# The Improvement Loop Is the Product

Source: https://www.youtube.com/watch?v=ZZ19zhYnDKY

Machine-generated transcript; recognition errors remain. Source-recording timestamps. Boundary-crossing subtitle ends are clipped to this session.

[03:59:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14386s) Uh, hey guys, this is Era. I'm a

[03:59:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14389s) founding engineer at Klein and today I'm

[03:59:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14391s) going to be talking about recursively

[03:59:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14392s) self-improving coding agents. So this is

[03:59:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14395s) a very interesting conversation in the

[03:59:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14399s) sense that everyone wants to understand

[04:00:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14401s) what RSI is, how we accomplish it, how

[04:00:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14403s) we go through it and how it affects our

[04:00:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14404s) career. So I'm going to start with

[04:00:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14406s) giving you just three goals of this

[04:00:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14408s) entire conversation. Just three simple

[04:00:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14410s) goals and that would help you understand

[04:00:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14413s) like what what we get out of it and what

[04:00:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14414s) the future uh for this looks like. So

[04:00:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14416s) the goal number first is that

[04:00:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14418s) understanding the intent of Frontier

[04:00:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14419s) Labs, right? Everyone who's outside of

[04:00:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14421s) the frontier lab is always trying to

[04:00:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14422s) speculate like hey what what is the

[04:00:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14424s) topic in open or what is deep mind going

[04:00:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14426s) to come out with or what what are they

[04:00:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14427s) trying to make and what's the future

[04:00:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14429s) going to look like who has cracked the

[04:00:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14431s) RSI so like this kind of helps you get

[04:00:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14434s) get a rough outline of what their road

[04:00:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14436s) map looks like and the second thing is

[04:00:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14437s) that like if the models are improving so

[04:00:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14440s) fast like how do you think about your

[04:00:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14441s) career or the next few years right that

[04:00:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14444s) is really what we want to understand we

[04:00:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14445s) want to understand like okay things are

[04:00:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14448s) changing so fast like the ground

[04:00:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14450s) underneath us is shifting so quickly.

[04:00:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14452s) Right now like it's like uh

[04:00:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14455s) there used to be a time when you could

[04:00:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14456s) like demonstrate your side projects on

[04:01:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14460s) your personal website or GitHub and now

[04:01:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14462s) it's just like you could literally

[04:01:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14463s) oneshot them using any Frontier Coding

[04:01:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14465s) Engine and with that it's like it's like

[04:01:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14468s) oh well how how do we demonstrate proof

[04:01:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14470s) of work? How do we demonstrate you know

[04:01:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14472s) that that that we know what we're doing

[04:01:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14474s) and how do we think of our career for

[04:01:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14475s) the next coming years, right? So it's a

[04:01:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14478s) very interesting times that we're living

[04:01:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14480s) in. And then the third and the most

[04:01:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14481s) important goal of the conversation is

[04:01:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14482s) escaping the permanent underclass,

[04:01:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14485s) right? So I know this seems like a joke,

[04:01:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14487s) but this really is a very valid

[04:01:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14489s) possibility that people are

[04:01:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14491s) contemplating, everyone's like, "Oh

[04:01:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14492s) well, if everything that we want to be

[04:01:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14495s) doing can potentially easily done by a

[04:01:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14497s) frontier model, like what are we what

[04:01:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14498s) are we good for?" And like where does

[04:01:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14500s) software go? Where where does the future

[04:01:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14502s) of technology going to go? And what are

[04:01:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14503s) we going to do as as things evolve so

[04:01:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14505s) fast, right? So I think that there's

[04:01:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14508s) like a lot of like fear-mongering but

[04:01:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14510s) also there's like some very valid merge

[04:01:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14512s) where people are concerned that like if

[04:01:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14514s) all the revenue of the planet is going

[04:01:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14516s) to go to um a few hardware companies

[04:01:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14519s) that make frontier GPUs and then a few

[04:02:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14521s) few uh AI companies that that make the

[04:02:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14525s) best models like how are we how are the

[04:02:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14527s) rest of us going to throw in what what

[04:02:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14528s) are we going to do here? So, I think

[04:02:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14531s) I'll I'll walk you through a few

[04:02:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14532s) examples and a few ideas that will like

[04:02:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14534s) really really help you reconsider the

[04:02:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14537s) career choices and and and the life

[04:02:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14539s) choices you're making and what your

[04:02:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14540s) problems you're picking up. And then and

[04:02:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14542s) then that will help you like realign

[04:02:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14545s) what it is that we should be working on

[04:02:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14547s) and what it is that you you know would

[04:02:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14549s) would that you could work on that will

[04:02:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14551s) stand the test of time and help you both

[04:02:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14553s) just not just like on a psychological

[04:02:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14555s) but like financial and other levels as

[04:02:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14556s) well. So uh to give you a very precise

[04:02:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14558s) example, let's start with something very

[04:02:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14560s) precise that I had done. So a couple

[04:02:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14562s) months ago I uh came up with recursive

[04:02:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14567s) improve improvements on a coding

[04:02:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14569s) harness. So this was the first probably

[04:02:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14571s) public example of uh an automatic eval

[04:02:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14574s) improvement. So to give you an example u

[04:02:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14577s) we [snorts] used to do this thing where

[04:02:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14578s) we would measure the quality of a coding

[04:03:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14581s) agent using evaluation scores.

