AI researchers debate how close we are to recursive self-improvement

Today I'm chatting with three of my AI researcher friends from whom I learn a lot every time we talk, and who also happen to be at somewhat open-ish labs and companies, so you guys can actually say things on the record.

I'm joined by Beren Millidge, who is the CTO of Zyphra, which is developing open-source models. John Schulman, who is the chief scientist at Thinking Machines, previously the co-founder of OpenAI, led the RLHF work that led to ChatGPT. And Charlie O'Neill, who is head of model training at Baseten.

The first question I have: if we're in 2036, it's been 10 years, and we don't have billions of crazy superintelligences that are running around that have radically transformed the world, what is the most likely reason that that doesn't end up being the case?

Other than sort of exogenous political shocks, or like there's a war, or they ban AI, or something. But what is the most likely technical reason that we don't— that 2036 isn't like a crazy alien superintelligence world?

I mean, like, my reason would just be like, it's got to be the sort of— like, there's been a classic thing, almost like Marvex paradox,right? Where like, we see like, you know, we think of the AI being like, if it can do this, it's going to be amazing,right?

Like, if it can solve these hard math problems, if it can win at chess, blah, blah, and then it solves these things, and then it's like, not that impactful. Obviously it's somewhat impactful, but like, not everything. It's like, if somehow that continues and like, there's never like the true spark of generalization that occurs, I think that could lead to like, the AIs just being like, extremely good at kind of everything that people like put into a benchmark, put into an environment, but like, there's still some persistent like sim-to-real, which is somehow blocking everything.

I think this is kind of unlikely. I think we do actually see this kind of generalization even from RL in practice already. But like, if it is just like, ridiculously hard to like, generalize meta-learning, plus like, we don't solve continual learning, it's just like, super hard and impossible.

Yeah.

Like, this would be my like, default scenario in that case.

Yeah, I agree with that. Humans have a lot of advantages over models now, and each time a new model comes out, it'll sort of, it'll catch up in some of these areas. But like, you end up getting bottlenecked by the places where the model is weaker and where it has worse judgment, or the models can't check themselves well enough.

Yeah, so there's this cycle that keeps repeating where people think, where a new model comes out and people are blown away and they're like, this is it, this is AGI, but then they use it a bit and then it starts to feel dumb after a month or so.

So that cycle just might keep going, and it's hard to predict how many times it's going to repeat. And like,right now you don't get explosive growth in capabilities because you still get bottlenecked enough when you're trying to do research and engineering that even if the model can write way more code than a person, it doesn't make you like, 100 times more productive.

But yeah, so maybe there are just more of these cycles than we would expect.

For me, it's like a question of how far off, like, this global optimum of a learner you could have on a chip is like, the transformer plus like, RL, basically like the current recipe. So like, I think people imagine that even once, like, once you have an agent which is better than all humans at AI research, even if it's like, 0.1% better than all humans, then the fact that you can run like, you know, hundreds of thousands, if not millions of these in parallel, you can run them much faster, like chips kind of speed up, that's going to outweigh every other like, bottleneck, and like, you're eventually just going to like, hit this like, very fast takeoff with recursive self-improvement.

I could imagine that if we continue along the trajectory that we're currently on, with that paradigm where, you know, it's basically just like, self-attention, RL, scaling up RL environments, the, I guess like, if you think about what happened with Moore's law,right?

Like, we had this very like, nice straight line and that held for a really, really long time. But there were so many like, discrete like, discontinuities and innovations that had to happen to keep that scaling law going. And the same thing has kind of happened with LLMs.

Like, we had this pre-training like, scaling law, and then that was kind of like, you know, hitting the diminishing returns. And then we came up with like, RL and solved that. And then we got this new like, you know, diminishing returns curve to hit that made it keep looking like a straight line going up.

And so like, if it requires another one of those discontinuities to solve, like, I'm not sure that like, the current method of like, training LLMs with these RL environments, even like, RSI targeted RL environments, would be able to discover that discontinuity.

And if not, like, we're probably going to hit this like, asymptotic like, curve where it like.

Sorry, but do you think the discontinuity will be harder than anything that's come since 2012?

If we had the answer to that, we'd kind of have the ability to implement it. But like, maybe there's the distinguish, we should distinguish between a discontinuity which adds to the current paradigm. Again, it's like cumulative. Like, there's some thing beyond the RL that we have to discover, and maybe they're capable of like, you know, connecting the dots in that straight line.

Or like, but like, again, how far off the global optimum are we? Do we have to go back and throw out like, you know, gradient descent and like, neural nets in general? And I don't think like, if you continue to scale up the current paradigm, an LLM, no matter how many LLMs you're running, are capable of necessarily discovering that if it's too far away.

Yeah, the only hope really is if deep learning just can't get us to an AI which is at least, can dominate human research and human development, including the human ability to come up with new paradigms and so forth.

Or like, I don't know, maybe humans would never, also never have discovered the next learning architecture, but to the extent humans could have discovered it eventually. But it just seems like, I don't know, if you just look at the progress that's happened since 2012 till now, and you just continue that on, I mean, I know that that's been powered by huge amounts of compute scaling and so forth, but it would be weird if like, it just didn't get to the point where it could like, dominate humans at least in R&D.

Especially over the next few years, there's going to be Ryan Greenblood was on the podcast recently, and he made this point that you could imagine as AIs get more and more capable and are capable of making progress on simulations which incentivize getting better at not only AI R&D, but generally at science.

So this is a thing that all the labs targeting, many startups are targeting. Or another intuition pump is if you look at the ELO score of chess bots since the '80s, there's just like a very linear increase in ELO over time, but there's this huge discontinuity as they cross the human range of human experts always win against AIs to like, human experts never win against AIs, as this linear increase in ELO happened.

And you could think, I agree with your point that so far, AI capabilities have not been that big of a deal in terms of their end economic impact in the world, but that's just because like, they're slowly rising in ELO relative to humans.

Yeah.

Yeah, I mean, I agree it would be very, I mean, the only way for this to not happen is if like, as you said, somehow asymptotic like, just before basically, because we're already pretty close, in my opinion, to like, where we'll start crossing like, the human ELO score.

And so we'll need to asymptotic before that. And like, that's the only way, you know, in the scenario you pose where like, somehow we're sitting here in 2035 and like, everything is normal for this to happen, I think.

Yeah, I mean, the only other way is like, there's like, some dramatic like, regulation on AI. It's like, this is kind of what I see as like, the most likely way for this scenario to happen, actually, rather than the technical thing.

Yeah, I think there's different kinds of research. There's like, research where it's like the order research style where the objective is already specified very cleanly.

Oh, for sure.

And you're optimizing that objective. And I think everyone is picturing like, if we continue along this path of like, you know, making pre-training loss go down, making RL environments work, you know, what on earth go up, that's going to lead to like, improvement.

But like, you know, maybe what Ryan is talking about is like, this much more open-ended type of science, which is required for like, paradigm shifts, where we can't specify the objective and the AIs are definitely not able to specify that objective either.

Like, we have to be really, really careful about how we specify objectives for any of these things.

And maybe your point is that like, the nature of the breakthroughs that have happened since 2012 is that we have found like, in 2012, people weren't saying, or I'm assuming, I don't know, you guys were there, or at least John, you were there, but I was not.

