:08:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4091s) control the process, adhere to

[1:08:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4093s) skills.

[1:08:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4095s) You have a developer who has already thought about it,

[1:08:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4097s) what exactly should be encapsulated.

[1:08:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4101s) And if too many people have too much

[1:08:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4104s) skills, this can slow down processing

[1:08:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4106s) data.

[1:08:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4108s) So there is no clear answer, but this

[1:08:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4111s) need to be managed.

[1:08:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4113s) This is my recommendation. Yes. I think, how

[1:08:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4116s) the general trend we are observing

[1:08:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4118s) , our team focuses mainly on

[1:08:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4120s) on API usage options. That is,

[1:08:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4123s) use cases where you know that

[1:08:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4125s) are going to use the API for

[1:08:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4127s) getting answers.

[1:08:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4135s) This is a different surface, but I work with OpenAI

[1:08:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4137s) for several months now, and now, for

[1:08:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4139s) such a short period, it is clear that many

[1:09:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4141s) options that were previously possible

[1:09:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4143s) implement only through writing

[1:09:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4146s) code using API and keys,

[1:09:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4148s) can now be perfectly reproduced by

[1:09:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4150s) using a plugin.

[1:09:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4159s) To give an example that is a bit off

[1:09:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4162s) beyond capital management.

[1:09:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4165s) A few months ago we created

[1:09:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4167s) internal case study on how to create

[1:09:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4169s) advanced financial research.

[1:09:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4174s) So, this can be seen as a profound

[1:09:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4176s) research, but specifically for

[1:09:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4178s) financial problems with using API

[1:09:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4180s) .

[1:09:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4182s) It was like the orchestration of several

[1:09:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4184s) agents.

[1:09:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4185s) The scheme was quite simple if you look at it

[1:09:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4188s) at a high level. So, you have what

[1:09:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4190s) we call it a planner.

[1:09:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4193s) This is essentially an agent that accepts

[1:09:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4195s) question, breaks it down into sub-questions

[1:09:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4198s) and for each subquery creates

[1:10:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4200s) a separate subagent who will conduct

[1:10:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4202s) independent research.

[1:10:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4208s) And then the critic agent appears, the agent

[1:10:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4211s) feedback that collects

[1:10:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4213s) the results of all these independent

[1:10:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4214s) studies, combines them and checks for

[1:10:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4217s) absence of conflicts.

[1:10:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4223s) If there is a difference between the different conclusions

[1:10:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4224s) contradictions, it automatically creates

[1:10:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4226s) a new subagent to investigate these

[1:10:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4229s) discrepancies until the final report is

[1:10:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4231s) will achieve the required quality and not

[1:10:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4233s) will answer all questions asked.

[1:10:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4238s) So think of it as orchestration.

[1:10:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4241s) subagents and overall dynamic

[1:10:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4243s) process management.

[1:10:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4247s) Now you can do it because Dom

[1:10:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4249s) actually recreated the same case for

[1:10:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4252s) using a plugin.

[1:10:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4254s) Speaking of in-depth research.

[1:10:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4259s) And it's much more convenient for customers,

[1:11:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4261s) because the end user does not

[1:11:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4263s) need to leave the interface to

[1:11:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4265s) switch to another product, it can

[1:11:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4268s) just stay in the Codex.

[1:11:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4272s) He can actually see what they are doing.

[1:11:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4275s) subagents.

[1:11:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4276s) What are they researching? You can even

[1:11:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4279s) create a special interface for

[1:11:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4281s) the user so that he understands that

[1:11:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4282s) is happening.

[1:11:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4286s) And with current capabilities

[1:11:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4288s) multi-agency you can see everything in

[1:11:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4290s) Codex, do everything in Codex and deploy

[1:11:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4292s) solutions instead of creating

[1:11:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4294s) new codebase—just create a plugin

[1:11:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4297s) and add it to the Codex.

[1:11:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4300s) Therefore,

[1:11:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4309s) I think we can ask the author.

[1:11:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4311s) Yes.

[1:11:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4312s) Yes, and then there's Matthew, Zach, and Conan. I think

[1:11:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4315s) , you guys are leading the way here. This is interesting.

[1:11:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4317s) thing. Let's see at the end.

[1:12:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4320s) This is a serious race.

[1:12:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4323s) Yes, as far as where, why and when

[1:12:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4325s) use the capabilities of plugins, then

[1:12:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4327s) a typical initial example was

[1:12:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4329s) in-depth stock research—it was an app

[1:12:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4332s) , which I created to perform deep

[1:12:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4334s) parallel proven work in this area

[1:12:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4337s) at a professional level.

[1:12:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4338s) Ultimately, this required a significant level of

[1:12:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4341s) orchestration. It meant a lot.

[1:12:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4342s) deterministic and non-deterministic

[1:12:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4344s) workloads within this

[1:12:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4345s) process. And it was a separate independent

[1:12:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4347s) addition. It became clear that many

[1:12:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4349s) the people who did this work, not

[1:12:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4351s) used separate disposable

[1:12:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4353s) applications for this. They started

[1:12:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4355s) consolidate the place of execution of this

[1:12:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4357s) work, and this place became Codex or Jet

[1:12:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4359s) B. So it makes a lot of sense to move there

[1:12:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4361s) this workspace. And the choice of how we

[1:12:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4363s) we want to rebuild it, it came down to

[1:12:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4365s) Question: Can I use

[1:12:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4367s) Codex capabilities, including its

[1:12:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4369s) functions of subagents, its ability

[1:12:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4371s) to direct dynamic agent structures

[1:12:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4373s) to connect plugins and various

[1:12:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4375s) data sources, instead of creating

[1:12:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4377s) special connections. And it became completely

[1:12:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4379s) it's logical to say, "Hey, this works like

[1:13:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4381s) a plugin specifically focused on

[1:13:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4383s) using the capabilities of the language model

[1:13:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4385s) to handle things like dynamic

[1:13:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4387s) call of tools corresponding to that

[1:13:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4389s) "like models do." So we

[1:13:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4390s) we call this a distributed call

[1:13:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4392s) tools. This is one of the things that

[1:13:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4394s) makes work faster and more accurate with

[1:13:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4395s) from the point of view of using these tools

[1:13:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4397s) . So, all of this came together to

[1:13:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4399s) instead of creating a separate agent

[1:13:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4401s) build this functionality into the set

[1:13:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4403s) work processes—rigid or flexible,

[1:13:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4406s) depending on the filters you need, —and

[1:13:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4408s) use Codex as the main working

[1:13:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4410s) area for this. In fact, we take

[1:13:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4413s) old, manually configured application,

[1:13:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4415s) who used large language models

[1:13:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4418s) in some nodes and traditional

[1:13:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4419s) deterministic workflow in others,

[1:13:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4422s) and combine them into a native one for Codex

[1:13:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4424s) a process that simplifies the user's work:

[1:13:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4426s) you don't need to go to this anymore

[1:13:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4428s) separate site. This becomes local

[1:13:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4430s) generated. It gets personalized.

