# Chasing capabilities while waiting for the frontier

Alexander Shevchenko · AI Engineer NYC 2026 · October 14

[Watch recording](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1229s)

[00:20:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1229s) Hey, everyone.

[00:20:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1230s) I'm Alex from Ramp.

[00:20:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1232s) I'll be speaking about chasing capabilities

[00:20:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1234s) while waiting for the frontier to advance.

[00:20:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1237s) So a quick intro about myself.

[00:20:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1239s) I joined Ramp around three years ago

[00:20:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1241s) on the broader Applied AI team,

[00:20:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1243s) working on a bunch of different surfaces and products,

[00:20:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1246s) applying LLMs and ML to make a great product even better.

[00:20:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1253s) and to caveat my presentation on ramp labs ramp is an extremely bitter lesson build company

[00:20:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1259s) over the last year the number of use cases that use lms has increased 123 percent

[00:21:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1264s) while at the same time a lot of our top features are growing in token spend but on a unit economic

[00:21:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1272s) basis they're actually becoming cheaper because there are frontier models that are being released

[00:21:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1278s) all the time now, and these models are cheaper, faster, and better.

[00:21:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1282s) And so we're able to provide the same level of accuracy

[00:21:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1286s) at a much lower cost.

[00:21:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1288s) And so this is a really bitter lesson-pilled pattern.

[00:21:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1292s) And so given all this, a question that I sometimes get,

[00:21:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1297s) why not just wait?

[00:21:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1298s) Why have an applied research function inside of a company like RAMP,

[00:21:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1303s) an application layer company, right?

[00:21:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1306s) And to answer that, I would like to draw a parallel to 2023.

[00:21:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1311s) Back in 2023, you had access to kind of like GPT-4,

[00:21:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1316s) and creating an agent was a lot of work.

[00:22:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1320s) You would have to create very, very strict guardrails.

[00:22:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1323s) Models weren't RL'd to backtrack.

[00:22:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1325s) They didn't have thinking tokens.

[00:22:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1327s) And so if it went in the wrong direction,

[00:22:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1329s) it would just kind of charge ahead

[00:22:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1331s) and not realize that it made mistakes.

[00:22:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1334s) You had to sweat over all of the prompts to the point where some synonyms could change the behavior of your agent

[00:22:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1343s) because they weren't RL'd properly on generalizing to the user's intent.

[00:22:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1348s) You had to reduce and scope down the context as much as possible

[00:22:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1352s) because giving too much context to the agent would just confuse it.

[00:22:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1358s) And even back then, we decided, well, we need to ship these things.

[00:22:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1362s) We can't wait for GPT 4.5 or GPT 5 to come out.

[00:22:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1366s) We need to serve our financial professionals now.

[00:22:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1369s) And we should, things like RAM Tour Guide and many other agents,

[00:22:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1373s) even though it was not the easy thing to do.

[00:22:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1376s) And fast forward now to 2026, building an agent is a lot easier.

[00:23:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1380s) The models have gotten so much better that you can kind of give them MCPs

[00:23:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1383s) and access to kind of any context.

[00:23:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1386s) And they do a very, very good job of understanding

[00:23:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1388s) what they need to do given a task.

[00:23:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1390s) But I still believe that we can dig a layer down and create even better experiences for our users.

[00:23:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1399s) just it requires a bit more effort now.

[00:23:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1402s) So to present this, some of our work from Ramp Labs, three specific cases of how we've applied research into Ramp itself,

[00:23:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1413s) starting with creating a specialist model.

[00:23:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1417s) So in November of last year, we launched Ramp Sheets,

[00:23:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1420s) which was one of the really early spreadsheet agents.

[00:23:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1425s) This was before the two frontier labs had launched their versions of spreadsheet agents.

[00:23:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1429s) And we had a very good launch in November.

[00:23:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1432s) A bunch of users from great companies had signed up and started using it.

[00:23:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1436s) And we realized that, well, this is a very general purpose spreadsheet handling agent.

[00:24:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1442s) and a lot of the work that people do on RAMP,

[00:24:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1445s) both inside of the company,

[00:24:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1448s) but also all of our customers, accounting, and finance professionals,

[00:24:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1451s) are spending most of their days inside of spreadsheets.

[00:24:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1454s) So we can kind of take the harness and the brains behind RAMP sheets

[00:24:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1459s) and add it into a bunch of other surfaces within RAMP.

[00:24:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1465s) And as we were building this out,

[00:24:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1467s) we realized, well, we can probably apply some research from RAMP Labs

[00:24:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1472s) to make this even better.

[00:24:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1475s) And we sat down and we started looking,

[00:24:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1477s) where do models fall short on this?

[00:24:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1479s) And back then, this was like Opus 4.6.