I was in primary school.

Actually, John, I'm curious for your like, sort of wisdom of the ages of, or wisdom of like, being in the trenches way back when, but presumably a big breakthrough was realizing that next token prediction is the, like, you wouldn't have thought that Nano GPT speedrun is the thing to be optimizing for in 2014.

But now that we have come to this new paradigm, that's, you wouldn't think to do a speedrun on that and have AIs get really good at that. But maybe there's like a next inner loop to optimize that the AIs wouldn't anticipate.

And there's an outer loop of like, revenue or something that eventually should be strong, but it's a very slow outer loop.

Yeah, in fact, I remember in the early OpenAI days having the intuition that actually just doing like, minimizing log loss wasn't going to get you to intelligence because like, the important bits are accounting to, for such a small fraction of the loss that like, it was going to be overwhelmed by noise.

So just training a language model on next token prediction just wasn't going to learn the interesting things you want it to learn. And we needed to craft better objectives that would put more emphasis on the important things. And like, you can make all sorts of arguments for this, and you could say, oh, humans probably don't learn how to like, we don't learn how to model everything in our environment.

We can't, most like, people can't create a photorealistic reproduction of some kind of scene they've looked at. So there must be, we must need a better objective, but then it turned out that it just worked anyway.

As you were pointing out, the inner loop, even in current AI research of like, post-training benchmarks or whatever, doesn't necessarily translate into what users like.

Oh, yeah. I mean, the whole field relies a lot on generalization, and it's very hard to predict when you're going to get generalization, or when you're going to get some kind of out-of-distribution generalization. So we know that if you train on the task you care about, you're going to do better.

But like, the most important advances are often the

types of generalization that we have noright to expect. So for example, from just pre-training on this very naive next token prediction objective to like, various tasks of interest where some very, that require understanding of the input in some deep way, or learning some skill from pre-training that's like, very rare and like, not very like, heavily represented.

And then like, also generalization from these verifiable tasks to less verifiable ones. This is also a type of generalization that there's no reason a priori to expect it.

Yeah, yeah. So this is an interesting question because one intuition pump that you could have for why you would see some sort of singularity very rapidly without even scaling up the inputs to AI progress that are not just AI labor, is that before every single experiment you run, that's, you know, like a seven-figure experiment, you spend an equivalent amount of compute on AI labor.

And so you just have automated versions of you guys spending a century thinking about like, what is the optimal experiment to run, doing like, small scale ablations, developing literally like, a century's worth of theory. So going back even before like, deep learning, before you decide what experiment to run, doing extremely optimal like, setting up of the experiment, then you do a century of thinking after the experiment is over, where you're like, analyzing what happened and what the next experiment to run is.

Yeah. Well, I think if you think hard enough, you probably could have expected some of these things beforehand. Like, there is probably some very clever way to do a small scale experiment that'll let you build the theory that then will generalize to the large scale experiment.

So I would expect that like, we're nowhere near the ceiling of how well you can do research. And like, I would imagine a future where AI is like, is doing a lot of like, analysis and theory building, like, spending a comparable amount of compute to the amount that you're spending on the experiments themselves, doing various kinds of analysis and building a theory around what we've seen so far.

I think there's like, really concrete examples of this when the objective is well specified. So again, like, all thinking can do is like, update your posterior based on like, you know, the bits that you've gotten since you formed your prior.

Like, you can't gain any new bits from like, just thinking. But when the objective is well specified and there is like, this data sitting around, like, I imagine there will be this big speed up in the current paradigm we're in.

And like, good example of this is like, you know, if you got an AI to think about like, the CAPLAN scaling laws, like, an AI at this point would have noticed that like, you know, oh, like, they've just taken these intermediate checkpoints and didn't account for like, the annealing.

And so like, this is wrong. And like, that would have caught that like, years earlier, we would have made like, you know, progress like, like, would have cut off a year or two of progress just from that like, observation from an AI.

And like, again, once the objective is well specified, which is like, lower pre-training loss or whatever, like, there's many, many good examples where if you just thought about it a bit more, you would have been able to like, cut down a significant on things that you've done.

So like, mu p and like, how learning rate scales with like, model size and like, realizing the model width is important in that as well. Like, I feel like you can, you can really back out a lot of these things and cut off like, a lot of low-hanging fruits.

I would imagine like, a 10 times speed up if our thing is just like, maximize the objective we're currently on. But I don't see that how that generalizes at all to, you know, come up with theright objective in the first place.

Like, just thinking doesn't necessarily buy you theright objective in the first place.

I mean, yeah, I think this is really the key question to like, any kind of like, very rapid RSIs like, from current AIs is like, how well can AIs generalize to like, learning their own objectives? Because to have any kind of like, self-propelling automated loop, you need the AI to like, propose objectives, optimize them, figure that out, propose a new objective, and like, have this like, not go off the rails at like, any point for like, a long, long time.

Coming back to Moravec's paradox, there might be like, a case of Moravec's paradox where like, we think this kind of like, autonomy and sort of like, being like, self-encapsulated, so we can, you know, think of what we should do ourselves and then go do it and like, have this loop is like, super easy because we always do this.

And like, obviously, evolution needs to create creatures that can like, survive on by themselves like, long period of time. And like, this just might be something that for some reason is like, really hard for the AI in the same way that like, locomotion stuff is really hard, whereas like, math is super easy despite being super hard for us.

I don't know, doesn't the time horizon increasing suggest that that's.

You would, yeah, exactly. I mean, this is another possibility which like, but I agree like, there's no obvious evidence for this. Like, in fact, the fact that, you know, RL agents now like, super persistent and it's quite easy to do this, it's kind of evidence against this.

But like, this would be, you know, potentially like, one of the reasons why like, we just don't get this like, immediate takeoff is like, if this is hard.

If you look back from 2012 till now, or maybe from when you started doing your research till now, what part of

all the innovations that have happened since that time, including purely engineering ones, including purely conceptual ones, what seems like the thing that is the thing that would be the last things humans would have to do before AIs totally automate AI R&D?

Probably just like, iteratively asking theright questions. Like, if you can get the AI to like, do any experiment, but like, you need to decide what experiments to do. And like,right now I think AIs are not very good at this, compared to coding the experiments at all.

Like, whenever we talk about research, they propose like, a bunch of like, miscellaneous things which are like, very, very tiny steps.

Or even going from like, you know, DeepMind's approach of like, we're going to solve intelligence by learning to play games at a superhuman level. That's going to be the approach to like, one random researcher like Radford being like, I'm going to try and just predict the next token off a very wide swath of data.

Like, and then even once Radford had discovered that,right, like, it took a while before people decided to scale it up because we had to come up with the idea of scaling laws and the fact that like, you could very reliably predict these things.

Yeah. Yeah, I would say that the last job for humans, or the role for humans that will last the longest is like, defining the objective and like, deciding what we actually want. So like, in that vein, something like deciding what, how like, the AI, the assistant should behave or what it means to be helpful or what's like, the objective when we're doing our all from human feedback is one such thing.

And then like, then later, like, defining like, constitutions and model specs is another one. And I think even if the AIs can do all the technical work, we'll have to still do a lot of that and decide what we actually want.

Yeah, alignment is the final job.