[04:03:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14584s) on terminal bench 2.1 Kim K3 with client

[04:03:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14587s) would score um usually like you know um

[04:03:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14590s) 60 70% and then depending on like how

[04:03:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14594s) good the model is like you get better

[04:03:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14595s) scores on on terminal bench 2.1 so we

[04:03:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14598s) noticed that when Kim K3 came out our

[04:03:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14599s) score was 77.5% which is a good score

[04:03:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14602s) but we were like oh is this is this

[04:03:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14604s) truly is it truly the best score that

[04:03:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14606s) this model can get on this benchmark and

[04:03:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14608s) also like what does the model release

[04:03:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14611s) paper claim about what their best score

[04:03:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14613s) is so the model release paper said that

[04:03:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14614s) they could score up to 85% or even more.

[04:03:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14617s) So we were just like okay okay so we are

[04:03:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14620s) not getting the sort of score and often

[04:03:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14621s) what we do with model releases in client

[04:03:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14623s) is that we try to like optimize your

[04:03:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14625s) score and try to get the best out of the

[04:03:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14626s) model. So when we started hill climbing

[04:03:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14630s) on the on the model like we we realized

[04:03:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14633s) is we realized that like you know

[04:03:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14635s) there's a lot of steps that we do says

[04:03:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14637s) that we do it gracefully says that we

[04:03:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14638s) pursue the right kind of problems and do

[04:04:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14640s) it the right way and we started thinking

[04:04:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14643s) like okay if we there's a lot of ground

[04:04:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14645s) work here how do we automate it and this

[04:04:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14647s) is a demonstration of like how we

[04:04:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14648s) automated it. So I'll I'll just leave

[04:04:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14650s) that as a placeholder for now and we'll

[04:04:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14652s) get get back to it later where we talk

[04:04:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14653s) about in detail of like how does like

[04:04:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14655s) this automatic improvement loop of a

[04:04:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14658s) client harness work. So this basically

[04:04:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14660s) means that like you have a coding a

[04:04:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14661s) harness that could just like basically

[04:04:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14662s) improve itself and become better on on

[04:04:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14665s) the same set of problems. So let's start

[04:04:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14667s) with the ground terms. What is recursive

[04:04:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14669s) self? Recursive self-improvement is the

[04:04:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14671s) idea that an AI system that helps

[04:04:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14673s) [snorts] build a more capable successor

[04:04:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14676s) which repeats the process you know until

[04:04:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14677s) you hit artificial general intelligence

[04:04:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14680s) right you have better systems you have

[04:04:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14681s) better improvements you have a better

[04:04:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14682s) system and then you know you have

[04:04:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14683s) singularity so a classic example is like

[04:04:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14686s) a self-reing coding agent right

[04:04:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14688s) self-reving coding agent is just like a

[04:04:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14690s) coding agent that can improve itself so

[04:04:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14692s) you know this this is all like a pie in

[04:04:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14694s) the sky idea like it's like it's like

[04:04:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14695s) you know yeah sure this looks cool but

[04:04:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14697s) how how did this even come out to be

[04:04:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14698s) like how we even get to here. So if you

[04:05:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14702s) notice in topics road map in 2021 in

[04:05:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14705s) 2021's road map they very specifically

[04:05:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14707s) said that uh uh they wanted to develop a

[04:05:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14710s) general AI assistant that can interact

[04:05:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14712s) conversationally with a user to create

[04:05:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14714s) edit and explain software. It would

[04:05:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14716s) display all the capabilities of a

[04:05:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14718s) software engineer colleague. It would

[04:05:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14719s) write functions. It would identify bugs.

[04:05:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14721s) It would design an API discuss style and

[04:05:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14723s) best practices. Right? So we don't

[04:05:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14725s) expect to reach this goal right away but

[04:05:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14726s) it's in the direction which we're

[04:05:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14727s) striving. It's interesting how this was

[04:05:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14729s) more than five years ago. More than five

[04:05:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14731s) years ago, people had predicted that

[04:05:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14733s) there exist a day where you will have

[04:05:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14736s) like a full-blown journal assistant that

[04:05:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14737s) can basically act like a colleague,

[04:05:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14739s) which we do have right now. Um, then in

[04:05:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14741s) 2022, Chad Deputy got released. We all

[04:05:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14743s) know how that happened. It was it was a

[04:05:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14745s) fun moment. It was like the first time

[04:05:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14747s) things got sparked for us. Um, then in

[04:05:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14750s) 2023, 2024, there was the first time

[04:05:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14753s) chat models could call functions. These

[04:05:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14756s) functions would be simple functions like

[04:05:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14757s) hey what's the weather or you know

[04:05:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14759s) what's the what's you know what's this

[04:06:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14762s) competition over this computations

[04:06:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14764s) result in Python. Um and then finally in

[04:06:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14766s) 2024 when s 345 was released that was

[04:06:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14769s) when the agents could like truly call

[04:06:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14771s) like actual tools the way we call like

[04:06:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14774s) think of tools. So it's like reading a