[1:13:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4432s) according to their needs. They can

[1:13:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4434s) just add, "Hey, here are the data sources,

[1:13:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4435s) which I want to use instead."

[1:13:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4437s) These are types of checks, for example, barriers

[1:13:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4439s) quality. So, it stopped being "black"

[1:14:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4441s) "box" and a predefined

[1:14:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4443s) application, turning into an evolutionary

[1:14:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4445s) a process that logically coexists with

[1:14:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4446s) user.

[1:14:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4449s) Thank you, Don.

[1:14:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4450s) Yes. So, this is something that could be a little

[1:14:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4453s) add for developers here.

[1:14:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4456s) I think that "under the hood" we now have

[1:14:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4459s) Codex harness, which is essentially

[1:14:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4460s) open source.

[1:14:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4464s) So, this is also a significant competitive advantage.

[1:14:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4467s) advantage, I believe, because you have

[1:14:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4469s) full access to the source code, which

[1:14:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4471s) runs everything I showed here today

[1:14:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4473s) . And then you run it on the desktop

[1:14:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4476s) table. This is one of the options with a controlled

[1:14:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4479s) product, and that's what you see here.

[1:14:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4483s) In fact, you can also take this and

[1:14:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4485s) transfer using SDK to the server on

[1:14:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4488s) App Engine.

[1:14:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4491s) There is a way to do this programmatically.

[1:14:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4494s) And, of course, you have the same opportunities

[1:14:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4496s) developer.

[1:14:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4499s) You see, you have plugins that essentially

[1:15:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4502s) , become a way to connect to

[1:15:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4504s) the existing system, which is now very

[1:15:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4506s) popularly.

[1:15:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4510s) And that gives you a few options when you

[1:15:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4512s) you have such a research workflow

[1:15:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4514s) , which my colleague just mentioned,

[1:15:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4516s) it is possible to create it very quickly.

[1:15:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4518s) Maybe a business technologist, a quant or

[1:15:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4521s) the researcher will be able to do this. But then you

[1:15:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4523s) want to scale this to a program on

[1:15:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4525s) server side. What are you doing? How are you?

[1:15:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4528s) you can take the work you've already done and

[1:15:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4530s) scale it for the team. So, you

[1:15:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4532s) you can also get this using Codex SDK

[1:15:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4535s) connect to the real interface if

[1:15:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4537s) this is necessary, isn't it?

[1:15:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4539s) But you don't... It's just a way of thinking. It

[1:15:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4542s) becomes very flexible when you can

[1:15:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4544s) reuse your work and

[1:15:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4546s) where will you place it in

[1:15:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4547s) in the future, right? Yes.

[1:15:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4554s) So, let's quickly check how much

[1:15:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4556s) We have time. 40 minutes, yes. Perhaps,

[1:15:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4559s) 50 minutes for questions at the end or

[1:16:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4561s) Yes, it is possible. We will have about 50

[1:16:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4564s) minutes for questions and answers at the end

[1:16:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4566s) , but feel free to interrupt us at any time

[1:16:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4569s) -what time is it?

[1:16:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4573s) So, I'll continue to talk a little about

[1:16:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4575s) Codex to give you a little more

[1:16:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4577s) context about why we think this

[1:16:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4579s) a great product.

[1:16:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4581s) And, as we mentioned, the basis of the Codex,

[1:16:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4584s) i.e. uprunner, is completely open

[1:16:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4586s) source code.

[1:16:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4591s) So even to deploy this

[1:16:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4593s) product at the enterprise level this gives

[1:16:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4595s) lots of transparency to see clearly:

[1:16:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4598s) of course, your own app is proprietary

[1:16:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4600s) , but the engine that runs behind it is

[1:16:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4602s) completely open.

[1:16:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4606s) So it's not only transparent, but also completely

[1:16:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4608s) customizable.

[1:16:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4612s) So I think it's really interesting

[1:16:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4614s) aspect.

[1:16:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4617s) So, at the enterprise level, you can

[1:16:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4619s) choose the uprunner settings that will be

[1:17:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4621s) use your own interface

[1:17:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4623s) according to your needs.

[1:17:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4625s) If you need any level

[1:17:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4627s) settings, it is available.

[1:17:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4631s) In addition, we believe that compared to

[1:17:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4633s) other types of interfaces, our models

[1:17:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4635s) literally train with this

[1:17:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4638s) tools.

[1:17:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4640s) So, for our model,

[1:17:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4642s) many stages of preliminary and final

[1:17:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4645s) learning, but when we talk about

[1:17:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4647s) final training, it is based on

[1:17:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4649s) our own tools.

[1:17:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4654s) So our models are mostly learning

[1:17:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4656s) work with many tools, but

[1:17:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4658s) this is where they show the highest

[1:17:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4660s) performance for a simple reason: under

[1:17:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4662s) during the learning phase they are already familiar with

[1:17:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4664s) this codebase.

[1:17:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4673s) Especially lately, if we

[1:17:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4675s) let's remember the possibilities of multi-agents

[1:17:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4677s) or what Don talked about, a lot

[1:17:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4679s) workflows, the most advanced of

[1:18:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4682s) they use subagents, and the agent

[1:18:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4684s) decides when a subagent is needed, and

[1:18:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4686s) when not.

[1:18:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4695s) Of course, if you switch to a mode that

[1:18:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4697s) we call "ultra", we force

[1:18:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4699s) use the agent more often

[1:18:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4701s) subagents' capabilities, and working with

[1:18:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4703s) the subagent code provides significantly

[1:18:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4705s) better performance.

[1:18:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4712s) And on top of that, in the background

[1:18:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4714s) Compression mode is running.

[1:18:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4720s) Compression mode is essentially a secondary

[1:18:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4723s) a model that takes the context of the current

[1:18:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4725s) conversations and constantly squeezes it to

[1:18:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4727s) remove redundancy and keep the size within

[1:18:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4730s) within limits, and this is a very high-quality model.

[1:19:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4740s) We noticed, especially in research

[1:19:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4742s) tasks where you need extremely

[1:19:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4744s) long context what compression mode

[1:19:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4746s) allowed us to solve problems that

[1:19:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4748s) required a much larger volume

[1:19:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4750s) context.