Yeah, alignment is sort of the answer, but it's also alignment itself can be kind of decomposed into like, specification of the objective or figuring out what theright objective should be. And then like, actually like, achieving or optimizing the objective you've defined.

And I think the first one is not going to go away anytime soon. And like, if I think about like, a post-training team and why you need a lot of people to be on the team, it's just because there are a lot of different like, areas where you have to figure out like, how the model should behave.

And like, there's no way of like, there's no way of, it would be very hard to automate the whole thing just because someone has to think about how should the model behave in this area.

Jane Street started using Antithesis to test their software in early 2025, and they were so impressed by the product that they decided to invest in the company. I recently caught up with Ron Minsky, who co-leads Jane Street's tech group, to ask about how Antithesis actually plugs in.

The thing that I think is most impressive about Antithesis is we started using it in a team that was building high-assurance software and being really careful. And nonetheless, it was able to shake out bugs that were otherwise going to be really hard to find.

And that's important both because it helps make those systems more reliable, but also because it helps the teams that build it to just move faster.

This matters more and more as code production is increasingly automated.

I think in general, as we've been using agents more and more, the key problem that you run into is the verification bottleneck, just the time it takes from people to look at code and figure out, is that actually something you want to accept in your production software?

And tools that make testing better are just incredibly helpful there. They just ease the verification bottleneck and make it possible for you to get more stuff done and move faster because you can have more confidence that the code generated by the agent is actually not introducing new problems.

To see how Antithesis fits into your development process, go to antithesis.com/dwarkesh.

What is the story for why there isn't huge consolidation in model providers? There's just so many things that point to centralization here. Is there, yeah, if you step back over the course of years, is there something that is going to prevent that?

Yeah, I think distillation is the main thing that fights against the centralizing force. Because basically anything that can be learned through RL can be distilled very easily because

like, it's a small number of bits. It's something that you can learn from a small amount of data. So if you can collect, if you can get trajectories from the model that show a behavior, you can easily distill it.

So I think, I think distillation is one of the things that fights centralization. There's also, I mean, there is a possibility that there'll be company-specific models that it'll be possible to learn from deployment and have a company continually improving its own model.

And such a system could be provided by the current oligopoly of model providers or some other currently smaller company. But I think that'll change the game a bit.

Yeah, and I also want to point out that like, continual learning and RLC doesn't stop distillation,right? Like, even if your model is improving every day, like, people could be distilling it every day. So it's like, the loops could just operate at the same pace.

Right, that makes sense. Okay, so copying model behavior,

you, I guess you need to know yourself what theright distribution to prompt is in order to get like, the relevant model behavior.

Oh yeah, for just distilling with supervised learning, the prompt distribution is extremely important. So it's very non-trivial to distill a model, even if you have full access to it and have the cut, the chain of thought and everything.

Yeah, it's non-trivial to distill all of the useful capabilities from it because you need to prompt the model with something. And you need to prompt it with like, realistic prompts. You need to have a really wide distribution of realistic prompts.

So yeah, one thing that's been coming out recently is some of the Chinese companies are probably using these router services, which are designed to allow people in China to use the US frontier models, which would otherwise be blocked in China.

But there are all these router or proxy services that allow people in China to use these models mostly for coding. And these router services are collecting and selling some of the data. So I think this is like, a very useful data set for distillation because it gives you the perfect prompt distribution.

Yeah, I think this is one of those things where AIs help a lot here. Like, if you actually look at like, you know, the frontier pipelines that say like, the Chinese models that they actually put in their papers, it's a lot of like, humans or like, they get seed prompts from somewhere, which is in combination with humans, this kind of data.

And then they like, synthesize a vast coverage from those seed prompts using their existing models or like, the other frontier models. And so it's like, you can automate like, an awful lot of this like, prompt distribution gathering and like, environment creation.

It's just like, humans need to provide like, increasingly fewer amounts of bits as like, the models get better.

Right, or somebody, it still seems your bottleneck by like, having a service which has users or users are going through. So like.

Not necessarily. I mean like, yeah, that's obviously very helpful, but like, theoretically you can just think about like, what users want or like.

No, but it's like, the whole point is that we don't, the user says, make me an application like this. Oh, that didn't work. I actually want you to make this new feature. But actually, let's step back and do this other thing.

And capturing that whole trace is the, or to the extent you could have done that anyways, then you just have like, RSI anyway.

Yeah, I mean like, ultimately like, if you have this like, fully automated loop, that is basically RSI,right? Like, the AI is deciding the data, it's deciding the training, that is the loop. But yeah, I mean like, it depends how much human information you need.

Like, at some point if you're just like, I want traces that look like this, you prompt that to the model, the model will be able to like, come up with like, a pretty good approximation.

But what if you want to do like, make me a really good politician and it has to like, anticipate de novo, like, how would a discussion in like, the Senate halls go or something? I just feel like, there's going to be a lot of things which are.

Ironically, this is actually, I think, easier for the distillers than the frontier labs,right? Because the distillers just like, I want a good politician, they go to the like, frontier model. The frontier model already knows how to be a good politician, so it just like, generates those traces.

Whereas like, if you actually want to build the first model that does this, you have to like, actually somehow like, get data on like, what politicians do every day and like, build that. So it's actually much easier to like, say like, I want something like this and then like, get like, the AI to produce like, a billion variations than to like, actually create the thing like this to begin with.

I think you can actually make a really concrete prediction based off like, this observation that the Chinese labs have this router data. So like, I think the thing that just stated this originally was I was saying, isn't it weird how Sonnet 5 and Opus 5 are like, like, almost objectively worse models than like, GLM 5.3, Kimi K3, even though they've had access to like, not only distillation but logit distillation from like, Mythos.

And so the counter here was that like, okay, the prompt distribution really, really matters. Like, you need to see what users are doing so that you can distill like, kind of these behaviors and things in. I think the prediction from this is that the frontier labs don't necessarily have much of an advantage, if at all, in RL environments now.

Because yes, like, user distribution matters for like, general like, behavior and so on, but like, the best measure of a capability is the very, very hard RL environments you've made at the frontier. And so if you have access to those RL environments as Anthropic and you have access to logit distillation and you've still made a worse model, then maybe like.

Then real-world deployment matters more than the environment. That's really interesting. So, but they had to incentivize those capabilities in the first place in Fable or the frontier model. And so it's weird that they can't incentivize them again

with like, a smaller model or something.

Maybe like, maybe we're just in this weird like, uncanny valley where, you know, like, actually trying to copy that frontier model too much, like the student-teacher gap or whatever it is, it's just like, too large. And like, I think people made this point with Opus is it's like, the difference between Opus 4.6 and Opus 5 is that Opus 5 really feels like it's got this like, AI as a judge checking every possible thing it's done.

It's like, that's why it uses so many tokens. It like, tries to think about all these things, but it doesn't necessarily have the big model smell of Fable to know when to like, stop doing that or like, when's a good path to go down or whatever.

The reach exceeds the grass.

Yeah.

Yeah, I would offer a slightly different hypothesis. So I would say there are a couple of different axes for the environments you can create. And like, one of them is difficulty and the other is realism. It's sort of easy to create, or it's comparatively easy to create a lot of difficult environments like, that are just like, involve like, doing a much more complicated task or doing something that requires a lot more cleverness.

