AI Engineer Code 2025
AGI: The Path Forward
Read the talk
From Ada to Rust: Poolside’s Path Toward More Autonomous Software Work
An Ada database conversion shows how Poolside connects generated code, constrained agent access, executable checks, and feature iteration to its broader ambition for machine intelligence.
From a talk by Eiso Kant and Jason Warner
Before you start: Familiarity with command-line programs, build output, and basic database queries will help; no Ada experience is required.
Closing the gap starts with a codebase
What would it take to close the gap between language models and human intelligence? Poolside’s starting conviction is that next-token prediction needs to be paired with reinforcement learning. Jason Warner describes two and a half years of building models from scratch around that idea. Malibu Agent, the second generation, provides the model behind the demonstration that follows.
The concrete starting point is an Ada database. Warner connects the language to critical infrastructure and satellites, although the demonstration does not identify this database as an operational satellite system. Before changing anything, Eiso Kant asks the assistant what the codebase does. Inside Visual Studio Code, the assistant inspects the project and produces a structured overview. Understanding the existing program precedes the migration request.
Warner knows other languages better than Ada, so the next request is straightforward: convert the database to Rust. The interface resembles other coding agents, but Warner emphasizes that Poolside supplies its own model rather than using OpenAI or Anthropic. He places the work in the context of Poolside’s preceding year serving government and defense organizations, where mistakes and access decisions can have serious consequences.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Visible edits, bounded permissions
As the conversion proceeds, a middle pane streams the changes. This makes the agent’s work visible while it is happening, rather than presenting only a final answer. Visibility accompanies a separate requirement: the agent must operate within appropriate permissions. In sensitive environments, access to a data source cannot mean permission to do anything with it. Warner argues for tightly constrained operation; he does not describe a complete permission-enforcement architecture.
Kant reports approximately 1,152 generated lines for this conversion. A command appears to start testing the result, and the live diff view exposes the files the agent has created. These are intermediate artifacts: the conversion has produced code and reached a point where execution can begin.
Kant describes inference as fast, without supplying a throughput measurement, and later reports ten completed tasks. Warner immediately asks whether the result works. The agent has written build commands, and Kant inspects their output, which looks promising. A completed task counter and reassuring output are useful signals, but the next step is to run the application itself.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Run the database, then inspect its behavior
The agent’s summary is not the end of verification. Kant asks how to run the program, receives instructions for both the Ada and Rust versions, and launches Rust. Output briefly looks like a demo bug, but turns out to be warnings. Warner’s question about a possible unwrap is a joke and a question, not the identification of a confirmed defect.
The Rust program opens a REPL, an interactive loop that accepts commands and displays results. Kant checks a small but meaningful path through the database:
- Run Help to discover the available commands.
- Create a
userstable. - Insert a record.
- Execute
SELECT * FROM usersand inspect the returned record.
The record appears. This smoke test establishes that the demonstrated create, insert, and retrieve sequence works in the Rust program; it does not establish complete behavioral equivalence with the Ada implementation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turn a missing interaction into a feature
Using the database reveals the next requirement: pressing the up arrow does not recall earlier commands. Kant asks the agent to add that behavior. This is a narrower task than the language conversion, with an observable acceptance condition: after entering a command, the user should be able to retrieve it with the up arrow.
The agent identifies a package, and Kant compares the edits with the earlier version. It adds RustyLine, a Rust line-editing library, and changes the relevant files. The mechanism is to let a line editor handle keyboard input and maintain command history instead of treating terminal input as plain lines.
A minimal Rust example shows the integration point: read a command, add it to history, then dispatch it. Entering help records that command so the next prompt can recall it. This example uses the current RustyLine 18 API; the demonstration’s dependency version is unspecified.
rust
use rustyline::{error::ReadlineError, DefaultEditor};
fn main() -> rustyline::Result<()> {
let mut editor = DefaultEditor::new()?;
loop {
match editor.readline("db> ") {
Ok(line) => {
let command = line.trim();
if command.is_empty() {
continue;
}
editor.add_history_entry(command)?;
match command {
"help" => println!("Commands: help, quit"),
"quit" => break,
_ => println!("Unknown command: {command}"),
}
}
Err(ReadlineError::Interrupted | ReadlineError::Eof) => break,
Err(error) => return Err(error),
}
}
Ok(())
}
In the database, the existing query handler would occupy the dispatch step. The feature changes command entry without requiring a new database engine.