[04:06:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14775s) file editing a file you know reading

[04:06:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14778s) searching something on web uh listing

[04:06:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14780s) listing a folder searching and grapping

[04:06:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14782s) through you know uh a file system things

[04:06:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14784s) like that. So and in 2024 we had the

[04:06:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14787s) intuition that like chatbot what if

[04:06:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14789s) chatbox could actually talk to each

[04:06:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14791s) other you know so then that's was the

[04:06:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14793s) first time when we had like you know uh

[04:06:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14795s) multi-agent conversations now at the

[04:06:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14797s) time these things were very janky right

[04:06:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14798s) like they were just like taped together

[04:06:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14800s) in this like weird weird uh bizarre ways

[04:06:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14804s) where you have just like these classes

[04:06:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14806s) and these links and then like these

[04:06:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14807s) hierarchies of classes and it was like

[04:06:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14809s) it was all like kind of like a very

[04:06:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14810s) janky attempt to like make us think that

[04:06:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14813s) like these these models to like have

[04:06:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14815s) these pipeline of flows of responses

[04:06:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14816s) that can all communicate with each

[04:06:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14818s) other. [snorts] A lot of it was like

[04:07:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14820s) mostly deterministic and it was like you

[04:07:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14822s) know it was not as free flowing as it is

[04:07:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14824s) right now. So and you know we started

[04:07:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14827s) getting the intention. So this was the

[04:07:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14828s) autogen project but like in only in 2025

[04:07:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14831s) we we started [snorts] getting the like

[04:07:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14832s) what if the agents just kept getting

[04:07:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14834s) better and faster and cheaper like what

[04:07:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14835s) what if that just kept happening and

[04:07:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14837s) then in 2026 uh we we had the intuition

[04:07:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14841s) that like what if what if the agent

[04:07:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14843s) could improve itself which is like the

[04:07:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14845s) final you know the final part of the

[04:07:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14847s) whole process. So let me give an example

[04:07:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14849s) like what does it mean for an agent to

[04:07:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14850s) improve itself right? So like it's like

[04:07:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14852s) a lot of times when we think of like

[04:07:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14854s) evaluating an agent, we initially used

[04:07:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14856s) to think of evaluating an agent in the

[04:07:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14858s) form of like just a single turn

[04:07:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14860s) conversation. So we would tell we would

[04:07:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14862s) tell like um um agents that like ask a

[04:07:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14865s) simple question like how many toes does

[04:07:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14868s) a dash animal has? It's like it could be

[04:07:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14870s) let's say how many toes does a cat have

[04:07:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14871s) and then al would go through and the

[04:07:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14873s) element would like prepare response and

[04:07:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14875s) then we would get the response. So like

[04:07:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14876s) a cat has maybe 10 toes, 18 toes,

[04:07:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14878s) however many. I don't I actually don't

[04:08:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14880s) know how many toes a cat has. But

[04:08:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14881s) anyway, um yeah. So then we create the

[04:08:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14885s) response as like a yes or no. But like

[04:08:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14887s) when you work with an agent, there's a

[04:08:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14889s) ton of things that you're doing. So let

[04:08:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14890s) me give an example. Let me give an

[04:08:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14892s) example. So for example, you're going to

[04:08:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14893s) ask an agent like hey set up this MCP

[04:08:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14895s) for me. So now it's not just like it's

[04:08:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14898s) not just like you know a trivial

[04:08:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14900s) oneshotable like thing like it will like

[04:08:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14902s) search online and go what's the MCP?

[04:08:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14904s) what is the config? So, I'll like edit

[04:08:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14906s) and download some soft files. Like maybe

[04:08:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14908s) I need to add database update the MCB

[04:08:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14910s) tool. Maybe I need to read and edit some

[04:08:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14912s) files. Maybe I need to run some bash

[04:08:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14914s) commands. Right? So, you do all of that

[04:08:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14916s) and then the agent response like, hey, I

[04:08:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14918s) did it. And then you grade that with a

[04:08:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14920s) bunch of tests to gauge like, hey, did

[04:08:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14922s) the agent actually accomplish the thing

[04:08:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14924s) that it said it would do? So to do that

[04:08:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14927s) thing the second part which is like more

[04:08:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14928s) intricate where you're actually like

[04:08:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14931s) truly you know testing agents like uh to

[04:08:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14934s) their best extent like it's like you

[04:08:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14936s) know um we use this thing called

[04:08:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14938s) terminal bench. terminal bench is a

[04:09:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14940s) benchmark for you know AI agents in

[04:09:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14941s) terminal environments and that is

[04:09:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14943s) something that we are we have a lot of

[04:09:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14945s) like experience with at client and uh

[04:09:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14948s) it's like a it's like one of the best

[04:09:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14949s) ways to like really gauge like is your

[04:09:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14951s) coding agent accomplishing uh in general

[04:09:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14953s) the task that you're doing uh that you

[04:09:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14955s) want to be doing with age model. So

[04:09:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14958s) there's a couple there's a many

[04:09:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14959s) different parts of evaluation suit. So

[04:09:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14961s) ideally you want that to run in um you

[04:09:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14965s) know [snorts] you you want that to run

[04:09:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14967s) in um individual sandboxes. Then you

[04:09:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14969s) have like independent graders that like

[04:09:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14971s) check everything like how many turns it

[04:09:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14973s) took, how many tokens it took, how much

[04:09:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14975s) latency it took, how many how many how

[04:09:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14977s) many tool calls did it make, what was

[04:09:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14979s) the trajectory looking like and then

[04:09:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14980s) grading all the results for things like

[04:09:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14982s) hey did you not reward hack? Did you not

[04:09:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14983s) cheat through the solution? Did you not

[04:09:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14985s) just like look up the solution online?