[1:19:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4755s) This allowed us to be extremely efficient

[1:19:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4757s) to perform tasks that we call

[1:19:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4759s) long-term.

[1:19:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4762s) As a team, we often work with hedge funds.

[1:19:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4764s) funds and firms engaged in property

[1:19:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4766s) trading.

[1:19:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4772s) As an example that I have experienced myself: I

[1:19:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4774s) took one of the articles from quantitative

[1:19:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4777s) finance about Brownian motion and

[1:19:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4779s) options pricing, and literally

[1:19:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4781s) asked for a model—not even an Astro, but

[1:19:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4783s) version 5.5 a few months ago—

[1:19:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4785s) recreate the code base from this article.

[1:19:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4792s) So, essentially, recreate the codebase

[1:19:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4794s) articles, generate a numerical result

[1:19:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4796s) and try to see if it matches

[1:19:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4799s) with what is mentioned in the article.

[1:20:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4802s) So there's this task that, you know, I think

[1:20:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4805s) , was performed for several hours, and which was

[1:20:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4807s) would be impossible to run without the mode

[1:20:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4809s) compression.

[1:20:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4812s) And then, like, 2 hours later I was

[1:20:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4815s) a little impressed that the model was able to

[1:20:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4817s) completely recreate this, you know, very

[1:20:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4819s) code saturated with quantitative indicators

[1:20:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4822s) base with, say, a simulation

[1:20:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4824s) stochastic Brownian motion and

[1:20:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4826s) by calibrating the models in addition; and this

[1:20:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4828s) was, in fact, a task that without the regime

[1:20:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4831s) compression is not feasible. And mode

[1:20:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4833s) compression is already available first as well

[1:20:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4835s) through the API, but you know, if you

[1:20:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4837s) use the Codex harness, it

[1:20:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4839s) available in harness, and you don't even have to

[1:20:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4840s) need to think about it.

[1:20:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4843s) Yes, I mean I'm running

[1:20:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4845s) in parallel. I think this is a hint, some

[1:20:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4848s) people have already done it. Yes, he works,

[1:20:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4850s) creating, in fact, a plugin extension

[1:20:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4853s) right now.

[1:20:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4855s) Yes, that's basically what you see on the right.

[1:20:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4857s) the way in which in this example my

[1:20:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4859s) customized user interface can

[1:21:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4862s) be connected to the surface.

[1:21:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4865s) So this is just a small example, but yeah,

[1:21:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4867s) it's pretty easy, of course, just a hint

[1:21:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4870s) .

[1:21:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4870s) But yes, it shows you some options

[1:21:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4873s) interface, and if you look at our

[1:21:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4875s) website, there's a lot more there

[1:21:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4877s) examples of how our partners, such as

[1:21:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4879s) Figma or Adobe, created slightly more complex

[1:21:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4881s) apps where you can do more complex things

[1:21:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4884s) . But I really think, I mean, I

[1:21:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4886s) I have said many times that I think that

[1:21:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4888s) This is probably what it's heading towards.

[1:21:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4890s) industry. Now think about it more.

[1:21:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4893s) : you have an AI platform and you want to

[1:21:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4895s) connect to the app, and then this, by

[1:21:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4897s) In essence, the market and enterprises have their own

[1:21:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4899s) own inside.

[1:21:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4901s) And I think that's my personal opinion, of course.

[1:21:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4904s) , but yes, I see a strong similarity here. AND

[1:21:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4907s) yes, this is, I think, the developer; you must

[1:21:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4909s) just take it as an opportunity, not

[1:21:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4911s) as something that can be used.

[1:21:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4915s) Yeah, and I would say that when I joined

[1:21:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4918s) to OpenAI, many people still

[1:22:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4920s) used Codex CLI.

[1:22:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4924s) I was like, you know, most people with

[1:22:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4926s) engineering background, we are all used to

[1:22:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4929s) CLI, and I see that the percentage is extremely

[1:22:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4931s) technical people who are switching to

[1:22:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4933s) own application and almost never

[1:22:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4935s) leaving Codex as an add-on, everything becomes

[1:22:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4937s) bigger and bigger.

[1:22:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4942s) So, as a former quantum, I still

[1:22:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4944s) I'm doing some experiments, and if

[1:22:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4946s) I always had it open before.

[1:22:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4949s) Jupiter notebook, now it is already there

[1:22:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4951s) part of the Codex.

[1:22:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4957s) So, we believe, and this is one of

[1:22:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4958s) messages we want to convey:

[1:22:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4961s) even for a Codex technician—it's

[1:22:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4963s) an amazing environment where there is everything you need

[1:22:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4966s) .

[1:22:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4972s) And you, as a technical specialist, can

[1:22:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4974s) create plugins and provide them to your

[1:22:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4976s) knowledgeable but non-technical users

[1:22:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4978s) .

[1:23:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4983s) The process of implementing these plugins in

[1:23:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4985s) work can be quite optimized

[1:23:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4987s) thanks to corporate control, and you

[1:23:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4989s) can provide truly powerful

[1:23:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4991s) tools to end users for

[1:23:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4993s) weeks, not months.

[1:23:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5002s) I believe that, especially for financial

[1:23:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5004s) services where you need constant

[1:23:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5006s) load data when appears

[1:23:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5008s) new source, does anything need to be changed,

[1:23:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5011s) whether to perform a new analysis, this new

[1:23:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5013s) an AI-based approach will help

[1:23:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5015s) maximize productivity the way we

[1:23:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5017s) seen starting in 2025 in

[1:23:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5019s) programming—the same famous 10-

[1:23:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5022s) multiple, 100-fold increase

[1:23:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5023s) productivity in engineering.

[1:23:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5031s) Our goal is to provide the same acceleration

[1:23:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5033s) for knowledge workers.

[1:23:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5036s) This, in fact, is the vision behind

[1:23:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5039s) product.

[1:24:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5042s) And financial services thanks to this

[1:24:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5044s) combining artifacts, data sources and

[1:24:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5046s) technical analysis looks perfect

[1:24:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5048s) option for starting.

[1:24:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5052s) But, I apologize.

[1:24:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5055s) Do you have any questions about topics we have already covered?

[1:24:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5057s) considered?

[1:24:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5063s) I know we're here to discuss

[1:24:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5065s) financial services, but do you see

[1:24:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5067s) similar in other industries? Is it financial?

[1:24:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5069s) services because the data there is enough

[1:24:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5071s) structured? You create an ontological

[1:24:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5074s) layer. Is OpenAI doing something similar for

[1:24:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5076s) health care? Is something happening?

