Ian Fischer - An Optimization Perspective on Recursive Self Improvement

0:02 Okay. yeah, my name's Ian Fisher. I'm a co-founder, co-CEO of Poetic. and Poetic, if you haven't heard of us, is a startup we founded about a year and a half ago that's entirely focused on recursive self-improvement. So, we're building AI that improves itself. and in the past year, and a little less than than a year, our focus has taken us from from nothing to a system that automatically can set state of the art on a wide variety of major public benchmarks. we're we're 10 scientists and engineers, mostly from Google DeepMind and Apple, Microsoft, Amazon, ByteDance.

0:39 some of our early results we had on reasoning and knowledge extraction, setting state of the arts on all these and and coding. So, let me dive into this. I want to define carefully what RSI is. and so, you know, it should be really simple. There's only three letters. and really two of them are doing all the work here. So, recursive means that the new improvements at each step are going to drive the next round of improvements.

1:10 and self is the system really is going to improve itself, not some other target. improvement is really obvious. Everything you know, everybody's doing improvement of some sort. but R and S really come together as a pair. if you don't have both of them, you you really aren't doing the right thing. and the reason that I'm being a little bit pedantic here is that common use of RSI is both too permissive and too restrictive. so, some people have described approaches as as being RSI that are really just a a standard iterative improvement loop.

1:47 and other commentators other people think that RSI is something that's only RSI if you are improving the parameters of a large language model. and this is too restrictive because there are a lot of other pieces of of an intelligent system that can be improved quite effectively. and so why does RSI matter? we think that this is, you know, RSI is the most important frontier in AI research because the improvements of real RSI systems compound.

2:19 and this is going to lead to superintelligence much more quickly than than human-driven improvements are likely to achieve. For you know, the obvious reasons that that humans don't become inexorably smarter while improving models, but RSI systems can. so I'll talk a bit about the the landscape here. We can split RSI approaches on two axes. On the Y axis we're going from cheap at the top to expensive at the bottom, and on the X axis going from things that really aren't RSIs, they're just standard iterative iterative optimization approaches on the left to things that are RSI on the right. and so, you know, the upper right quadrant here is a a nice place to be. Of course, we put poetic there, but that's not arbitrary. We we put ourselves there because our loop does in fact compound.

3:11 It really is RSI, and each step of RSI only costs an inference run rather than a training run for an LLM. in the lower right are where the best-funded efforts are. This is what people generally think of when they're talking about RSI. Again, the kind of this LLM centric perspective, where Anthropic, OpenAI, and Google are putting a lot of effort. And these are generally genuinely RSI, but they do require to train an LLM from scratch at every step.

3:41 In the upper right, we have a few neighbors, and they're they're also genuine RSI, and they're also cheap, but they're a bit narrower than our approach. Like Darwin Gödel machines and and Seeca are both targeting coding primarily and the minimax harness trains on some of its own generated data. So as I mentioned you you you basically you can you can partition this this landscape into two pieces and on the bottom we have LLM centric approaches that focus primarily on improving a particular LLM and then on the top we have more systems oriented approaches that view the LLM just as a component of a broader system that should be improved and shifting from an LLM centric approach to a view to a systems oriented view opens up a lot of possibilities for improvement that you that you would otherwise miss.

4:34 Also they tend to be faster and cheaper. So I'll I'll talk a little bit about two of the really prominent approaches. One is what the recent Anthropic post AI that builds itself describes where you take a version of Claude, you put it inside of Claude code and then humans plus Claude code improve Claude and so this is partial RSI clearly you know clearly there's a you know Claude is part of the system that's being is the thing that's being improved and is part of the system that's that's improving it.

5:11 But there are humans in there so I kind of don't want to count them and Anthropic doesn't count them either. On the right we have another common approach called an automated AI scientist. This is similar also to auto research and this is typically not going to be RSI. There's there's a number of nodes here that are doing different things but basically you know you generate the system generates a hypothesis and then it implements that hypothesis and then it measures it and then it like writes a a paper about it and puts that paper back into some knowledge store.

5:44 And this is not usually is not targeting a piece of its system itself. So, it's not actually going to be RSI. But, we can make both of these into pure RSI systems. In the case of Anthropic, we can just get rid of the humans. They've kind of indicated that that's that's where they're headed and that's maybe something that they're worried about. For the AI scientist, it's also a simple fix. You just target the LLM that you use in your ASI, excuse me, in your in your AI scientist loop. Of course, doing that makes it a much more expensive than what most people are doing with automated AI scientists.

6:29 So, what does poetic do? So, what we're doing is what we call self-optimizing optimizers. So, what's this look like? We we start with a general-purpose optimizer where a general-purpose optimizer is one that doesn't rely on rely on any specialized signals like gradients. It just requires measurable feedback from the optimization target. So, this is more general even than like black-box optimizers like hill climbing because it doesn't have to be you know, quantitative in nature.

7:04 So, then given that optimizer, that general-purpose optimizer, let's use it to optimize something. So, we'll have some optimization target. It's providing feedback to the optimizer. And as I said, that the feedback here can be quite arbitrary. So, you know, it could be standard approaches like accuracy, cost, things like this. Or more unusual feedback like rubric evaluations or reasoning targets. So, once we have these two pieces, this clearly is still not RSI. This is just standard iterative optimization. But, we have this this framework of a general general-purpose optimizer that we can then apply to the optimizer itself. So now the optimizer is optimizing the optimizer that's optimizing the optimization target.

7:51 And this is now RSI. So in order to have an optimizer that can optimize itself, you only need two things. It needs to be a general purpose optimizer and the optimizer needs to provide feedback about its performance. So since we started with a general purpose optimizer and every optimizer can do that, we don't have a problem here. But this is, you know, I I want to be clear this is very different from saying that the optimizer is Adam or SGD, right? That would not work.

8:23 So this is what we call the Poetic Meta System. It is RSI and we use it to optimize everything. So every task that the Meta System optimizes helps it optimize itself, which makes it become a more powerful optimizer. And again, since it is a general purpose optimizer, we can point it at any part of the Meta System itself as well as at any other measurable task like benchmarks, the the benchmarks that I showed you at the beginning.

8:51 And customer data, things like this. So talking about what we you know, empirical results here, our most recent empirical results. We have this blog post where we pointed out that benchmarks are dead for us. And what this is really referring to is not that benchmarks aren't useful, but that in the presence of a properly recursively self-improving system, static benchmarks have fairly limited value because it is we are able to fully automatically get soda results, state-of-the-art results on you know, every benchmark we've we've turned it on. And so again, as I said, It's automatically so zero human interventions and usually quite quickly.

9:39 So concretely here's a table bunch of results the the interesting thing one of the interesting things here is that on half of these benchmarks which are across a wide variety of different different domains that we've never worked on before. We got soda using older cheaper models than you know the whoever held the previous soda typically fable 5.

10:09 So what's next after now that we're not focusing on benchmarks we are working with with early customers in variety of variety of domains and I want to leave you with just this perspective on the frontier that the the next phase transition in AI is going to come from a system that invents its own improvements and so at we're we're intentionally casting a very wide net for that phase transition. Again by taking this optimization perspective on our side and using it to optimize everything including the optimizers themselves.

10:47 Thank you and we're hiring.