[04:09:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14987s) Things like that. Um so we we basically

[04:09:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14990s) like a lot of these uh data sets of

[04:09:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14992s) evaluation problems have a set of tasks.

[04:09:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14995s) So when a terminal bench 2.1 they have

[04:09:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=14997s) 89 tasks. So what we do is like we spin

[04:10:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15000s) up 89 parallel containers where they all

[04:10:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15003s) run online and they run in parallel and

[04:10:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15005s) they're sandbox and then they each try

[04:10:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15007s) to solve a problem. So then they're all

[04:10:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15010s) uh in parallel and independent of each

[04:10:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15011s) other. And then at the end we get get a

[04:10:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15014s) score. So our score could like give us

[04:10:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15016s) measures like is is a is this is the

[04:10:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15019s) model performing well and what kind of

[04:10:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15021s) failures it has right so like you run

[04:10:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15023s) the evaluation you get a score so you

[04:10:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15024s) get an aggregate score and there

[04:10:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15026s) aggregate score could reveal a set of

[04:10:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15028s) failures okay failures could be of any

[04:10:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15030s) kind failures could be hey the read file

[04:10:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15032s) tool is not working sometimes the

[04:10:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15034s) entrance is so slow that like you know

[04:10:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15036s) the agent's working it's doing it's

[04:10:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15037s) making progress maybe it's running a

[04:10:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15039s) circle but it's so slow that it

[04:10:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15040s) literally like runs out of time it runs

[04:10:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15042s) out of like 45 minutes or where that

[04:10:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15044s) were given for the task and just like

[04:10:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15046s) stop working. So once you do that once

[04:10:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15048s) you understand hey this is what we ran

[04:10:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15050s) this is what the failures we got we

[04:10:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15052s) bucket the failures in different

[04:10:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15053s) categories. So like 10 failures of you

[04:10:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15055s) know this kind of tool called 10 10 20

[04:10:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15057s) failures of like say inference not

[04:10:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15058s) working or something else right and then

[04:11:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15060s) we once we bucket the failures into

[04:11:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15062s) different categories where like we try

[04:11:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15063s) to make an improvement to the harness

[04:11:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15064s) because like what if there's like

[04:11:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15066s) something wrong with the harness that we

[04:11:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15068s) could change such that the model score

[04:11:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15069s) would improve right and on and on and on

[04:11:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15071s) and you do this iteratively it takes a

[04:11:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15073s) while it's just like you know hill

[04:11:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15074s) climbing is like a very groundworky kind

[04:11:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15076s) of process and you know usually usually

[04:11:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15079s) we we took we used to take a lot of time

[04:11:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15082s) trying to do this but it was a great way

[04:11:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15083s) for us to like learn how to improve

[04:11:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15085s) models. So [snorts] one way to hill

[04:11:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15088s) climb on top is that like you know we we

[04:11:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15092s) could have a thing where you know we

[04:11:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15094s) started doing this thing where we could

[04:11:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15096s) just like what if we could have like a

[04:11:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15099s) longunning agent that runs for a while

[04:11:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15100s) and automatically improves the model

[04:11:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15102s) score and that's what we did. So we ran

[04:11:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15104s) GPD 5.6 soul that led the work and that

[04:11:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15107s) was like the main agent and then uh Ky

[04:11:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15110s) K3 was just running eval. So what it was

[04:11:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15112s) doing is that like GBD 5.6 was like hey

[04:11:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15114s) run these events get these cores let me

[04:11:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15117s) look at the failures let me look at

[04:11:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15118s) where where it went wrong and then I'll

[04:12:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15121s) make changes to client's harness and

[04:12:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15123s) then I'll run these eval again and then

[04:12:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15125s) I'll keep trying. So again, you can see

[04:12:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15127s) how like very quickly this can like go

[04:12:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15129s) off the rails because this is supposed

[04:12:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15130s) to be like the running and evaluation

[04:12:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15131s) itself takes like an hour. Then you have

[04:12:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15133s) to read the feedback. Then you have to

[04:12:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15135s) like see if you're properly interpreting

[04:12:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15137s) the feedback to make reasonable changes

[04:12:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15139s) to the harness that are like not reward

[04:12:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15141s) hacking that are just like actual

[04:12:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15142s) genuine improvements and then run the

[04:12:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15143s) eval again. So we ran that for 17 hours

[04:12:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15146s) and uh when we ran that for 17 hours

[04:12:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15148s) there was like a legitimate hell

[04:12:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15150s) climbing and what we found that was very

[04:12:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15152s) interesting. This happened with a very

[04:12:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15154s) detailed prompt that accomplished this

[04:12:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15156s) is that like we had baseline score of