[1:24:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5078s) other? Therefore.

[1:24:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5080s) The answer is yes.

[1:24:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5084s) We work on different verticals, and

[1:24:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5087s) healthcare, of course,

[1:24:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5092s) is one of the most important priorities,

[1:24:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5094s) also in terms of research.

[1:24:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5096s) I would say that another area that we

[1:24:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5098s) We encounter a lot of it — it's cybersecurity.

[1:25:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5101s) Cybersecurity is changing dramatically.

[1:25:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5105s) In this area, for example, in addition to

[1:25:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5107s) specialized model for cybersecurity

[1:25:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5110s) , which we offer—our offer

[1:25:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5112s) called Daybreak, —we also, in essence

[1:25:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5114s) , we are developing the plugin again.

[1:25:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5117s) And again, this is the same concept that

[1:25:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5120s) and before.

[1:25:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5121s) The idea is that your team with

[1:25:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5123s) cybersecurity doesn't have to be experts

[1:25:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5126s) from artificial intelligence, at least on

[1:25:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5128s) initial stage.

[1:25:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5129s) So, we're essentially giving you a model.

[1:25:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5135s) We provide you with tools that are open

[1:25:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5137s) tools for work, but we also

[1:25:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5140s) we provide a plugin, for example, for

[1:25:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5141s) static asset scanning, which

[1:25:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5144s) tuned to our model.

[1:25:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5149s) This will help you create a baseline,

[1:25:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5151s) so you can use it right away

[1:25:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5154s) ready-made models, tools and plugins

[1:25:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5156s) to start everything.

[1:26:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5160s) We actually have a library of skills,

[1:26:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5162s) that helps you do everything from scanning

[1:26:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5164s) until final review and correction, but

[1:26:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5167s) we focus a little more on

[1:26:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5169s) financial services, although our focus is

[1:26:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5171s) different industries.

[1:26:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5175s) Of course, there are industries where profit is easy to come by.

[1:26:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5177s) measure: finance, cybersecurity, security

[1:26:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5180s) health, research, and there are others where

[1:26:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5182s) the usage options are not that wide.

[1:26:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5187s) Therefore,

[1:26:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5190s) So, I think this is

[1:26:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5194s) for platform developers—it's more than

[1:26:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5196s) on me, digital businesses or

[1:26:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5198s) software providers,

[1:26:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5201s) who usually look for new ways to

[1:26:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5203s) dissemination of their solutions.

[1:26:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5207s) So I think our team is working with

[1:26:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5209s) there are much more of them.

[1:26:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5212s) I think we're here because we're in New York, and

[1:26:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5214s) there are also financial services here.

[1:26:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5217s) I think you're right, structured data itself

[1:26:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5219s) are suitable for use in themselves

[1:27:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5221s) artificial intelligence.

[1:27:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5225s) AI works with their data, and in my opinion

[1:27:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5228s) , it is always difficult to deal with these "data in

[1:27:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5230s) "isolation" in such companies.

[1:27:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5234s) And then how can you effectively

[1:27:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5236s) use all this data in context

[1:27:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5238s) , which does not exist? There are no ontologies.

[1:27:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5241s) etc.

[1:27:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5244s) So sometimes the strategy I see is

[1:27:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5246s) is that it will take a lot

[1:27:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5248s) a lot of time to create a proper

[1:27:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5250s) context level for all data stores.

[1:27:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5255s) But why not start from the bottom up?

[1:27:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5257s) Bottom-up—this is when you have a strategy

[1:27:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5260s) GPT works with Codex.

[1:27:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5264s) And then the teams gradually connect

[1:27:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5267s) to all data sources and build this

[1:27:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5269s) context, while modernizing

[1:27:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5271s) data architecture.

[1:27:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5275s) This would be extremely effective. I think

[1:27:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5277s) we have the luxury of having it internally. IN

[1:28:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5280s) We don't have this outdated legacy. We

[1:28:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5282s) we already have this level of context. It

[1:28:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5284s) It's just amazing what you can do,

[1:28:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5286s) when the potential of the workforce is truly

[1:28:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5288s) open. But you need to move from below

[1:28:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5290s) upwards, because the approach is top-down

[1:28:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5292s) will take more time to scale.

[1:28:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5295s) Yes.

[1:28:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5300s) So, let's see. Call Matthew and

[1:28:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5302s) Zaku.

[1:28:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5303s) Yes, it seems that this is my screensaver.

[1:28:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5307s) I want to discuss something else.

[1:28:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5313s) I think it's also worth mentioning that when you

[1:28:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5315s) install any extension in Codex

[1:28:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5318s) , some code can be executed on

[1:28:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5320s) another person's desktop.

[1:28:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5327s) Or on the website.

[1:28:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5330s) We discussed this a lot with our

[1:28:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5332s) security teams: how to make sure that

[1:28:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5334s) won't they cause harm?

[1:28:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5338s) There should be a certain "sandbox", and I think

[1:29:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5341s) that the concept of isolation in the Codex is very

[1:29:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5343s) an important aspect to understand,

[1:29:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5346s) because this is another advantage of this path.

[1:29:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5350s) Let's say you have a plugin that needs

[1:29:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5353s) execute the tool call.

[1:29:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5355s) You want to restrict the tool's access to

[1:29:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5358s) certain URLs.

[1:29:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5359s) You should always have a multi-level

[1:29:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5362s) CONTROL. You need to start with

[1:29:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5363s) corporate controls,

[1:29:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5365s) network security and end-to-end inspection

[1:29:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5368s) points. But when the application runs on

[1:29:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5370s) platform, you also want to have the means

[1:29:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5372s) control within it.

[1:29:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5374s) So, Codex provides policy files

[1:29:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5377s) configurations, TOML files that are managed

[1:29:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5379s) on the side of the enterprise administrator and

[1:29:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5382s) deployed during implementation.

[1:29:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5387s) And you can decide, "Okay, I don't want to

[1:29:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5389s) to allow this plugin to access a specific

[1:29:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5391s) website.

[1:29:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5394s) I want to be sure that certain actions

[1:29:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5396s) only run in my workspace

[1:29:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5398s) spacious.

[1:29:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5399s) Let's say a team is developing something in

[1:30:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5401s) within the enterprise, but this should not

[1:30:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5404s) have access, for example, to sensitive

[1:30:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5406s) project folders.

[1:30:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5408s) So, with the help of policies in Codex, you can

[1:30:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5411s) restrict corporate rights.