The updated build succeeds with warnings; warning cleanup is deferred. The agent writes a Bash script to test history and then a final demonstration script before returning its summary. Kant closes the loop himself: rebuild, rerun, enter Help, and press the up arrow. This time command recall works. The requested interaction is checked in the running program, after the edits and generated tests.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Beyond the coding interface
Kant next asks for a poem about the changes, shifting from software work to creative writing. He describes the models as useful for broader knowledge work, emotionally intelligent, and enjoyable for bedtime stories. Warner adds his own anecdote about model-written love letters. The poem request is part of the demonstration, though its output is not shown in the spoken walkthrough.
VS Code is one interface into the platform. Warner also describes web access and a downloadable agent for existing customers. He then turns to the next model generation: additional compute would support training, followed by a planned public release very early the following year—early 2026 in the recording’s November 2025 context. Distribution was planned through Poolside’s own API and Amazon Bedrock. These are announcements made during the talk, not established availability; the inspected AWS page still describes the offering as coming soon.
The intended consumers include both engineering assistants—such as Cursor, Windsurf, Cognition, and Replit—and applications outside coding, such as Harvey and Writer. Warner positions Poolside as an additional model supplier at the leading tier. That is the proposed market role, rather than a comparative benchmark demonstrated by the database exercise.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Scaling the model and the duration of work
Kant identifies compute as a previous constraint on capability. He reports more than 40,000 GB300s coming online to support further scaling in software development and other long-horizon knowledge work. The claim concerns incoming infrastructure, not a measured improvement in model performance.
Longer work also changes the interface problem. Kant says he and Warner already have agents performing tasks for hours, and forecasts tasks lasting days in the coming years. The asynchronous interaction shown in VS Code is one expression of that capability. As the duration and independence of work grow, he expects the surrounding interfaces to keep changing.
Poolside’s response is vertical integration: a multi-gigawatt campus under development in West Texas, teams building models, and the infrastructure to serve them. Kant treats the demonstrated interface as Poolside’s own expression of the underlying intelligence, while expecting application builders to create other ways to turn it into economically valuable work. Warner makes the physical commitment explicit: the company is putting shovels in the ground and hiring people to build the data-center campus.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An invitation to build on model checkpoints
Kant adds an exploratory invitation for companies doing reinforcement learning and fine-tuning. He names Qwen, Kimi, and MiniMax as model families others are building on, then proposes working with partners on Poolside checkpoints ranging from early training stages to current models. No particular versions or licensing terms are specified.
The partnership approach is not yet defined. This is an invitation to discuss collaboration, not an announced open-source release or an established checkpoint-access program. Its purpose is broader than consuming an inference API: Kant wants others to participate in building on the models themselves.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From early completion systems to Poolside
The informal Q&A turns to how the founders met. Warner recalls joining GitHub in 2017 with the ambition of turning a collaborative code host into an end-to-end development platform infused with intelligence. He describes the role retrospectively as CTO; contemporary GitHub material identified him as SVP of Technology. His account traces the platform’s expansion through Actions, Packages, Alerts, Notifications, Codespaces, and eventually Copilot, naming Nat Friedman and Oege de Moor among the collaborators.
In Warner’s recollection, Kant already had working code completion using LSTMs in 2017, before transformer architectures had fully landed. Warner tried to acquire Kant’s company, and Kant declined. The prolonged discussions about what neural networks could mean nevertheless became the basis of a friendship that continued over the following years.
Developments around ChatGPT and Anthropic in 2022 brought them back to the question of starting a company. Kant initially preferred learning to paint and sail, Warner jokes. Continued conversations eventually produced a different answer: they would build Poolside together.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Making raw capability useful
Kant closes by returning to the founding technical conviction. He says the company’s first 18 months concentrated on combining reinforcement learning with LLMs, a position that felt contrarian to him then and no longer does. He places the work in a progression from completions to chat to agents, with greater autonomy still ahead.