[04:12:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15159s) 27% and then you know we we found issues

[04:12:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15163s) like loop detection loop detection is

[04:12:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15164s) when a harness just runs in circles

[04:12:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15166s) around a problem right uh we found like

[04:12:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15168s) background process grinding where

[04:12:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15170s) background processes would run too long

[04:12:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15172s) uh we found some file index issues and

[04:12:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15174s) other things as well I mean when I say

[04:12:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15176s) we what I actually me 5.6 solve right so

[04:12:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15179s) that found those issues that found and

[04:13:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15181s) bucketed those issues it's like hey how

[04:13:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15183s) about I make those changes in the

[04:13:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15185s) harness and I run the eval again and it

[04:13:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15187s) iteratively built the hill climbing

[04:13:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15188s) process completely autonomously on its

[04:13:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15190s) own without my interference and 13 hours

[04:13:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15192s) later when I come back to it like it has

[04:13:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15194s) improved the harness's performance in so

[04:13:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15196s) many ways so see what does that mean

[04:13:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15199s) what does that mean it just it just

[04:13:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15200s) basically means that like we're at a

[04:13:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15203s) point where the experience could make

[04:13:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15206s) the harness better that like there's

[04:13:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15207s) like a tool level learning loop where

[04:13:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15209s) the harness does some action, interacts

[04:13:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15211s) with the environment, understands

[04:13:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15213s) whether or not it did the right thing or

[04:13:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15214s) the wrong thing and gets some reward and

[04:13:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15217s) and like almost and then has there's

[04:13:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15219s) like a meta agent that's observing the

[04:13:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15221s) main harness and finding the changes to

[04:13:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15224s) it and just like incorporating those

[04:13:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15226s) changes and improving. Now it's I think

[04:13:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15228s) in my personal opinion it's much harder

[04:13:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15229s) to do in real time because if you do it

[04:13:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15231s) in real time you overindex on solving

[04:13:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15234s) certain problems or you're just like too

[04:13:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15236s) cartel to a specific thing. But I do

[04:13:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15239s) think that this is something that works

[04:14:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15240s) like on a broader scale of like working

[04:14:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15242s) on a wider range of problems [snorts]

[04:14:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15244s) and you want to be able to like you want

[04:14:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15246s) you want to be able to justify the

[04:14:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15248s) changes you're making on a broader

[04:14:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15250s) applicability so that the changes you

[04:14:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15252s) make here will apply to a broader range

[04:14:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15254s) of problems as well. So what does that

[04:14:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15258s) mean? Like what am I trying to get at

[04:14:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15261s) here? The point I'm trying to get at is

[04:14:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15263s) that the models are so good right now

[04:14:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15264s) that they can if they can improve the

[04:14:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15265s) harness, they can improve a lot of other

[04:14:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15267s) things. So then what else can they

[04:14:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15268s) improve? So this is this is a this is a

[04:14:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15271s) post uh primarily from OpenAI where GP

[04:14:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15274s) 5.6 autonomously rewrote and optimized

[04:14:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15277s) its own production kernels and those

[04:14:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15279s) production kernel improvements remove

[04:14:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15281s) the end to-end serving costs by 20%. you

[04:14:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15284s) know it also had 15% higher to token

[04:14:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15287s) generation efficiency and this was like

[04:14:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15289s) completely done by 5.6 soul on itself,

[04:14:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15292s) right? So what that means is that like

[04:14:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15295s) inference is becoming much cheaper and

[04:14:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15297s) that means that like if on if if for a

[04:14:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15299s) company like OpenAI if you're having 20%

[04:15:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15302s) low lower serving cost for a frontier

[04:15:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15304s) model that could be a billions of

[04:15:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15306s) dollars of improvement in price and you

[04:15:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15308s) could you could do that as well for your

[04:15:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15309s) own models. Um another interesting thing

[04:15:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15312s) is that like when people have kernel

[04:15:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15314s) writing uh competitions so core auto is

[04:15:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15316s) a company which has a kernel writing

[04:15:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15318s) competitions and you'll notice that like

[04:15:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15320s) they they mentioned that like the person

[04:15:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15323s) who won the kernel writing competition

[04:15:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15325s) the explic said that I have never

[04:15:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15327s) handwritten a GPU operator it's all ALM

[04:15:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15329s) and agent driven so notice that like how

[04:15:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15332s) like the best uh inference improvements

[04:15:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15335s) kernel improvements are at this point

[04:15:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15336s) completely exclusively done by a swarm

[04:15:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15339s) of agents or generally like agents doing

[04:15:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15341s) some magic to you know magically improve

[04:15:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15343s) like the kernel kernel code and have

[04:15:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15345s) faster inference cheaper inference. Um

[04:15:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15348s) yeah so basically like the improvement

[04:15:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15351s) surface that you think of is not just

[04:15:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15354s) like the harness improvement it's

[04:15:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15355s) literally the whole surface stack right

[04:15:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15357s) so the whole stack from like evaluation

[04:16:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15360s) to harness improvements to training data

[04:16:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15362s) to post training to world interaction to

[04:16:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15364s) inference like literally the entire pip

[04:16:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15366s) part of the pipeline like all the way

[04:16:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15368s) from like electricity going through your

[04:16:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15371s) machine to like making a network request

[04:16:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15374s) going through the model models doing

[04:16:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15376s) inference um making a response and

[04:16:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15378s) getting back to you like every single

[04:16:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15379s) part of the pipeline we can optimize it.