[1:30:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5416s) Determine what is locally accessible

[1:30:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5419s) via the Codex client, as well as which networks,

[1:30:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5421s) URLs and even plugins can

[1:30:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5424s) use certain people.

[1:30:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5427s) I think this is important to understand. For

[1:30:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5429s) there is a complex and effective

[1:30:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5432s) mechanism.

[1:30:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5433s) This is essentially a configuration that you can

[1:30:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5436s) see here.

[1:30:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5439s) This is a user-configurable

[1:30:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5441s) terminal. Your corporate

[1:30:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5442s) the administrator determines this once per

[1:30:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5444s) enterprise levels for different teams.

[1:30:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5446s) It sends these settings, which

[1:30:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5448s) include everything I just mentioned.

[1:30:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5450s) Safety isolation.

[1:30:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5453s) Because this will then run everything that is placed

[1:30:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5455s) inside the Codex client, it

[1:30:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5457s) will work in it.

[1:30:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5459s) You will still have to defend

[1:31:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5461s) tools etc. on their own, but I think

[1:31:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5464s) , this is already an advantage of this architecture.

[1:31:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5468s) And the next thing—we all know that's pretty

[1:31:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5471s) difficult, I mean, the security team

[1:31:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5473s) goes to extreme measures.

[1:31:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5477s) We want to protect, we want

[1:31:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5479s) perform any confidential

[1:31:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5481s) operations, but the business user says: "

[1:31:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5483s) "Hey, that's so slow." I always

[1:31:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5486s) have to click. "Please,

[1:31:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5488s) go to this plugin, go to

[1:31:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5490s) this site". And I think there is also

[1:31:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5492s) uniqueness, we have a kind of way

[1:31:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5495s) implementing protection, right?

[1:31:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5498s) And we can automate a little bit

[1:31:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5500s) decision-making when something is critical, and

[1:31:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5502s) when not.

[1:31:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5505s) This is a kind of way to make sure that

[1:31:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5507s) when using Codex settings you

[1:31:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5509s) balance between what politics demands

[1:31:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5512s) security, and what is convenient for the final

[1:31:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5514s) user.

[1:31:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5517s) This helps to understand when the data

[1:31:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5519s) confidential, and if not, is it necessary

[1:32:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5521s) I need to go to this working folder.

[1:32:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5523s) area or not. So I think it's very

[1:32:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5526s) important to highlight because it's something you also

[1:32:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5528s) must explain to their users,

[1:32:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5530s) when creating plugins.

[1:32:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5533s) No, yes.

[1:32:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5535s) In almost every large institution with which

[1:32:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5537s) we work, they set up theirs

[1:32:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5539s) configuration for approvals and permissions.

[1:32:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5545s) This is really the first thing we usually do

[1:32:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5547s) discuss when we talk about

[1:32:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5549s) deploying Codex in the enterprise, and here

[1:32:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5551s) again there is a lot of flexibility.

[1:32:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5555s) Any more questions?

[1:32:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5558s) So, maybe we're almost done.

[1:32:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5562s) Perhaps we should invite our guests and

[1:32:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5564s) ask a few questions: Cohen, Matthew and

[1:32:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5567s) Zach seems to be the leader, right?

[1:32:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5575s) Like any good technological

[1:32:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5577s) company, we brought some souvenirs.

[1:33:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5582s) Therefore,

[1:33:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5586s) should we say that Cohen, Matthew and Zach—

[1:33:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5589s) these are the ones who will get a great bottle for

[1:33:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5591s) water with the OpenAI brand?

[1:33:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5597s) Picked it up from the office this morning. Therefore,

[1:33:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5600s) If you're still here, let's see. IN

[1:33:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5603s) otherwise, therefore,

[1:33:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5605s) Let's start, for example,

[1:33:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5608s) Cohen? I don't know how this is correct.

[1:33:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5610s) pronounced.

[1:33:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5611s) Yes, come on. Come up here, okay?

[1:33:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5615s) Yes, thank you for answering the question. It

[1:33:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5618s) It wasn't too difficult, right? So, yes

[1:33:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5621s) , maybe

[1:33:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5622s) Tell me about yourself.

[1:33:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5624s) Yes, my name is Cohen.

[1:33:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5627s) I work as a data engineer in the Netherlands.

[1:33:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5631s) So, it's quite interesting to see, uh, also

[1:33:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5634s) the local audience. I was on AI

[1:33:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5636s) Engineering in London. Um, and also with

[1:33:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5639s) I look forward to meeting many

[1:34:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5641s) people here to see, uh, what they

[1:34:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5643s) do, especially in the financial industry

[1:34:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5645s) . Um, this is all about me. Yes?

[1:34:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5647s) And you're from this company, uh,

[1:34:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5650s) Xebia.

[1:34:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5651s) Xebia Data. So, we're a consulting, uh

[1:34:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5654s) , the company, uh, we work on everything

[1:34:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5656s) world, and I work, uh, in

[1:34:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5658s) Dutch, er, branch.

[1:34:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5662s) Thank you. Thank you.

[1:34:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5665s) Now we have Matthew.

[1:34:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5670s) Yes, great. Yes, now you must too

[1:34:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5673s) to say something.

[1:34:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5673s) I hope this was fun.

[1:34:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5675s) Yes.

[1:34:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5676s) Yes.

[1:34:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5676s) Uh, thank you. It was, it was wonderful. THERE ARE-

[1:34:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5679s) Hey, my name is Matt. Uh, company

[1:34:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5681s) x1wealth.com, and we build AI-

[1:34:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5682s) oriented family office. So, I

[1:34:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5685s) I work a lot with Codex, and it's very

[1:34:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5686s) helps.

[1:34:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5687s) Yes.

[1:34:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5687s) Thank you.

[1:34:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5688s) I think so, AI-oriented—that sounds

[1:34:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5691s) very good I think I would invest in that.

[1:34:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5694s) Yes, cool. Yes, thank you.

[1:34:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5696s) The last one—Deut.

[1:34:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5699s) Good.

[1:35:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5701s) Yes, come on.

[1:35:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5705s) Greetings. So, I'm Zach. I am the founder of The

[1:35:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5707s) Clockworks and create applications that

[1:35:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5709s) use AI to research new

[1:35:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5712s) things we couldn't do before. I

[1:35:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5714s) I focus mainly on development topics

[1:35:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5717s) , but, you know, I'm always looking for new

[1:35:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5719s) opportunities.

[1:35:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5720s) Thank you.

[1:35:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5721s) I have, yes, okay. Yes. I have an idea.