[00:24:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1482s) And the place that had the most friction

[00:24:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1485s) was when you would drop in a lot of spreadsheets

[00:24:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1488s) into RAM sheets and you would ask it to build a DCF

[00:24:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1491s) or model some financial scenario.

[00:24:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1494s) It would spend a lot of time creating a plan at first

[00:24:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1496s) and then researching all of the sheets

[00:24:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1498s) that it had access to,

[00:24:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1499s) to try to get a good understanding

[00:25:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1502s) and the lay of the land.

[00:25:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1504s) And Opus 4.6 is a great model,

[00:25:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1506s) but it was pretty slow,

[00:25:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1508s) and the thinking tokens are also pretty expensive.

[00:25:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1511s) And so we took a step back

[00:25:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1513s) and decided to train a retrieval model,

[00:25:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1516s) or kind of like a financial question-answering model

[00:25:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1518s) that would, while still keeping these frontier models

[00:25:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1521s) as orchestrators,

[00:25:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1523s) would then be handed off to this much faster model

[00:25:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1527s) that would do financial question-answering

[00:25:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1529s) for that planning agent.

[00:25:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1532s) And so we partnered with the company Prime Intellect to post-train a Qwen 3.5, 35 billion model.

[00:25:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1542s) And we actually got above Opus 4.6, which was the frontier back then, exact task accuracy,

[00:25:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1549s) while also being able to serve it at Haiku level latencies.

[00:25:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1553s) So a much faster but just a smart model that could then be used in tandem with the really smart frontier orchestrator.

[00:26:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1564s) And so this was a very successful example of post-training that we had actually sent to prod as part of our Ramp Labs effort.

[00:26:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1574s) The second big thing that we are actually exploring up to this day is KVCache

[00:26:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1581s) tricks and particularly cartridges are a really interesting family of techniques where you're able

[00:26:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1589s) to compress a lot of context up to 20 to 30 times more context into the kv cache offline and serve it

[00:26:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1597s) for your agent and so if you compare it to the ramp sheets example for example

[00:26:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1603s) there's a lot of tasks that you know are going to happen at some point right like month end close

[00:26:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1608s) a month finishes and on the first of the next month a financial professional will go and try

[00:26:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1613s) to reconcile and close their books you know kind of the ledger the aging report and the documents

[00:26:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1619s) that they will need to be retrieving and so you could prepare them as context but the agent will

[00:27:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1623s) still need to agentically retrieve and create a good picture and understanding of them with

[00:27:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1630s) cartridges we're able to compress all of that work and build one overnight before the the

[00:27:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1636s) financial professional starts their work on the first of the month and once they

[00:27:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1641s) open up the the agent it knows exactly what it needs to do because it has all

[00:27:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1646s) of the context already built in cartridges are also a much lower lift

[00:27:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1650s) in post training because the way that they do self play is kind of a lot more

[00:27:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1654s) generalizable the caveat with stuff like this is that compared to post training

[00:27:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1659s) this is something that the labs might ship at some point in their API's in

[00:27:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1664s) their foray into continuous learning.

[00:27:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1669s) Even if they were to do it in the API,

[00:27:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1671s) we would actually be happy because we

[00:27:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1672s) are developing a lot of this tribal knowledge

[00:27:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1674s) inside of the company and just learning how best

[00:27:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1677s) to apply these techniques.

[00:27:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1679s) So even if it does ship as an API,

[00:28:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1681s) we can just hit the ground running on day one,

[00:28:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1683s) knowing how to use them.

[00:28:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1687s) And finally, the third example of applied research

[00:28:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1692s) that we are sending to prod is our work on agentic map reduce

[00:28:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1696s) and kind of agentic swarms.

[00:28:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1698s) So more concretely, we published an article on the Ramp Labs blog

[00:28:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1703s) on how we used 10,000 parallel agents to pentest all of Ramp.

[00:28:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1710s) So at Ramp, we have a cloud agent called Ramp Inspect,

[00:28:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1715s) which helps out with setting up all of the infrastructure and plumbing

[00:28:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1718s) for something as ambitious as this.

[00:28:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1720s) and we were able to spin up 10,000 parallel agents

[00:28:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1724s) and then kind of map this work out,

[00:28:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1727s) try to find any vulnerabilities,

[00:28:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1729s) and then also agentically reduce on the tickets that they would create.

[00:28:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1733s) So deduplicate them with LLMs

[00:28:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1735s) and then try to reproduce them adversarially with LLMs

[00:28:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1738s) to try to make sure that the vulnerabilities are real and important.

[00:29:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1742s) At the end of all of this,

[00:29:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1744s) we have actually found seven highest severity vulnerabilities

[00:29:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1747s) that we all patched.

[00:29:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1749s) Interestingly, one of the vulnerabilities that we described in that article is a Lua

[00:29:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1754s) script injection that you could send over via email.