That progression requires more than increasingly capable models. It requires bringing those models into contact with real problems and building applications that make their capabilities useful. Kant calls the unfinished transition the “awkward teenage years ahead of AGI.” In that framing, application builders bridge the distance between raw intelligence and practical value.
The closing commitment is to broaden participation. Existing enterprise customers have already had access to the capabilities demonstrated here; Poolside wants others to be able to build with them too. The work ahead includes both improving the models and making that intelligence available in forms people can turn into useful software and services.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Rust line-editing library with command history, completion, and an example REPL integration.
AWS's overview of the planned Poolside offering, including Malibu and Point; the page labels availability as coming soon.
Further reading
Poolside explains how it builds and serves executable repository environments for reinforcement learning from code execution feedback.
- The finishing touchesArticle
Eiso Kant and Joe Rowell describe supervised fine-tuning and reinforcement-learning workflows inside Poolside's Model Factory.
- The carrier and the beaconArticle
An engineering account of Atlas, Poolside's inference system, and its integration with model evaluation.
The October 2025 announcement details the planned GPU cluster and the phased West Texas campus partnership.
Read the complete timestamped transcript
- 0:00
[on-hold music] How many people here know what Poolside is and does?
- 0:24
Anyone? Anyone? Yeah. [laughs] So let's talk about that real quickly. [laughs] Poolside exists to close the gap between models and human intelligence. That's literally it. That's what we're here to go do.
- 0:37
We're building our own models from scratch to do this. We're based on the idea two and a half years ago that we thought next-token prediction was an amazing technol- technological breakthrough, but it needed to be paired with reinforcement learning really to make that leap.
- 0:52
So that's what we've been doing for the past two and a half years. So we're on our second generation of models now, Malibu Agent, and instead of kind of like walking you through some slides and all that, we just thought maybe, I don't know, let's kind of show you what we're doing here.
- 1:07
So Eiso, are you there?
- 1:10
I got you, Jason.
- 1:11
So as I said, you were supposed to see him today, but there's...
- 1:16
I don't know. Our airline system kind of works sometimes maybe. So he's stuck in California. But, uh, we thought we'd just walk you kind of through some, um, some demos here today.
- 1:26
So what you're looking at here is a very modern programming language that the government uses to run all the world's critical infrastructure called Ada. Anyone familiar with Ada? Yes.
- 1:37
Yes. Okay, so everyone I saw put their hands up for Ada either has [REDACTED:physical_attribute] or [REDACTED:physical_attribute] like me. [laughs] So that should tell you what's going on here.
- 1:45
So Eiso, why don't we, uh, why don't we figure out what's going on with this code base here?
- 1:51
Well, let's start asking what the code base is about.
- 1:54
That's great. And what you're seeing here is obviously our assistant in, in Visual Studio Code backed by Poolside Agent, a model we trained from scratch using our proprietary techniques.
- 2:06
Um, and you can see what's going on here, kind of the stuff you expect from an agent. Uh, and obviously the form factors of all of these things are going to change a couple of times over the next couple of years, but, you know, people seem to like VS Code.
- 2:17
Uh, so we're gonna, you know, show you this demo here today. So you can see from this, it kind of went through, told you what this code base is all about.
- 2:24
But, um, you know, these things run on our satellites and, uh, I don't know anything about Ada, but I do know a lot about a couple of other programming languages.
- 2:33
So, uh, Eiso, what do we want to do here? Why don't we, uh, see what this thing might look like in Rust?
- 2:39
Let's do it. Let's ask it, convert this database to Rust.
- 2:47
So obviously you're gonna see what's going on here. Again, if you guys have used other tools, you're not gonna expect too much of the difference for what's happening here, except that, again, we're backed by our own model.