[1:35:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5724s) . So, I think your homework is—

[1:35:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5727s) create a plugin extension and deploy

[1:35:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5729s) its on the marketplace or as a plugin, yes

[1:35:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5732s) ?

[1:35:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5737s) Oh, you're already there, right? Therefore,

[1:35:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5740s) x1wealth

[1:35:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5744s) Yes, he, he understands everything.

[1:35:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5746s) Yeah, so, yeah, that's another thing.

[1:35:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5751s) I mean, with all these

[1:35:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5752s) marketplaces, it takes a little

[1:35:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5755s) work to apply because

[1:35:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5757s) this may also affect external customers,

[1:36:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5760s) but I think if you see it that way, it's

[1:36:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5762s) signal for people—this is perhaps an external

[1:36:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5765s) B2B marketplace.

[1:36:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5768s) But also don't forget about the internal

[1:36:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5770s) corporate system, this is the same

[1:36:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5772s) concept, yes.

[1:36:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5775s) But I think it's really a signal when

[1:36:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5777s) you should use

[1:36:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5779s) distribution channels that are currently

[1:36:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5781s) emerging, i.e. new channels

[1:36:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5783s) distribution, and you saw it with App

[1:36:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5785s) Store, and how viral it is now.

[1:36:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5788s) So, basically, consider this another

[1:36:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5790s) way, and of course, checking SM if

[1:36:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5792s) It's public, yes. So, we just

[1:36:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5794s) need to make sure everything works

[1:36:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5796s) smoothly, safely and reliably, and that we

[1:36:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5798s) now in the correct category. But yes, I am

[1:36:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5800s) I think it's just a signal.

[1:36:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5802s) That's what I wanted to say, yes.

[1:36:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5804s) Yes, I know, I think that's it.

[1:36:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5806s) The beauty of an open source codebase is

[1:36:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5808s) source code. And I would say that again

[1:36:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5811s) yes, today we cover more,

[1:36:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5813s) you know, the surface of our product, like

[1:36:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5816s) Codex, ChatGPT.

[1:37:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5820s) But we also interact a lot with

[1:37:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5822s) clients, working on special

[1:37:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5824s) cases, in particular over cases

[1:37:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5826s) using the API, which is of course a bit

[1:37:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5828s) more complicated, and there are many topics for

[1:37:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5830s) discussion, but a significant part of our

[1:37:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5832s) work is to understand

[1:37:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5834s) use cases and help

[1:37:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5836s) implement best practices, because

[1:37:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5839s) we are the first internal users

[1:37:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5841s) our product.

[1:37:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5845s) We communicate with many people in

[1:37:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5847s) industry. So, for example, as we already

[1:37:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5849s) mentioned, this magic word is "assessment" (

[1:37:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5852s) eval), I think is the most important.

[1:37:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5855s) So we literally help customers

[1:37:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5857s) understand how to develop a proper assessment

[1:37:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5859s) for a specific problem.

[1:37:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5862s) And the question is always, what do we

[1:37:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5864s) we are talking about increasing productivity,

[1:37:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5866s) but especially when you want to go out on

[1:37:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5869s) production stage, you need to have

[1:37:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5871s) a way to measure real performance

[1:37:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5873s) your product.

[1:37:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5875s) And again, when we talk about

[1:37:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5877s) a specific chat thread, there are no need

[1:37:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5879s) assessments, but when you create your own

[1:38:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5881s) agents, they become necessary.

[1:38:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5886s) And, for example, best practices regarding

[1:38:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5888s) understanding how to create these assessments

[1:38:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5891s) based on the evaluation data set, why

[1:38:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5893s) performance does not meet requirements,

[1:38:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5896s) or that perhaps it's not the model itself that's the problem

[1:38:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5898s) , but in how to improve the system,

[1:38:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5900s) prompting or communication with

[1:38:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5902s) tools.

[1:38:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5905s) And sometimes it's also a limitation of the model itself

[1:38:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5907s) . Therefore, we also turn to

[1:38:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5910s) researchers to understand where our

[1:38:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5912s) models reach their limits, where they

[1:38:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5914s) behave unexpectedly even in

[1:38:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5916s) ideal environment.

[1:38:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5918s) And ideally, thanks to your feedback, in

[1:38:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5920s) in the next iteration of the model it will be

[1:38:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5922s) corrected or improved. We have

[1:38:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5925s) question. Sorry.

[1:38:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5927s) Yes, on the topic of evaluations (evals), I

[1:38:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5929s) It was very interesting to know what yours is.

[1:38:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5931s) perspective on market maturity regarding

[1:38:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5933s) implementation of assessment, understanding

[1:38:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5936s) fundamental principles and transition to

[1:38:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5938s) more complex topics, such as automated

[1:39:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5940s) research workflows.

[1:39:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5943s) Okay, this is a pretty serious question.

[1:39:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5946s) Let's start with the assessment.

[1:39:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5949s) I would say that evaluation is usually

[1:39:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5951s) the biggest obstacle, because the creation

[1:39:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5954s) tests—this is my personal opinion, I would be happy

[1:39:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5957s) Hearing others is not an easy task.

[1:39:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5963s) The difficulty lies in the fact that the assessment

[1:39:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5965s) usually includes a lot of business knowledge from

[1:39:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5968s) specialized experts (SME), although

[1:39:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5970s) engineers create them.

[1:39:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5975s) Therefore, unite the team where they work

[1:39:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5978s) specialized specialists and engineers for

[1:39:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5979s) creation of evaluation systems, —

[1:39:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5981s) The task is not easy.

[1:39:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5986s) Usually I always am, and here it is again

[1:39:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5988s) bias, because in the previous

[1:39:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5990s) I was a so-called quantitative worker.

[1:39:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5992s) researcher.

[1:39:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5994s) To me, when you create an agent,

[1:39:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5996s) evaluation is the most important element,

[1:39:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5998s) because it is impossible to optimize,

[1:40:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6000s) if you don't measure performance.

[1:40:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6003s) And it's not just one line like, "Oh, I

[1:40:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6006s) I have the evaluation results, they are on par

[1:40:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6009s) 70%".

[1:40:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6012s) It is also an understanding of basic metrics:

[1:40:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6015s) the first one where we show good results

[1:40:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6017s) , where—not very, and so on.

[1:40:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6023s) So I think that evaluation is really

[1:40:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6025s) the difficult part, but not because of technology

[1:40:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6028s) , but because you need to combine

[1:40:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6030s) expert knowledge with a data set,

[1:40:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6032s) created by engineers.

[1:40:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6037s) Regarding best practices, in short, I

[1:40:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6039s) I think it's always worth starting simple.

[1:40:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6042s) .