And you could say this is like the benchmarking distribution because a lot of the most prominent benchmarks just involve doing some very hard puzzle-like task that's easy to verify. And then there's sort of like, the realism axis where you want the model to be good in the realistic coding agent setting where there's like, multiple back and forth with the human and there's like, multiple objectives.

And like, I'd say

like, the people, like the labs who are crafting the model behavior for the first time need to push in both directions. And to get good model behavior, you need to really push on the realism axis and have like, rubrics or some kind of human feedback that's informing the reward function you use there.

But I think when, if you try to do distillation naively, you end up just sort of matching the teacher on the benchmarking distribution. And, but if you don't have enough of the environments that really exercise the capabilities in these like, trickier realistic settings, then you're not going to, you're not going to get those into your student model.

And I think maybe one thing that's happening is the big models generalize better from the like, the tricky narrow tasks to these sort of more realistic tasks. So if you have a really good like, realistic prompt distribution for distillation, you can match the big model really well.

But if you only have this like, this distribution of easily verifiable tasks, then you can match the big model on all the benchmarks, but you do worse on this broader distribution. So you might, that might even explain something about the smaller Anthropic models like Sonnet 5, though it's hard to predict exactly what they're doing to post-train those models.

It could also be that they're always changing their post-training stack and they just made, they just got a few things wrong in some of these models. So they, like, I don't know, they turned up like, something too high and created some quirks that people really don't like.

So it's like, really easy to screw up post-training in some way that doesn't show up in benchmarks.

I mean, just one of the sort of very basic points is just like, the AI, like, the frontier AI labs buy all their data from big data companies. And like, the Chinese can also just buy the same data from data companies.

And they are,right?

And like, they are, exactly. There's a lot of people like, you know, being annoyed about this, but like, if they have exactly the same data and like, they can buy that, they can also distill. It's like, it means it's quite easy to like, keep up really.

Yeah, yeah, yeah. Okay, the other question I had is how the first models that are capable of automating AI R&D will actually be trained? Because there's a toy version, which is this thing that Ryan was talking about, which is you just have GPT-8 try to build GPT-3 size models that are really good at like, inner loop type challenges of beating video games that require continual learning or just getting to a certain loss with like, the least amount of compute, et cetera.

But John, I think you had an interesting point that maybe that's not the way it actually will happen in practice. So I'd be curious about, yeah, by the point in which you have AIs that are actually capable of automating AI R&D, how are they probably trained?

Yeah, I think we'll probably do some combination of learning from human feedback to absorb like, the researcher's taste and just like, creating a lot of practice environments, which involve like, doing multi-step research projects. So I think, yeah, people will in practice do some combination of those two things and just each iteration like, patch whatever seems to be most broken in the last iteration.

So like, researchers will be using the AIs a lot and will notice that they have some consistent weaknesses and then those things will either be patched by like, collecting human feedback or like, creating environments.

Yeah, makes sense.

Maybe useful when you think about this is like, how much of the lineage we roll back and then let self-play from there. Like, I think in the limit like, you're picturing like, you know, just giving them like, a GPU and maybe neural nets or something and saying like, okay, figure out how to train a model to like, do this particular task.

Like, the way it currently works is like, we go up to the very like, edge of the lineage and say, okay, like, here are the bugs like, you know, Anthropic has found and they're training stuff in the last few months.

We'll turn those into environments like, you need to train and get better at on the frontier. And so you obviously lock in all the previous history of the lineage. But you could imagine a world in which you roll back to like, you know, before GRPO or something.

And then you have environments which like, trying to get it to discover like, the best form to like, RL models on and then maybe roll further and further back. But I think we will be still so compute bottlenecked that like, people will just keep like, staying at the frontier and like, diffing essentially the bugs and whatever improvements they found since the last model version turning those into training environments.

Which is also really good for having non-steal, like, new data between model generations. It's just, again, this is basically continual learning within the AI lab of distilling the last three months of AI research progress through environments and like, RLHF type stuff back into the model itself.

And it is distilling,right? And that's maybe why some of us feel like it's asymptotic. It's like, you're always like, just trying to get the last three months of progress. And that progress is being contributed to by AIs, of course, but it also still has humans in the loop.

And it feels like, you know, you're just constantly inching closer and closer to what the human researchers are like, finding and capable of doing.

Yeah.

I mean, the one thing I will say though is like, obviously if you're just distilling on like, trajectories, you can never go above. But environments can go quite a far way above what a human can do. Like, it's very easy to design an environment that like, no human can solve, but the AI can obviously still try and solve it.

And so that would be the path to like, go ahead of just like, what the human AI researchers do.

Do you have like, an example of like, in terms of RSI or like, you know.

Like, maybe a nicer speedrun by doing it even faster than a human speedrunner.

Yeah.

I mean, I feel like in AI research especially, it's very easy to define like, goals which like, you know, you could say like, the loss needs to be like, 1.3 or something and like, no human can get that, you know, now.

But like, that's a very extremely measurable, verifiable task. And the AI, if the AI gets that, then great.

Right, and I'm building like, 100 million parameter model that beats Minecraft. That's maybe too easy, but like, it beats like, a much more complicated game or something.

Isn't it crazy that 100 million parameter model to beat Minecraft, we're calling that too easy? Like, I imagine if you said that like, five years ago.

I would say a lot of research is not exactly like that though, where it's like, hill climbing on a well-defined goal. It's sort of more like, here's an intuition we have about some way models should be better. And then we also have some idea for an algorithm that seems to go a little bit in this direction.

So let's come up with a task that is sort of designed to show signs of life on this approach and like, see if we get some, get those signs of life. And then if we do, we can make successively more realistic versions of the task.

Right, it's like a lot more guided by intuition. And then the inner loop is to elicit the, or make tests for that intuition rather than like, the test itself leading to the insight.

Right, like, you're not directly optimizing for the eventual objective you care about or the practical like, production objective. It's sort of,

you're relaxing your objective a little bit. You're saying, yeah, let's relax on the realism axis a little bit and find some methods that actually work. And then like, then try to get back to realism later after the method matures a little bit.

Yeah, yeah, yeah.

And then there's also like, more, there's research that's more oriented towards explaining things and like, developing a theory or a sort of, yeah, often we don't have like, mathematical theories in machine learning that are that predictive, but we have like, a lot of like, more informal theories for what's going on.

Yeah, I mean, like, presumably the models will be trained on like, some combination of all of these tasks and like, some will be very easily verifiable. Some will be like, LLM is judge or like, just ask the human, like, does this look reasonable?

And then you will, the hope would be that like, these would all generalize to like, these much sort of hard, sort of more vague, fuzzy kind of tasks. And like, it probably will to some extent, whether it generalizes enough that like, we could, the loop can become like, self-sealing without humans being in the loop at all.

It's like, unclear.

Yeah, yeah, yeah. Maybe taking a step back, here's what I, here's what it seems to me that the plan for AI research going forward is. And you tell me if you think it's going to work or if you agree with this characterization.

So the bet is that we will scale up RLVR training across millions of diverse environments across hundreds of different kinds of domains. And what will emerge on the other end is an agent which has like, learned these basic skills or less than basic skills around being persistent, being able to triage information and context, eventually having like, end-to-end optimization of working with other agents and things like that.