- 2:57
We're not using OpenAI, we're not using Anthropic. This is Poolside, and Poolside is, uh, a bottom and top stack that is right now, if no one's touched it, and I know no one in this room has touched this unless you work for a three-letter agency, a defense contractor, or you've sent missiles somewhere that we're not gonna talk
- 3:14
about in this session, um, 'cause that's where we're working. We're working in high consequence code environments for the last year inside the, the government and the, the defense sector, um, as you can see from this demo.
- 3:25
Um, so what you see here is, is kind of going through doing the conversions. What you see in the middle pane is something that we built to kind of show you as the streams come through all the different changes that are happening.
- 3:37
Um, one of the tricky parts about working on inside the defense sector and things like that is you can't have an agent that's just gonna run around and do stuff.
- 3:44
I mean, like, I can't walk into half of these buildings. You can't give an agent access to these data source and just say, "Hey, go nuts." You need to have the right permissions.
- 3:52
You got to actually really ratchet these things down to do things inside those environments that, you know, they feel comfortable with. So, uh, where are we on this now?
- 4:02
What is... Is it trying to fix itself yet?
- 4:05
Yes. So it's-- it cr- wrote about eleven hundred fifty-two lines of code, uh, and it just popped up a command, star and tested. [coughs]
- 4:15
Excuse me. Uh, so we see here all of the files on the left-hand side that it created. Uh, this is essentially our live diff view that's available.
- 4:24
Uh, and as we see, it's currently starting to actually test it out.
- 4:34
So this is the part where we just sit here and watch this for three minutes and I say nothing. No, what you see-
- 4:38
Well, the good thing is that this is a very fast inference.
- 4:41
Yes.
- 4:41
So eleven hundred lines of code.
- 4:43
Did it, uh-
- 4:44
Ten tasks completed.
- 4:45
Do we know if this works yet?
- 4:48
Well, let's have a look. So it actually wrote some build commands to test it,
- 4:53
and when we check out the output of those,
- 4:56
this actually looks pretty good.
- 4:58
Can we, uh-
- 4:58
Why don't we ask-
- 4:59
Can we verify that?
- 5:00
-it to run it. Let's go verify it. So of course, our agent came back and gave a summary of what it did, but let's just ask how to run this.
- 5:18
Okay. So I'm gonna go open up... So it says this is how I can run the Ada version, and this is how I can run the Rust version. Let's run the Rust version.
- 5:32
Perfect. Let's have a look at... Ooh, we might be hitting an actual bug.
- 5:38
An actual demo bug?
- 5:39
Let's have a look.
- 5:40
Let's see what happens.
- 5:41
Oh, no. No, no, just warnings.
- 5:43
Just warnings.
- 5:44
Sadly.
- 5:45
Do we have an unwrap in there that we need to take care of? I heard that those things are dangerous.
- 5:49
So right now there's a REPL. Uh, let's hit Help, see what we're able to do. So it looks like we have a set of commands. I'm gonna be lazy.
- 6:00
I'm gonna copy-paste these queries to create table users. Okay. So far, so good.
- 6:07
Let's insert a record. Okay, well let's find out if it actually did its job. So let's start from users.
- 6:16
Okay.
- 6:17
We've got a record here.
- 6:18
That's nice.
- 6:19
Now, now I want to actually... Uh, you see if I use the up arrow,
- 6:25
it doesn't actually allow me to cycle through commands. Let's ask it to add a feature.
- 6:30
Uh, allows me to use the up arrow to cycle through. I think it will understand my intent here.
- 6:42
The one thing we know about Eiso is he actually does know how to read and write, but he can't type. So all those errors that you're seeing in there, uh, yeah.
- 6:52
So it looks like the agent's identified a package that we can use.
- 6:57
Let's just quickly look here, compare this to version one.
- 7:02
And it looks like it's adding a library called Rusty Align and changing the files accordingly.
- 7:09
It's currently built it, and it looks like the build output is successful. There's some warnings. We'll ask it to clean those up later on, and it's now starting to test it.
- 7:25
Okay. Okay, apparently it works. It's gonna... It wrote itself a little bash script to test the history.
- 7:33
It's wrote itself a little final demo script.
- 7:36
So let's let it... Okay, so and it gave us the summary. Well, now how do I rerun this? I do kinda know that, though, so let's just-
- 7:46
You should know that. That was thirty seconds ago.