[04:16:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15381s) So the whole RSI loop has like um

[04:16:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15384s) there's a couple different parts so it's

[04:16:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15385s) like pre-training um uh neatron examples

[04:16:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15388s) of like you know AI curated uh

[04:16:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15390s) pre-training data that was good um

[04:16:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15392s) there's like post- training examples

[04:16:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15393s) which are you know where there's like

[04:16:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15396s) constitutional AI examples of like post-

[04:16:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15398s) training uh improvements then inference

[04:16:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15400s) I gave you the example of GP 5.6 So for

[04:16:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15403s) eval generation, I give you our own

[04:16:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15406s) example of how we improved ourselves on

[04:16:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15407s) our own eval. Um, in general, what you

[04:16:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15410s) want to think of is that like anytime a

[04:16:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15412s) model is trying to do something and it's

[04:16:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15414s) like not able to do it or not able to

[04:16:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15416s) accomplish it. So for example, like

[04:16:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15417s) imagine you're having a long

[04:16:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15418s) conversation where you ask the model

[04:17:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15420s) something obvious. So like it's like or

[04:17:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15423s) something that you think is not

[04:17:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15425s) necessarily obvious, but it's like it's

[04:17:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15427s) hard, but you would assume that the

[04:17:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15428s) model could potentially do it. So to

[04:17:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15430s) give an example, let's say you want to

[04:17:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15432s) make add a new column to a database and

[04:17:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15435s) then fill that column, right? And then

[04:17:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15437s) put a cache in front of it. So what

[04:17:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15439s) happens is it does the whole thing, but

[04:17:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15441s) for some reason the way it pushes the

[04:17:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15443s) cache in front of the whole uh database

[04:17:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15444s) like it kind of doesn't do it right. So

[04:17:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15446s) at that point you're like, "Oh no, set

[04:17:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15448s) this cache differently. Set this as the

[04:17:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15450s) TTL." And then you give like a

[04:17:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15451s) correction signal and then that kind of

[04:17:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15453s) like fixes your problem and you kind of

[04:17:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15455s) go about your day.

[04:17:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15456s) The way I want to think of it is that I

[04:17:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15458s) want to think of like every single time

[04:17:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15460s) like a human correction to the digital

[04:17:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15462s) landscape is a part of a delta that

[04:17:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15467s) needs to be covered. And what that means

[04:17:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15469s) is that like every single time you're

[04:17:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15471s) finding like the edge of model

[04:17:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15472s) capabilities of like how far it can push

[04:17:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15474s) for infants, how much it can push for

[04:17:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15476s) this, how much it can push for, you

[04:17:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15478s) know, solving your problems. You want to

[04:18:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15480s) find that delta and we want to like push

[04:18:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15482s) that frontier a little bit further. a

[04:18:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15484s) little bit further and that's what post

[04:18:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15486s) training does, right? So post training

[04:18:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15488s) is is this is in that post training we

[04:18:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15491s) find these like really interesting and

[04:18:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15493s) really hard long horizon problems for uh

[04:18:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15495s) for coding agents or for other kinds of

[04:18:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15497s) problems and then just like slowly hill

[04:18:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15498s) climb the models on on those kind of

[04:18:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15501s) problems. So anything that has a

[04:18:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15502s) verifiable loop can be optimized but is

[04:18:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15506s) this is not just limited to software

[04:18:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15509s) like literally like the physical world

[04:18:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15510s) itself could have a verifiable loop. So

[04:18:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15513s) to give an example like let me let me

[04:18:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15515s) give you a very funny kind of bizarre

[04:18:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15517s) example where like it's like you you

[04:18:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15520s) look at this and uh you um

[04:18:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15524s) um look at this uh like robot and it's

[04:18:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15526s) it's typing something and it's like

[04:18:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15529s) really really like interesting what the

[04:18:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15531s) robot is typing.

[04:19:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15541s) So notice how like notice how the robot

[04:19:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15543s) is itself typing like how it wants to

[04:19:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15545s) like solve robotics and imagine if the

[04:19:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15548s) robot itself could like interpret the

[04:19:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15550s) results of of like the answers and just

[04:19:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15553s) like keep improving then that means that

[04:19:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15555s) like this is you know basically like an

[04:19:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15557s) ASI robot. Another example that I could

[04:19:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15559s) give you for my uh personal life is that

[04:19:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15562s) like um I have this um chip that has the

[04:19:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15565s) ability to see and has a speaker and has

[04:19:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15568s) a microphone inside of it. So I was

[04:19:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15569s) trying to set up this uh with my machine

[04:19:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15572s) and it was um a gift that I had for my

[04:19:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15576s) girlfriend. And uh what I found very

[04:19:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15578s) interesting about this is as I was

[04:19:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15580s) trying to set it up, it was very slow.