And such an agent will be very sample efficient within the context. You know, you've done research on how you actually scale up in context learning to make it like, arbitrarily long, but you keep scaling it up. And so what comes out the other end will something, will be something that basically functions like a drop-in remote worker over the course of a week or a month.

First of all, do you agree that that is the bet the labs are making? And second, is it, is that enough? Like, basically learning how to learn within the simulacra within a data center and then getting deployed into the real world, but not actually like, learning from real world deployment, only learning these meta skills from the simulated environments in the data center.

Yeah, I think it's now hard to separate out like, how much of the lab's effort is going towards like, direct RSI versus like, making generally intelligent models that they can continue to deploy to collect revenue to fund the next big training run.

I think for the latter, like, yes, that's probably just the bet they're making. Like, and it's very clear like, the pattern of like, where these environments are going over the last few years. I mean, like, Anthropic's lineage of environments is like, a very clear example of this.

Like, you know, firstly like, we just focus on coding and like, we're going to get really, really good at that. And then the task horizon that we've got from coding, which is probably the lowest hanging fruit in terms of like, data available on the internet to create environments, like, their own internal stuff that they can turn into environments.

Then we're going to generalize, we're going to go after finance next. And like, literally like, just so much Excel data and all that sort of stuff in the RL training. And then, you know, it's PowerPoints, it's like, this long tail of like, the working economy.

And like, that seemed to work really well. And like, a lot of the other labs and things, even the open source labs have now realized that that was the correct bet.

What is the implication from that? When I had Dario on the podcast, the thing I asked him was, if you truly expect models, which will be human-like in their ability to learn on the job, why would you try to bake in all these skills of like, working with PowerPoint or something?

Why didn't you just expect the model to be able to pick that up on while it's deployed? And so, yeah, there's multiple different explanations. One is just that this is, we expect models to get there soon, but they're not there yet.

So why not amortize these skills into the model training? Another is that we're not concentrated on making it really good at widely deployed work. We just want it really good at RSI. And this is just like a way for us to like, get revenue so that we can pour it back into a model that is actually like, really good at doing RSI development.

And then like, once the singularity happens, the thing that comes out the other end will be really good at all the things which seem like bottlenecks to the current generation of models. Yeah, John, I don't know if you have takes on like, what, how much you construe why there is so much task-specific knowledge in these models if the path is like, this kind of generalization.

Yeah, I mean, if the models were good enough at learning in context, then in theory, you wouldn't be, you wouldn't need to train them on finance. They would just be able to figure out, read all the books on the fly and figure out how to do everything in the appropriate jurisdiction.

Yeah, and you could argue that you need to do a lot of this domain-specific training just to make them more efficient. So even if they were smart enough to figure this out on the fly, you still might want to do a bunch of RL and bake the, bake all these intuitions into the weights so the model would be more efficient at runtime.

Yeah.

Yeah, I'd say in practice, it does seem like model providers are going domain by domain and trying to strengthen the models in the highest value domain. And I'd say that that's one of the answers to why the models have gotten so much better.

It's just because the model providers have covered a lot of the high-value domains and the most common types of skills.

I mean, I think another thing is just like, it's not that expensive to do both at the same time,right? Because like, the models are massive. They can easily afford in terms of their parameters to like, learn everything.

Yeah.

And like, there is likely some transfer in sort of even just, even if like, finance is not specifically like, the information is important for like, RSI, just the general like, meta-learning of like, how to figure out what's important, how to have taste, how to like, do long-horizon work is potentially generalizable.

And like, there's not that much RSI-like data in the world as well. Like, it's kind of hard to generate and like, that requires a lot of effort to like, if you can sort of amortize in this other data, get some transfer from it, you already have masses of compute and masses of parameters based on like, why not do that as well as like, obviously the direct like, commercial incentive of like, selling a model.

That makes sense.

Oh, yeah, I'll add that, I mean, there's one question about whether this current paradigm of doing like, sim-to-real will be the dominant one forever. So basically you look at what the real world tasks are like and then you try to create a bunch of environments that can be simulated like, in the data center and you can do RL on them.

And I think obviously this has been very successful, but it has a lot of weaknesses because a lot of things are just kind of hard to simulate, especially if they involve like, interacting with a bunch of humans in real time.

Yeah, so there's some question about like, simule, like, whether sim-to-real will be the dominant framework forever.

I think sim-to-real has to be the dominant framework while like, sample efficiency is kind of low. Because like,right now you need like, you know, thousands and thousands of interactions with the humans and no human is going to sit there and like, deal with this, basically be in the loop of RL training.

Yeah.

And so like, we kind of have to simulate that now to like, get the samples you need. But like, obviously if sample efficiency improves a lot, you'd expect learning from deployment to like, become like, a much bigger part of it.

Though there are also other things you could do, like, you can learn off-policy so you can take all the traces and even without resimulating everything, you can potentially learn something from them.

Jane Street just launched a new competition and it's their most ambitious one yet: Design a protocol-emulator ASIC. Basically, if you have a chip that you want to test, you can connect it to this ASIC and then this ASIC will simulate realistic traffic.

That way you can see how the chip responds without having to plug it into a live system. Jane Street is looking for flexible, general-purpose designs, not single protocol emulators. When I was chatting with them, they suggested that I start off by trying to implement what are apparently three very common protocols: UART, SPI, and I2C.

Jane Street also mentioned that they hoped that more ambitious designs will also tackle low-speed USB and Ethernet and any other protocols that flex your chip's specific architecture. Importantly, your design should be reprogrammable rather than smashing a bunch of specific protocols onto a chip.

If a new protocol comes out after your ASIC is taped out, your chip still needs to be able to handle it. How exactly it does it is up to you, but there is one hard constraint. Your design must target an open source 130 nanometer process node.

That's because Jane Street will pay to tape out the most novel submissions and send the physical copies to the winners. The competition is open till January 18th, 2027, and working in teams is highly encouraged. Go to janestreet.com/dwarkesh to download the template code and get started.

I want to ask more about this because it's sort of weird that you have 50% of compute that's spent on inference that is not directly helping the model become better. Like, one of the key advantages you'd expect eventually digital minds to have is, unlike a human who gets to have 50 years of like, real-world experience, a model will get to, through all its instances, will get to experience, I don't know, millions of years of deployment across all kinds of economically relevant work in the economy.

Andright now that data is just not, in a meaningful sense, helping the model get better. Like, it just seems so obvious that eventually models should be able to learn from this data. And once they do, you would have something that almost feels like a widely deployed intelligence explosion because the model is assimilating so much information across all these deployed instances.

But when do you expect this kind of hive mind, kind of crazy shit to be starting to happen?

I mean, I think broadly, like, at a very basic level, this is already happening,right? Like, just in the next generation of models. So like,right now you can obviously take your deployment data and put this in the pretraining or the mid-training of like, future models, especially if you do like, some kind of filtering or some kind of like, judgment or annotation or like, recent, you know, synthesization of that.

How much do you think that explains the generation over generation improvement?

I think it explains like, quite a bit.

Really?

I mean, especially like, I mean, this is, you know, I don't know whether the labs do this because, you know, theoretically they claim not to train on people's data. But like, the Chinese 100% do. And like, they definitely get this advantage, both like, obviously deploying, this is basically what distillation is.