[1:40:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6043s) Don't start by using your LLM as a judge

[1:40:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6046s) .

[1:40:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6048s) Start simple and then add

[1:40:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6050s) complexity, otherwise you will never

[1:40:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6052s) understand whether the quality of your tests

[1:40:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6054s) Is the quality of the agent itself insufficient?

[1:40:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6059s) For me, this is a major best practice.

[1:41:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6064s) And another tip that we always

[1:41:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6066s) we suggest: when using LLM as

[1:41:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6068s) agent, do not ask to rate on a scale from

[1:41:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6071s) one to five—always ask the question "

[1:41:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6073s) "yes" or "no". This is much better.

[1:41:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6076s) Try to expand this scale from one

[1:41:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6078s) up to five for the options "true" or "false"

[1:41:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6080s) ". And you know, now, doing

[1:41:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6083s) promotion, we also release

[1:41:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6085s) Decision API so you can literally

[1:41:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6087s) to have a model that is, you know, pretty good

[1:41:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6090s) . It is based on Auto Luna and will prompt

[1:41:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6092s) you know, very quickly, very quickly, "

[1:41:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6094s) "yes" or "no", as well as the probability and

[1:41:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6096s) a kind of certainty in that answer.

[1:41:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6099s) So I see that there are many possibilities

[1:41:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6101s) applications for Decision API can be

[1:41:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6103s) precisely within the framework of evaluations (evals). Therefore,

[1:41:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6105s) this is the first part. The second part is

[1:41:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6107s) automated research. This is really

[1:41:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6110s) big topic. This, uh

[1:41:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6114s) something that I think is really significant

[1:41:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6116s) started this year.

[1:42:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6121s) And I saw that there were benchmarks: we achieved

[1:42:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6123s) very good programming capabilities

[1:42:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6126s) last year, but our models,

[1:42:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6128s) in particular, significantly improved mathematical

[1:42:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6131s) abilities recently, especially with Astra and

[1:42:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6133s) 5.6.

[1:42:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6137s) So I think that especially with quantum

[1:42:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6139s) perspectives, the intersection of programming and

[1:42:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6141s) mathematical research really allows

[1:42:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6143s) perform such more automated

[1:42:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6145s) research tasks. So I

[1:42:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6147s) I would suggest to you, for example, one of

[1:42:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6149s) such tasks—that's what I was referring to

[1:42:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6152s) previously: reproduction of scientific results

[1:42:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6154s) articles.

[1:42:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6157s) Of course, there are different levels of complexity,

[1:42:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6159s) but if you're thinking about a specific

[1:42:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6162s) the metric you want to improve for "

[1:42:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6164s) "hill climb" as we

[1:42:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6166s) in AI terms, we have what

[1:42:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6168s) called the "golden mode" in the Codex.

[1:42:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6174s) So, essentially, the golden mode means that

[1:42:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6177s) you must specify the goal, and it is better if

[1:42:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6179s) this goal will be a numerical indicator.

[1:43:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6184s) One simple example: imagine you have

[1:43:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6187s) agent, you have a data set for

[1:43:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6189s) evaluation, this benchmark.

[1:43:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6193s) Try calibrating the prompt to

[1:43:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6195s) maximize the performance of this

[1:43:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6197s) agent.

[1:43:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6201s) And what's happening there: the agent is looking at

[1:43:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6203s) target, and then continues to iterate over new ones

[1:43:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6205s) ideas and hypotheses.

[1:43:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6211s) Of course, this is the most simplified form of the regime.

[1:43:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6213s) automated research, but some

[1:43:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6215s) the results we have published, in particular

[1:43:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6218s) in mathematics earlier this year, were

[1:43:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6220s) created using the gold mode in

[1:43:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6222s) Codex.

[1:43:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6225s) In fact, for one of them we also

[1:43:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6227s) published a prompt.

[1:43:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6229s) So I think if you're interested in this area

[1:43:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6232s) , it is worth reading the public prompt, which

[1:43:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6234s) we gave the Codex to prove the theorem.

[1:43:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6237s) It was a hypothesis.

[1:43:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6239s) Um, uh, there's a slight difference here,

[1:44:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6241s) Sorry, for the mathematicians in the room.

[1:44:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6246s) Hmm, I think this is the most interesting, really very

[1:44:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6249s) interesting content because it's a little better

[1:44:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6251s) shows the logic of multi-agent approaches

[1:44:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6253s) and subagents.

[1:44:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6256s) So, um,

[1:44:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6260s) for example, when you execute

[1:44:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6262s) research assignment, you are most likely

[1:44:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6264s) , you will use the "ultra" mode

[1:44:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6266s) , and you will see how the agent automatically

[1:44:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6269s) involves subagents for parallel

[1:44:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6271s) conducting independent research, and

[1:44:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6273s) then merges them and iterates.

[1:44:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6277s) So I would say those are two very good ones.

[1:44:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6279s) entry points for assessment and

[1:44:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6281s) automatic research.

[1:44:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6284s) And I don't want to downplay the importance of "

[1:44:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6286s) "golden" regime, because even despite its

[1:44:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6288s) simplicity, we achieve significant results with it

[1:44:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6290s) results.

[1:44:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6292s) So it's quite powerful.

[1:44:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6295s) Yes, I fully support the "gold" regime

[1:44:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6298s) . Yes, I think, six months ago you

[1:45:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6300s) had to worry about prompts,

[1:45:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6303s) think about how to organize the work of agents

[1:45:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6305s) and all that other stuff. And now, I have to

[1:45:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6307s) mind you, you can move away from this. You

[1:45:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6310s) really need to think about what it is

[1:45:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6313s) agent's goal.

[1:45:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6315s) And I find that very powerful because

[1:45:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6318s) models will become more capable of doing

[1:45:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6320s) automatically what we used to—and partially

[1:45:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6322s) We still do it by hand.

[1:45:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6325s) And I strongly recommend the "golden"

[1:45:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6328s) regime.

[1:45:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6331s) Also today we demonstrated

[1:45:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6334s) plugin for financial services, but now

[1:45:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6336s) our team, including my colleague Dom,

[1:45:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6338s) focuses on working with clients in

[1:45:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6340s) in the field of finance and fintech, to

[1:45:42](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6342s) implement this on a large scale and

[1:45:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6344s) solve complex problems.

[1:45:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6349s) And in the evaluation conducted by our

[1:45:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6351s) team, there are many nuances. So we're going nowhere.

[1:45:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6353s) We don't disappear. We just want

[1:45:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6355s) make sure you scale

[1:45:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6356s) use our models on the platform

[1:45:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6358s) . Today was pretty simple, but...