- 7:49
Let's build it and let's run it again. Okay, let's do a help. Let's... And oh, yeah, that's the up arrow. It works.
- 7:59
Very nice.
- 8:01
Now, our models aren't just capable coding agents. They're capable in lots of areas of knowledge work. They're also emotionally intelligent. They're fun. They're great to write bedtime stories with for the kids.
- 8:12
So I'm gonna ask it to write me a poem about all these changes, but that's just more for fun. [chuckles]
- 8:18
So as Eiso was saying, this is just an interface into our platform. There's other interfaces into it if you're inside one of those organizations that has adopted Poolside, so this is the coding interface into it, but we also have other plate-- ways in which you, you can interact with it, web as well as an agent that you
- 8:33
can download on your machine. But, um, yeah, we don't really tout the poem writing or the songwriting, though I did send this to my wife to see, and I have been sending her love letters written by Poolside, so I kinda hope that she did not enter this session to know exactly how I've been doing that for the
- 8:50
past six months. But, uh, yeah, so this is kinda Poolside. This is what we've been up to. Um, so as I said, Malibu Agent is as a second generation.
- 8:59
We've got a ton more compute coming online, and that's when we're train-training our next generation. That is be-gonna be the one that comes out publicly to everybody very early next year.
- 9:08
We're gonna have it behind our own API. It'll be on Amazon behind the Bedrock API. Anybody in the world who's building out any sort of, on a one side engineering assistance, like the Cursors, Windsurfs, Cognitions, Replits of the world, you can use ours.
- 9:22
Or if you use building out on any other side of the fence, the Harveys, the Writers, the whatever applications of the world, there's gonna be a fifth model out there that's gonna be at that level that you can, you can consume.
- 9:33
But we're dead set on doing this and bringing this out to everybody in the world and kind of advancing that state-of-the-art, and we're just gonna keep pushing that out.
- 9:39
So that's kinda who we are, um, and, uh, you can find, find out very little more at our website since we don't put much out there. [audience laughing]
- 9:47
But Eiso, anything else you wanna say before you, uh, try to go make your flight this time, please?
- 9:55
So I would say that it's been a pretty incredible journey for the last two and a half years of starting entirely from scratch and now building to a place where we see our models have grown up to become increasingly more intelligent.
- 10:07
And the kind of missing ingredient that we had was compute. And now that's unlocked for us and, and with a large number of over forty thousand GB300s coming online, we see how we can start scaling up some of those models, uh, to get even further, uh, in their, in their level of capabilities in software development and other
- 10:23
types of long-horizon knowledge work. What I think is exciting about this conference and this audience is of all the work that's happening of evolving the form factor, right? Right now what we looked at was this asynchronous way of operating with agents, but, you know, Jason, you and I, we have agents running that are doing tasks for, for
- 10:38
hours, and I think that in the near future we can see a world where they're able to start doing tasks in days in, in the coming years. And so I think the interface will continue to change.
- 10:47
Uh, we're really focused on the fundamentals, building intelligence and being able to scale up and serve it, and it's why we go full vertical. It's why we go from our multi-gigawatt campus in West Texas, where we're building out data centers to our team building out models, and the interface that you saw today is just our version of
- 11:02
an expression. But I think this audience is gonna do an incredible job at building lots of better versions of how to express using that intelligence, uh, into actually, you know, valuable, economically valuable work.
- 11:13
Couldn't have said it better. Can't wait to see what you guys build on this, uh, in the future when it's publicly available. And if anyone really does wanna build a data center campus, we are hiring for that.
- 11:22
Um, it is weird to be putting shovels in ground again like we did in the '90s and early 2000s, but that's what you gotta do to scale intelligence these days, so...
- 11:30
I would make one other non-scheduled statement.
- 11:33
Oh, no.