Like, they take out the models, they get some of their like, deployment data, they get some fraction of that by like, pinging the model, and then they train their next generation of models on it. And they can certainly do it on their own models as well.

Like, there's no reason not to whatsoever.

I completely agree with this. I think if you zoom out far enough, this is like, definitely happening. Like, you're picturing this like, and we're all picturing this, this is like what continual learning, like, the Holy Grail is. It's like, this very, very organic, like, live loop of like, an individual model like, getting an experience and like, live updating on the spot and learning from that.

And like, a lot of things break when you like, zoom into that level of granularity. But like, yeah, the big labs are doing this, like, the closed models are doing this. There's also early signs of life of like, people using open source models doing this at a much faster cadence.

So like, a good example is probably like, Composer. Like, you have some sort of model and you are able to, or like, you know, Harvey's doing the same thing with like, legal agents. Like, it is getting very specific environments from the data that you have for that particular task and things that like, you know, users are complaining about.

And like, all the feedback that you're somehow extracting from like, your specific deployments. And a lot of these companies have the advantage over the big labs in that they can use this data really, really well. And then they will create environments.

They will like, you know, do a big post-train of Kimi K3. They will go deploy it. They might do some online learning as well. Like, Composer did online, basically like, reinforce for a long time. So yeah, like, there's still a human in the loop.

There's still a human saying, okay, these are the signals we care about. Here's how we're going to create environments from the data that we have. And it's like, still a longer cadence than maybe the one that you're thinking of.

But like, it really is happening. And like, eventually that loop will become like, faster and faster. I mean, the Composer thing is interesting because this is where the model, like, in Cursor people like, press tab or they don't press tab on the next completion that the model suggests.

And based on that, every single day Composer gets better at like, predicting the next user.

So that was the old tab model. Like, they actually did the same thing for the actual, not just like, the tab model, but the actual like, generative model.

Oh, that's interesting.

And they, it was, it's hard because when you do online reinforcement learning, you don't have groups,right? You just have one user saying one thing and then you get one rollout. And so like, you have a big variance reduction problem.

Yeah.

And like, Cursor's kind of fuzzy answer to this was like, oh, you know, we have very good heuristics which are able to estimate like, like, what the, how much better than average like, this response was or how much worse than average this response was.

And then they would do like, this big, you know, like, reinforce update. And then their solution to like, whether it got worse or not was like, if it improved on Cursor bench, they would deploy the new model like, every five hours.

Yeah.

And if it didn't, they would like, throw that version out.

Interesting.

Yeah, I think your biggest problem is actually just not knowing what the reward function should be from natural data. And if you use some kind of superficial signal like, did they accept the code, the edit, you might, that might get reward hacked in some way.

But isn't it, this seems like a bigger issue with the sim-to-real thing where the longer and longer horizon tasks get, the harder they are to simulate within a data center,right? It seems to me already, potentially, at least even in coding, we're getting to over the point where there's like, not some year-long coding task that doesn't eventually require you to like, talk to a client or interact with the company or interact with users.

And if you think about the gamut of things we would want AI to be capable at, you want eventually superintelligence to be able to like, run a business or like, start a new business and make it profitable or like, have a profitable day trading in the markets or win a court case.

And these are all things which are very hard to simulate in a data center. Like, an inherent part of the learning there is interacting with the real world. And so maybe, maybe they meta-learn how to get better at these things from like, the transfer between sim-to-real.

But alternatively, maybe you do need weight updates from these kinds of interactions in order to get better at them. And then if that is the case, if transfer isn't strong enough and you do need weight updates, then the fact that the models are quite sample inefficient is like, maybe a deeper problem.

And the reason I'm curious about this, I feel like by default, I don't see how you don't get some kind of crazy recursive self-improvement within the next 10 years. But the one reason why that might not happen is in terms of like, weight updates, the sample efficiency of weight updates, they just seem way far behind humans,right?

Like, plausibly million-fold behind humans in terms of how much data a human sees from birth to adulthood versus how much a model sees from, you know, like, cold start to like, finishing training. And so, yeah, this is all to say, first of all, is there going to be a good transfer between simulations and extremely long horizon, really complicated real shit that we want the AI to do in the real world?

And if not, does that really mean that like, the lack of sample efficiency in these models comes to bite us?

I think maybe the way I'd break down like, the two types of tasks in which models get good and models will like, still continue to struggle is whether the task is like, cumulative or like, you kind of have to, you have this like, non-stationary distribution, you have to keep learning and like, relitigating a bunch of stuff.

So like, maybe an example of a cumulative task might be RSI. Like, it's theoretically possible to maybe like, have a less than a million token like, you know, Python file which like, from scratch, trains a model that is capable of recursive self-improvement.

And like, every discovery that you make is kind of a line in the sand that you hold. Like, if it's true that, you know, for RSI, we don't need to discover a new attention variant or whatever. Like, once you've discovered attention and then once you've discovered, you know, mixture of experts and once you've discovered GRPO, then you just add that to the training stack and like, that's there.

And like, a good example of this is like, you know, 5.6 soul training, 5.6 tera or whichever one OpenAI told it to train. Like, it didn't have to go back and discover attention. Like, it basically probably would have called a bunch of like, scripts which is like, pretraining.sh and posttraining.sh and just did that.

So like, that's an example of like, a cumulative task. I think the real world and the reason like, people are thinking so much about like, continual learning is it's not really a cumulative task. Like, imagine like, in a law firm, you have an agent acting as a legal associate.

Like, that's a very non-stationary distribution. You have to be able to fit in your context like, all the relationships between all the important people at that company which are also changing all the time.

Yeah.

You have like, all these like, implicit like, ways about how things are done, where to find information, et cetera. And like, that's not as clean of an example of a cumulative task as like, RSI is. So I think that there will be this breakdown between tasks.

But you know, like, if the labs realize that and they do believe that RSI is cumulative in the sense that like, we don't need to go back and discover some brand new like, architecture or whatever.

Yeah.

Then maybe more and more effort and compute gets focused on that versus the.

It's so unfortunate that RSI happened to be easier than paralegal.

Yeah.

Yeah. Yeah, I don't know if you guys have thoughts on this.

Yeah, I would say there's, like, models, our today's models are weaker than humans in a lot of different ways. And some of them might have to do with sample efficiency in a certain regime where, I mean, in some regimes, models are very sample efficient, like learning in context.

But then there might be some like, medium-length regime where they're less sample efficient because humans can do some kind of weight update more efficiently than models. So I think like, being less sample efficient in certain regimes might be one of the sources of weakness.

But then I think there are other sources of weaknesses that are completely different than that. For example, having lower diversity of thought than humans or, yeah, being bad at certain kinds of long horizon judgments. I mean, I think a lot of what people call taste is something about behavior that works in the long run and that people have realized works in the long run.

Not everything, but like, some aspect of taste, like, especially for something like software engineering. Like, I think a lot of taste is like, what are the systems that are going to be maintainable and work well, yeah, in the long run of this project.

So yeah, I think the weaknesses of humans which limit

RSI along with other things are, yeah, there's a variety of them and some of them are related to sample efficiency and some of them aren't.

Maybe an interesting thought experiment is like, if you were able to give a model like, a context window of, I don't know, a trillion tokens or whatever you would have needed to fit in like, your experience prior to like, let's say RLHF.