[1:46:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6360s) contact us if you have

[1:46:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6362s) specific cases. We will be happy.

[1:46:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6364s) to cooperate with you.

[1:46:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6365s) Yes, I know. I really want to emphasize

[1:46:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6368s) because it's a two-way conversation.

[1:46:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6371s) So, if you're already working with OpenAI,

[1:46:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6373s) please refer to your account

[1:46:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6376s) -director.

[1:46:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6377s) Your feedback is more than welcome.

[1:46:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6380s) And that's also part of our job—

[1:46:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6382s) make sure it's not just pretty

[1:46:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6384s) presentation, but what is real

[1:46:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6386s) will work in your companies.

[1:46:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6389s) So, um, that's actually why we're here.

[1:46:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6393s) But, are there any more questions?

[1:46:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6397s) Good.

[1:46:38](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6398s) For one of the questions, I probably can

[1:46:40](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6400s) answer.

[1:46:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6401s) Thank you. Uh, one of the questions I have is

[1:46:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6404s) arose during the presentation: there was

[1:46:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6406s) many skills and plugins just for

[1:46:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6408s) reading.

[1:46:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6409s) I wonder if you saw any skills in

[1:46:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6412s) customers who actually record data

[1:46:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6414s) back into the system, or, let's say, into the system

[1:46:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6417s) accounting that the team works with?

[1:47:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6421s) That's a good question.

[1:47:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6423s) I think Dom has a better answer.

[1:47:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6425s) Yes, hello, I'll come over here. Yes, therefore,

[1:47:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6428s) regarding the structure of plugins and that

[1:47:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6431s) how they can and should interact with

[1:47:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6433s) large systems, there are several ways to do this

[1:47:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6435s) make.

[1:47:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6436s) And it depends on what exactly you

[1:47:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6438s) trying to create. So, if I

[1:47:21](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6441s) I am creating a full-fledged plugin extension, I

[1:47:23](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6443s) just using Codex as a system

[1:47:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6445s) orchestration, but these instruments and

[1:47:27](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6447s) processes may well store data in

[1:47:29](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6449s) remote locations and interact with others

[1:47:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6452s) systems. The easiest way to do it

[1:47:34](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6454s) is to interact with the local file system

[1:47:36](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6456s) system. So if you say, "Hey, in

[1:47:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6457s) within the scope of this project where I am developing it"

[1:47:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6459s) , it can write data back to these

[1:47:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6461s) processes. Or you give instructions to the model

[1:47:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6463s) so that the tool output is

[1:47:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6465s) such, and the skill set instruction— "

[1:47:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6467s) Hey, handle this in this place." You

[1:47:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6469s) you will see many stable structures within

[1:47:52](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6472s) Codex. He became very, very well.

[1:47:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6474s) cope with long context,

[1:47:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6475s) effectively using your local

[1:47:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6477s) file system as a kind of point

[1:47:59](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6479s) recording. However, once you

[1:48:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6481s) are you switching from a traditional plugin or

[1:48:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6482s) reading-only skill set to

[1:48:04](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6484s) full expansion operation, I can

[1:48:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6486s) order him to contact the central

[1:48:08](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6488s) server. I tell him to start the base.

[1:48:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6490s) data or at least connect to

[1:48:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6491s) a specific database as part of

[1:48:13](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6493s) tool. So, like many options

[1:48:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6495s) uses that we implement, we

[1:48:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6496s) just made one about proxy-

[1:48:18](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6498s) voting—the possibility of actually

[1:48:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6499s) find your proxy systems

[1:48:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6500s) votes, what exactly they received and

[1:48:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6502s) Where were they sent? So, although the plugin in

[1:48:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6504s) the work area was Codex, the actual

[1:48:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6506s) the source of the recording was this application server

[1:48:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6508s) , connected through one of the extensions

[1:48:30](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6510s) plugin. He says, "Hey, write this report down."

[1:48:32](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6512s) , write down the conclusions, write down the date, write down

[1:48:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6515s) sources back into this database". And this

[1:48:37](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6517s) the database powers their production system

[1:48:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6519s) applications that are outside the

[1:48:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6521s) Codex. So, an operating system with

[1:48:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6523s) The user's perspective is Codex.

[1:48:44](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6524s) Operating system from the point of view

[1:48:46](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6526s) administration and from a legal point of view

[1:48:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6527s) verification was outside the scope of Codex. Therefore,

[1:48:49](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6529s) MCP extensions allow you to be

[1:48:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6531s) flexible about where these are located

[1:48:53](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6533s) different operating systems and systems

[1:48:55](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6535s) accounting. You can submit this to Codex and

[1:48:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6538s) leave everything local, or create

[1:49:00](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6540s) tools and paths to specific MCPs,

[1:49:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6543s) easily configure permission sets,

[1:49:05](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6545s) simply adding MCP where needed.

[1:49:10](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6550s) Yes, I would say that another common

[1:49:12](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6552s) use case—this is especially true when

[1:49:15](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6555s) when you perform certain processes that

[1:49:17](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6557s) require an audit.

[1:49:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6560s) As he recalled about voting for

[1:49:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6562s) power of attorney.

[1:49:24](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6564s) To have standardized logs that we

[1:49:26](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6566s) often create, which would allow

[1:49:28](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6568s) for a technical user to understand not only

[1:49:31](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6571s) the result of the analysis, but also how it was

[1:49:33](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6573s) every decision was made and every action was taken

[1:49:35](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6575s) analysis stage.

[1:49:39](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6579s) For example, such an extended journal is something

[1:49:41](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6581s) that we often save, especially in

[1:49:43](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6583s) highly regulated industries such as

[1:49:45](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6585s) financial services where it is important to track

[1:49:48](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6588s) each solution to the data source and logic

[1:49:50](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6590s) , which is behind him.

[1:49:54](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6594s) This was critically important, and usually,

[1:49:56](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6596s) when we create an artifact like

[1:49:58](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6598s) result, we also keep logs,

[1:50:01](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6601s) which are fully auditable, as

[1:50:03](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6603s) part of the process and the plugin.

[1:50:07](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6607s) Maybe one more question?

[1:50:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6609s) I think we're almost done.

[1:50:11](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6611s) Yes, we have about 5 minutes left.

[1:50:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6614s) Yes, thank you. Thank you for taking the time.

[1:50:16](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6616s) with us. And please contact

[1:50:19](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6619s) us. And thank you. I wish you a good conference.

[1:50:22](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6622s) And thank you for coming. Thank you very much,

[1:50:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6625s) guys.