- 11:33
I don't know if you're gonna be okay with this one, Jason. Uh,
- 11:37
as, as our models are, are getting more capable, we'd love to also see who wants to build with them. Right now the, the vast majority of, uh, you know, companies that are doing additional reinforcement learning and fine-tuning on top of models are, are doing it on what I would consider right now the, you know, best-in-class open source
- 11:54
models, the QEMs and SkimEs and Minimaxes of the world. And, uh, we'd like to start figuring out of how we can, you know, partner with you with our, our models anywhere from any checkpoint early on to where we are today for you to be building closer together with us on top of things.
- 12:08
Uh, we haven't really figured out the approach to it yet, uh, but I think since we have this audience, it's, uh, it's not a bad place to put it out there.
- 12:14
And so definitely reach out to us, uh- We think the world till date was built by intelligence. The world in the future is being built on top of intelligence, and so be a, a great way to partner.
- 12:25
Well, thanks, Eiso. Thanks everybody here. And now we do have five minutes left. I don't know if we're supposed to take questions, but I'm happy to, so if anyone does.
- 12:31
But if not, I'm just gonna go that way. Is Eiso AI? What was that? Is Eiso AI? Sort of. I mean, I think of him that way. Here, for, here's a fun story.
- 12:42
Here's how I met Eiso. I like to tell this story because, um, Eiso's a fun, fun dude. I met Eiso because it started with a failed acquisition at GitHub.
- 12:51
So back when I joined GitHub in 2017 as the CTO, I wanted to take GitHub from a kind of collaborated, collaborative code host with open source bent and turn it into an end-to-end software development platform infused by intelligence.
- 13:02
And so you know the, the products that we launched from '27 on, or '17 on, GitHub Actions, Packages, Alerts, Notifications, eventually Codespaces, um, and then Copilot was the last thing that the office of the CTO did before I left with Nat Friedman, Uga De Moore, and a couple of other folks inside there.
- 13:19
But Eiso, in 2017 when I joined, uh, he had working code completion before the transformer architecture was, had landed fully, he had it on LSTMs. And so I quickly tried to acquire his company and he just, he just said no. [laughs]
- 13:33
He just said no to me. Uh, but we had... That was a long, drawn-out process talking about what we thought neural networks were gonna mean for the world. And so during that process, which was a lengthy one, we became really good friends, and we'd stayed in close contact over the years.
- 13:47
And then '22 rolled around, obviously ChatGPT comes out, Anthropic's out, and we kinda saw the end game at play, and we said, "Do we jump back in or not?"
- 13:55
And of course, yes, we jumped back in. But I like to tell that story about how he just kept saying no to me, and I just kept asking him questions, and eventually he said, "Yes, we should found a company."
- 14:04
'Cause by the way, when I asked him if we should do this, he said, "Oh, goddamn no." [laughs] That was word, his exact words. He's like, "No, we should just learn how to paint and sail."
- 14:12
But here we are. [laughs] So.
- 14:15
Yeah. It's, it's been a great road journey together, Jason. I, I think the reason we ended up doing this is because of our, our opinionated view on what it was gonna take to build more capable intelligence.
- 14:27
And, and the first 18 months of this company, you know, obsessing and focusing on reinforcement learning combined with LLMs felt like one of the most contrarian opinions in the world.
- 14:35
But I think today it's absolutely not, and it's super exciting to see the, the progress that's continuing to make. Like, we're... In the coming years, we're gonna see the world that started in completions and went to chat and is now agentic increasingly approach more autonomous.
- 14:49
And we're... A little bit is stemming effectively from the combination of bringing highly capable models that are constantly evolving together with real world problems. And, and I think what we're starting to see now is we're entering these kind of awkward teenage years ahead of AGI, where everybody in this room who's building out incredible companies and applications is
- 15:08
bridging this gap of what it really takes to make intelligence that in its raw form actually be valuable. And we, uh, we wanna be a, a small, humble part of that.
- 15:18
We've got a lot of work still ahead of us. Uh, the team is growing. Uh, but hopefully what you've seen today, uh, is what our, our customers and enterprises have been having access to and seen for a while, is that we're, you know, hard at work at, uh, at really pushing those capabilities, but also wanna make sure
- 15:32
we make them available to build together with others.
- 15:35
Well, that's it. Thanks everybody. [audience applauds] [upbeat music]