[04:19:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15582s) It was slow to respond. It had higher

[04:19:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15583s) latency. It had an agent in and of

[04:19:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15585s) itself. Right? So what I did was I

[04:19:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15587s) connected the step to my um computer and

[04:19:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15591s) I let um I let um putting agents run on

[04:19:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15595s) this hardware and basically run on this

[04:19:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15598s) hardware automatically make tests to it.

[04:20:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15601s) So like my computer would talk to my

[04:20:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15604s) hardware give it some tests like hey do

[04:20:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15606s) this do that and just like what do you

[04:20:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15608s) see right now? So then it would notice

[04:20:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15611s) the end to end latency of like the

[04:20:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15612s) inference response of like how fast uh

[04:20:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15615s) my hardware is responding and then um

[04:20:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15618s) what I found is that like it was able to

[04:20:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15620s) do the breakdown of like every single

[04:20:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15622s) case and scenario where there was like

[04:20:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15624s) high latency um slower response or or

[04:20:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15628s) just the algorithm wasn't right and I

[04:20:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15630s) could then set up a recursive loop where

[04:20:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15633s) both the firmware could be updated the

[04:20:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15636s) um the tool calls the inference model

[04:20:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15638s) used um the speech to text model used

[04:20:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15641s) all of those could be improved

[04:20:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15642s) automatically with like an open router

[04:20:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15644s) key which I'd given to a coding agent

[04:20:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15646s) and then overnight with self-running

[04:20:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15648s) experiments with the hardware uh my um

[04:20:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15651s) this tiny little piece of beautiful

[04:20:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15652s) hardware that I have was like deeply

[04:20:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15654s) optimized so this also like even

[04:20:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15657s) hardware could be a part of the RSI loop

[04:20:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15659s) so what happens when you connect these

[04:21:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15660s) loops right that's the next question

[04:21:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15661s) like you could have cheaper inference

[04:21:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15663s) you could have like more experiments

[04:21:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15665s) budget you have like you know better

[04:21:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15667s) data and better models. So, is it

[04:21:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15669s) singularity? It's a very question to ask

[04:21:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15672s) and I think that the part of the parts

[04:21:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15673s) of the loop are already weird, right?

[04:21:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15674s) Like we're we we are really we really

[04:21:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15677s) are very close to parts of the RSI loop,

[04:21:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15680s) right? And what do we do like what do we

[04:21:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15682s) do about this and it's like how do we

[04:21:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15684s) think of that in terms of like our

[04:21:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15686s) career and how do you think of that in

[04:21:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15687s) terms of you know our lives and what do

[04:21:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15689s) we want to do next? So here's my advice

[04:21:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15691s) to everyone who is watching this like

[04:21:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15693s) please please please please move closer

[04:21:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15695s) to the RS group right so it's like right

[04:21:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15698s) now like people are working on all these

[04:21:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15699s) like different kinds of software which

[04:21:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15701s) is like you know a agents for you know

[04:21:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15703s) construction agents for this agents for

[04:21:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15704s) that and it's like I think that you're

[04:21:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15707s) you know what the way I see those things

[04:21:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15710s) is that like I think of those as like

[04:21:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15711s) you're missing the forest for the for

[04:21:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15712s) the trees you really are and it's like

[04:21:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15715s) you have this like very interesting

[04:21:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15718s) godlike technology that can change parts

[04:22:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15720s) of humanity. And then we're not

[04:22:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15722s) ambitious enough to pursue the best

[04:22:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15724s) problems. We really are not. And we're

[04:22:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15726s) just like I think that it's not just

[04:22:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15728s) like the problem isn't just that we're

[04:22:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15730s) not ambitious and we're also like what

[04:22:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15732s) we're pursuing. A lot of those things

[04:22:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15734s) are just like one model release away

[04:22:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15736s) from like being completely like wiped

[04:22:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15738s) out, right? So like it's also like it's

[04:22:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15740s) also like it may not be that

[04:22:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15742s) intellectually simulating because like

[04:22:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15743s) all you're essentially all we're

[04:22:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15745s) essentially doing is like sitting all

[04:22:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15746s) day. It's like hey cloud do this for me.

[04:22:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15747s) Hey cloud do this for me. And I don't

[04:22:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15748s) know if that's like I don't know if

[04:22:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15749s) that's like the most intellectual

[04:22:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15751s) stimulating or like or like um

[04:22:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15753s) thoughtful use of a person's time like

[04:22:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15755s) maybe maybe you know we're supposed to

[04:22:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15757s) be solving harder and more uh more

[04:22:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15759s) frontier problems and I think that like

[04:22:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15760s) it's like the ceiling is like our

[04:22:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15762s) ambition. So instead of like working on

[04:22:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15764s) these kind of things my suggestion to

[04:22:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15766s) people is to like you know do things

[04:22:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15768s) that of course like goes without saying

[04:22:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15770s) but you should do things you believe in

[04:22:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15772s) and you know I'm just a guy on the

[04:22:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15773s) internet but do things that help you

[04:22:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15775s) move closer to the RSI loop right? So

[04:22:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15778s) like anything that just like the end to

[04:22:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15779s) end pipeline of like hey this is where

[04:23:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15783s) like the models are lacking like it's

[04:23:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15785s) like whether that's inference whether

[04:23:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15786s) that's post training whether that's