[00:29:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1758s) And no single component of that stack was unsafe, and the kind of swarm had to patch

[00:29:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1765s) together multiple components and understanding of multiple components end-to-end in order

[00:29:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1770s) to be able to figure this out.

[00:29:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1772s) And the other nice thing with this is that this is all running automatically.

[00:29:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1776s) You don't need to spend anyone's time.

[00:29:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1778s) is just use started and the only thing that failed passes that get filtered out is the cost of the

[00:29:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1784s) tokens. The other thing that's interesting outside of just building out all of the infrastructure to

[00:29:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1790s) support these swarms is that we do research into how to best use models as part of these swarms.

[00:29:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1798s) One of the interesting

[00:30:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1800s) insights from that same article that we published is that you can actually use open weight models that are up to five times cheaper compared to the frontier reasoning.

[00:30:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1809s) Because you're setting it up as a swarm, the detection rate is not five times lower.

[00:30:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1814s) So, you can kind of spend more on these cheaper tokens and have almost as good of a detection rate.

[00:30:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1825s) If you follow through all of these examples, you kind of notice a pattern of a lot of them is attempts at saving on time and saving on money.

[00:30:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1835s) So reducing latency, which is time, and then reducing on token cost, which is money.

[00:30:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1840s) That's actually RAMP's slogan, saving our customers time and money.

[00:30:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1844s) And so I'd like to point to two products that we actually offer from RAMP that have been influenced by RAMP Labs' research.

[00:30:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1853s) The first one is token spend management that connects to a bunch of your API keys

[00:30:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1858s) and is able to give you a really good picture of how LMs are being used inside of your company.

[00:31:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1863s) Not only token spend on an API key basis or a user basis, but digging deeper.

[00:31:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1869s) Accounting categories and even things like tax credits that you can get by analyzing the spend.

[00:31:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1879s) It also goes deeper, and we create a taxonomy that contextually analyzes and semantically analyzes the LLM usage,

[00:31:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1887s) so you can actually understand how LLMs are being used outside of just attributing it to a team or a person inside of your company.

[00:31:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1895s) And then the second one is Ramp Router, which also leans into improving the latency of your agent,

[00:31:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1901s) where we have a bunch of routing strategies that we work with Ramp Labs to develop,

[00:31:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1906s) things like NVIDIA Switchyard, things like smart flex offloading of work to both improve latency

[00:31:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1913s) and cost per tokens that you spend. And there's a bunch of different strategies already available

[00:31:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1919s) for routing and even more coming out later. So if you're interested in any of these two products,

[00:32:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1926s) you can take a photo of the URLs or scan the QR codes there. I'll give you a second.

[00:32:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1939s) Awesome.

[00:32:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1940s) And then I'd also invite you to check out our RAMP Labs blog where we publish a bunch

[00:32:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1947s) of really interesting articles that I haven't had the chance to cover as part of this presentation.

[00:32:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1952s) Notably one of the really interesting ones is PorTAL where we figured out a way of transferring

[00:32:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1958s) learnings from one post train to another.

[00:32:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1961s) So you could do something like a LoRA post-train on one model,

[00:32:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1966s) so like the Qwen 3.5 that I was describing,

[00:32:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1968s) but then transfer it to newer versions of QAN

[00:32:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1971s) or even across different families.

[00:32:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1973s) So you're post-training this latent

[00:32:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1975s) that then can be projected onto different models.

[00:32:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1979s) And that really amortizes the cost

[00:33:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1981s) that you have to spend on post-training.

[00:33:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1983s) More blog posts on agentic map-reduce,

[00:33:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1986s) so not only the pen testing stuff that we did,

[00:33:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1988s) but also things like analyzing and creating taxonomies of LLM usage,

[00:33:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1993s) which is stuff that has gone into the token spend product.

[00:33:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1997s) Things like self-improving monitoring,

[00:33:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=1999s) where we use LLMs to automatically create monitors with Datadog

[00:33:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2005s) on a lot of our product surfaces,

[00:33:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2007s) and then as alerts come in,

[00:33:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2009s) improve those monitors automatically without a person in the loop

[00:33:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2013s) to try to basically create super, super robust and resilient software.

[00:33:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2018s) And finally, semantic tagging, which is this interesting foray

[00:33:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2022s) into creating this egentic map reduce over a bunch of user sessions

[00:33:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2027s) where we can annotate and create taxonomies that are open-ended,

[00:33:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2031s) so not knowing ahead of time what are the clusters of LLM usage

[00:33:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2037s) and trying to create a taxonomy that can then be reconciled

[00:34:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2040s) against the spend inside of the companies.

[00:34:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2043s) So if you're interested in this type of research,

[00:34:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2046s) x.com slash ramp labs where you can go and check out more of our research and

[00:34:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2054s) there's my Twitter if you have any other questions or want to talk about some

[00:34:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=2059s) use case as part of your applied research and that's it thank you