And like, it's got all that experience in the context window and it has the same sample efficiency and in-context learning ability as it does at a million tokens. Like, do you think taste is then solved? Like, would it be able to like, make the same judgments that you did or is there like, something fundamentally missing apart from just a longer context window with the same sample efficiency?

Yeah, I mean, it would have to be trained to learn from that context. So I'm not sure,

yeah, either it would have to be trained to learn theright update to make from that context or.

So you would have to, I think, you could just like, dump it all in. Like, your whole like, life like, research experience.

I mean, like, you still need the data to train it long context,right? Like, even if you could theoretically get like, a trillion context, you would need a trillion lengths of data to train it. Like,right now you have like, take context, you can't just dump it in there.

I'm just asking if you had that.

In theory, I think yes. I mean, this really just comes down to the question of like, how meta-learnable is taste from like, shorter horizon episodes. And like, I feel like there's no obvious reason it's super long because like, humans somehow developed taste with not having many long episodes.

Like, we don't live to be like, 10,000. We have like, you know, we develop pretty quickly,right? And so like, you know, if you think about like, even like, in a PhD, the difference between like, a first-year PhD student and like, a final like, postdoc or something, that's like, five years maybe.

And they've only done like, maybe like, 10, 15, 30 research projects in total. But somehow they develop taste quite quickly from like, a relatively short succession of like, small things. And so like, theoretically, it's, you know, possible to develop it like that.

The AI obviously will have vastly more experience in which to develop taste to like, meta-learn it. And then it's like, how well does that generalize to like, really long horizon things is I think the question, which I think is really unsolved at this point.

Like, we don't know. Going back to this question, eventually there should be a regime where AIs are learning a ton from each individual instance of deployment that they have. Where currently you could say there's some meta-fuzzy process by which models do improve for deployment.

But I feel like it's a very weak, very weak feedback loop. Do you see this around the horizon where there's this like, hive mind kind of learning that's very rapid? And if so, how exactly does it happen?

Actually, I would say that around, will we get a hive mind that learns from all of its deployment experience? I mean, a big part of that is actually about incentives rather than being a technical question. So like, companies aren't going to want to have the model provider learn from all of their deployment because that might just reduce the advantage of their business.

I think that maybe the economics of this will pressure not necessarily weight updates to one big like, common shared model, but like, kind of like, modules that get subbed in. So like, a very obvious example is the lore.

But it might be something else like, you know, there's been a lot of work to try and fit like, an arbitrary context length into a fixed size. Like, this is all the linear attention stuff and all that sort of stuff.

And like, cartridges which are essentially KVKs just trained to be very, very compressed KVKs just to fit in a lot of information. That's another example of like, you know, something that like, companies may be willing to sign up for if that gets subbed into the model and it's not like, actually changing the base on the line model itself.

So like, there's many different versions of like, learning from your data in real time. And like, the latter ones are not really helping the big labs because they are just these modules. But I think the like, the economic pressure will force like, the labs to go down that path first before they can embark on this like, you know.

Which economic pressure there? Because I feel like even if you have like, a bunch of cartridges or lores or whatnot, you can still just like, take all these traces and just like, distill this, stub this to the pretraining of like, your next generation of models.

Yeah, so it may be a more indirect form of learning that the big labs are getting. And that's obviously still really valuable to them.

Yeah.

But I can't imagine a world in which we start off with like, you know, we're going to just like, directly train this one big model on like, all the exact data that we're getting.

No, I think it will definitely like, go through stages. Because I mean, this is assuming there's like, one discontinuous event where it's like, suddenly we fix like, weight updates continuously. And like, in practice, I think it's much more likely to be like, the cartridges and stuff allow you to specialize in deployments.

Then you generate traces, you put that in your model. Like, three months later you come out with a model which is better at this stuff. You specialize it again and you like, consolidate it again. And then eventually we will just like, make this loop faster and faster.

So instead of like, every three months we release a model, now it's like, every week and then every like, day and then every hour. And at which point we basically have obviously solved it.

Yeah, and I think this is a good point as well because you asked like, kind of how far off the current paradigm we are from being able to do this. I think like, we've done a bit of research to this and people have done a lot of research.

Like, at a really large scale, like, when you wash out enough noise and you have large enough patches, like, this outer loop process of like, putting data into mid-training and creating RL environments, like, it does work in like, some sort of continual learning regime.

But the problem is like, when you zoom in close enough at like, a micro level of like, I've got one model and I'm trying to update it again for like, a law firm or something. And I'm trying to do that very continuously like, with a relatively small amount of data.

Like, all the methods kind of break down a bit. So like, if I SFT the model on just like, you know, successful traces, off policy, on policy, like, eventually in the very iterative regime, like, when you're doing like, you know, hundreds of these micro updates, you see both catastrophic forgetting.

You see forgetting of like, previous information I've learned on top of the base model that was much earlier on. And I see degradation of general like, use, general capabilities. You know, on-policy distillation seems to like, push this horizon out a little bit, but it still eventually succumbs to the same thing.

And RL is not very good at like, it is good at like, getting capabilities in, but it's not as good as getting like, knowledge in. And like, just this very explicit knowledge of like, oh, okay, like, this person does this at this law firm and like, this is a very specific process we find.

And you have to pour in a lot of compute to create theright environments to get the knowledge in.

Do you think that the fundamental issue here, why you get worse at these other skills or there's forgetting and stuff, do you think it's fundamentally an issue of capacity or it's an issue of techniques?

A little bit of both. I think like, SFT and even like, distillate, like, on-policy distillation can be like, way too destructive. Like, the reason RL is so nice is because like, yeah, it changes a very, very small amount about the model and there's like, a lot of evidence for why this is the case.

And so like, it kind of just like, tweaks it in this very, very, very small like, loss value to like, get it into theright point. But that also then limits what you can do with RL, like, how much you can actually change the model.

Whereas I'm sorry, you're saying like, the reason this isn't the winner-take-all potentially is that it is just like, very hard to distill that much information into the base model.

Without ruining something in an iterative regime.

Yeah.

Like, it's easy to distill it into like, a different base model. Like, this is why I think it's mostly technique. It's not like, it's definitely not like, just like, there isn't capacity. Like, if you had some modeling with all this data and you take like, literally the same size model and pretrain it from scratch with like, all of this stuff in mid-training, it will be better.

And I think that's a lot of what's happening today.

Yeah.

And so it's very much like, there's, you know, a bottleneck that stops us from just keeping training the same model forever versus just like, getting all the data from the old model and like, training a new model from scratch.

And this is exactly as Charlie was saying, like, some combination of like, plasticity and like, catastrophic forgetting. And that like, you know, if you just naively train on like, non-stationary data because you're adding new data as you go, basically this is messing with the data distribution.

So like, the old stuff is just forgotten.

I see.

And we don't really have good methods to like, stop that from happening.

And maybe at the end of the limit, you're like, just bottlenecked by retraining the model from scratch with all this new information.

Yes, which of course is like, very expensive. Like, training a model from scratch is expensive.

But you're going to do that anyways.

I mean, not necessarily. I mean, like, maybe eventually if you have continual learning, you never train a new model. You just like, just have a model and it keeps learning and like, expanding,right?