[04:23:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15787s) literally like physical interaction

[04:23:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15789s) whether that could be you know send

[04:23:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15792s) sending end to-end pipelines of truly

[04:23:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15793s) any kind you know whether that's

[04:23:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15795s) hardware problems whether that's um

[04:23:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15797s) networking problems whether that's um

[04:23:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15799s) building a sense community things that

[04:23:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15800s) just you know um things that just like

[04:23:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15803s) help um you as an engineer get closer to

[04:23:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15807s) the benefit benefits of the improvements

[04:23:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15809s) um in uh in uh um in in the model. So

[04:23:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15813s) this is a very interesting slide from

[04:23:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15815s) Sasha topic. So um um they mentioned

[04:23:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15819s) that like you want to have like this

[04:23:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15823s) you're in this space where it's like you

[04:23:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15824s) know you only you only have like one to

[04:23:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15826s) five years left to contribute to luxury

[04:23:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15827s) to the world. What do you do? Right? So

[04:23:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15829s) it's like um they they gave great

[04:23:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15831s) suggestions were like prefer ambitious

[04:23:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15833s) targeted and past collaborations right

[04:23:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15835s) and you know in general like the time

[04:23:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15838s) scale like all the timelines that you

[04:23:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15839s) give to yourself like they should be

[04:24:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15840s) just like completely squeezed out should

[04:24:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15842s) be completely squeezed out because it's

[04:24:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15843s) like a lot of times and you also have to

[04:24:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15845s) think through that like if if the model

[04:24:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15847s) if the problem that you're working on is

[04:24:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15848s) too ambitious like it's like is it

[04:24:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15850s) probably is it probably better to wait

[04:24:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15852s) for just like the next next frontier

[04:24:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15854s) model to come out or not like that's

[04:24:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15856s) that's something you also have to weigh

[04:24:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15857s) through Right. And uh you you ideally

[04:24:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15861s) want to work on problems that benefit

[04:24:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15863s) from scale and you know you also at the

[04:24:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15865s) same time want to avoid problems that

[04:24:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15866s) can be solved by sale alone because then

[04:24:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15868s) you know if even if you like it's like

[04:24:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15869s) if you're working on a problem and then

[04:24:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15871s) the next model release comes out it

[04:24:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15872s) solves the problem. So that's not that's

[04:24:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15873s) probably not a good thing to do. So it's

[04:24:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15875s) very like you have to like think

[04:24:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15876s) philosophically of like what is the

[04:24:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15879s) endeavor that benefits uh greatly from

[04:24:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15881s) auto improvements but like isn't also

[04:24:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15883s) completely wiped out from it. And then

[04:24:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15885s) ideally like it's like if the if the

[04:24:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15887s) endeavor could also uh be a part of the

[04:24:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15890s) the like the improvement pipeline uh

[04:24:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15893s) that could be amazing. So you see a lot

[04:24:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15895s) of companies raising on say um making

[04:24:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15898s) data pipelines and there's a lot of

[04:25:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15900s) money there. There's a a lot of

[04:25:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15902s) companies working on physical hardware

[04:25:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15903s) and stuff. So I think if you're a

[04:25:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15905s) software engineer watching these, it's

[04:25:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15906s) like I think that if you're if you feel

[04:25:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15908s) like you're like here the delta of your

[04:25:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15911s) intellect and ability in your

[04:25:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15913s) day-to-day, it's like not contributing

[04:25:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15915s) as in like you have the intellect, you

[04:25:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15917s) can make a difference. But like if cloud

[04:25:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15919s) is just like oneshotting all those

[04:25:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15921s) problems like in your day-to-day and

[04:25:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15922s) you're just like kind of like becoming a

[04:25:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15923s) me proxy, it's worth to think through.

[04:25:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15925s) Take a step back and think through. like

[04:25:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15927s) okay where are the problems where where

[04:25:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15929s) you're actually truly uh contributing

[04:25:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15931s) and that they will stand the test of

[04:25:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15932s) time and then that kind of reflection

[04:25:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15934s) and realignment right now would I think

[04:25:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15936s) like have like very interesting

[04:25:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15939s) consequences for you um in the long term

[04:25:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15942s) um so yeah so thank you so much you know

[04:25:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15944s) thank you so much for your time like I

[04:25:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15946s) think that uh I think that if you're

[04:25:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15948s) watching this and if you've gone this

[04:25:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15949s) far I think that you're definitely

[04:25:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15951s) contemplating like one of the most

[04:25:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15953s) interesting improvements in the course

[04:25:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15955s) of humanity And uh you know there's many

[04:25:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15958s) philosophical arguments here. Some

[04:25:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15959s) people might say hey you know I I think

[04:26:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15962s) I'm maybe I'm off on the timelines or

[04:26:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15964s) maybe maybe I'm I'm just being too

[04:26:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15966s) pessimistic or maybe I'm being too

[04:26:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15967s) optimistic. But I would love to hear all

[04:26:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15970s) those thoughts. So this is you know my

[04:26:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15971s) QR code. Please feel free to scan it and

[04:26:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15973s) let me know all your thoughts. Uh thank

[04:26:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15975s) you so much. Very kind of you to give me

[04:26:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=15976s) your time.