But there might be like, some deep technical reason why that's very difficult because of these like, plasticity laws.

I mean, that's the question. That's the question, yeah.

I think we have pushed back like, how much from scratch we need to do. Like, it is definitely possible now to take like, the pre-trained base and like, do very good mid-training on top of that, like, kind of continuously plus some RL from like, different checkpoints that are later on in the training.

And like, that's looking more like continuing learning, but it's certainly not the case of like, you know, take the most recent model, apply a couple of very small updates and like, iteratively like, never lose it. So but isn't this like, I'm a bit confused because isn't this literally what happens during training where during post-training or something, you just have like, you have a model that's already gone through so much training and then you like, distill some fork that's been further RLed or something.

Isn't that literally what happens? And like, why is that?

It's still at a large enough scale. I think that you're washing out a lot of like, the noise. And like, you're not just focused on one distribution, which as Beren said, is like, you know, that is now a very, if you're just like, focusing on one task,right?

Like, that's how much it is.

But sure. I mean, in the eventual regime, you'd be doing, I don't know, there's billions of deployed instances. You're like, doing, you're learning from all of them at once. And so hopefully there's some washing out of noise and stuff from that,right?

Maybe that scale, yeah.

Yeah, I mean, I think like, definitely as Charlie was saying, like, you can do continual mid-training for like, a long time and you can like, well back to a checkpoint, give any mid-training data.

Yeah.

But at the same time, like, you can't do this like, indefinitely. Like, if you just keep continuing mid-training the same base forever, it just like, get, it does, it's sort of asymptote at some point. Like, you can't just learn new stuff in that base.

And this is why people end up training new bases. Like, otherwise you would just keep mid-training the same base forever.

Whenever I finish recording an interview, I immediately brain dump all my thoughts into Slack. Things like, what was the most interesting and what should get cut? This ensures that my editors have all the context they need to start editing the episode.

But it's not like these brain dumps have any clear timestamps and my unedited recordings are many hours long. It can take a ton of editor time to even find the exact moments that I was referencing. So we decided to try adding a Grok Bot producer to our chat.

And now whenever one of my editors posts a rough cut of the episode, Grok Bot opens the transcript on its own computer and starts working, usually before I've even seen the message. It takes the notes that I dropped into Slack and it highlights the relevant snippets in the transcript.

It also uses a big taste file that I've compiled with all my preferences so it can suggest potential edits. And when it's done, it sends me its top clip candidates so that I can review everything from my phone.

This has worked really well. Being able to send informal messages like I'm texting my editor and then having the transcript immediately reflect my preferences has just been so helpful. Try Grok Bot yourself at x.ai/bot. Okay, let's talk a bit about data now.

So I'm generally interested in this question of how much of AI progress is just explained by data progress. It doesn't mean it will be necessarily hard to automate, but that's a separate question. So is there some data distribution which if you trained current architectures on would result in a superintelligence that totally dominates human experts across every single field?

Are we talking about like, pre-training plus post-training data?

Yeah.

Like, environments as well?

Yeah.

Like, I think the existence of this is obvious. It's just like, whether we can create theright environments to get that.

Yeah, I mean, in the trivial case, we could just train it to output the Python file which like, trains the actual superintelligence. Like, just how to memorize it and the weights.

Like, yes, there's probably like, a ladder of RL environments that is possible to construct such that you would get an AI researcher which is at least as good as a human researcher. But the effort to climb each successive rung grows like, kind of exponentially.

And that's going to be the two things that you have to trade off against as to whether like, like, how fast we're going to hit like, that final rung where it's better. I think that's fairly clear. And I think there's like, you know, we're still relatively early in like, RL environment creation.

Like, there's a lot of asymmetries that we exploit in order to create good environments. So one of those asymmetries which we've talked about before is like, it's, there's environments where it's easier to go backwards than forwards. And like, what I mean by that is like, it's very easy to define this like, complex data generating process.

And this is like, this kind of latent variable you keep hidden from the model. You can generate like, arbitrarily complex like, environments and the model has to do a lot of like, irreducible like, token spend and irreducible work to figure out what that data generating process was.

There's asymmetries in terms of like, you know, you can inject information from the real world. So like, Anthropic finds a bug through like, you know, tens of thousands of human and LLM hours combined and like, turn that into a very, very neat environment which a single LLM could theoretically find within like, you know, a few million tokens.

So like, there's all these asymmetries which we're cherry picking and like, we're counting on like, kind of this task horizon generalization. But I think, yeah, again, there's just going to hit diminishing returns at some point. Like, at some point it's diminishing returns and how hard it is to create these environments in the first place.

Like, coming up with them because you can't necessarily just like, have these really, like, these processes where it's easier to go backwards than forwards. Like, you actually have to sit down and construct like, something that looks like, with humans, like, you know, a long enough time horizon.

Like, it's going to be a really complex task to create. And then there's also going to be like, the compute and time bottlenecks for the agent to actually do those tasks. So like, I think you're just going to start seeing this like, curve to flatten out.

I saw something about how someone fine-tuned the talkie model, which is only trained on data up to 1930 on this like, modern coding agent data. And it did better than Claude 3 Opus on Swebench. So this model that has like, no knowledge of code whatsoever can be fine-tuned on a moderate amount of data and like, behave better as a coding agent than this much larger pre-trained model is pretty crazy.

And it kind of shows you that like, once you have an example of like, theright expert behavior, it's actually surprisingly easy to like, copy that into a like, relatively weak model.

Yeah, but a counterexample to like, that kind of is, there was a paper recently where they trained it up to like, fifth grade maths.

Uh-huh.

And like, also like, primary school like, English and stuff. So it was like, a decent language model. And they tried to RL it to do like, you know, late high school and college maths. And the gap was just too large.

Like, they couldn't get it to climb at all. Like, but if you did like, successive rungs of like, you know, year seven maths and then year eight maths and so on, like, you could obviously climb to year 12.

So like, again, it's just like, what is the distance between the rungs on those ladders and how hard is it to create?

Yeah, and this just comes back to like, the RL signal problem. Like, RL is not very good at like, exploringright now. And so like, if the model can't like, get in like, you know, 128 word arts, it's very unlikely to get signal to like, progress.

And this is why like, in RL we need like, curricula. Whereas like, in pre-training we don't because like, it's, that's not a problem for pre-training at all.

Yeah, and again, pre-training data is different to post-training data.

Yeah.

And I imagine as we continue on, like, yeah, humans will be involved less and less, but that doesn't change the fact that your bottleneck on like, how much signal you can extract from the real world. So like, there's a lot of signal in the world and that's true.

Like, you know, there's people doing like, spreadsheet tasks, there's people doing like, legal tasks and all this sort of stuff. But you know, the capability frontier of where the models are at now, like, how many bits in the world are actually like, really relevant to like, improving the model's capabilities?

Like, you know, how many new maths problems are being solved that like, couldn't, they're just beyond the reach or grasp of the current models? Like, how many new coding problems are being created or solved that are beyond the reach of the current models?

And like, I think that's why the diminishing returns kicks in because like, even the world as a whole is not giving you the bits that are useful for tipping you into the next like, basin of capability.

Yeah, I totally agree with this. It's like, really a question of like, where the signal is coming from. And so like, the signal doesn't, you know, in pre-training the signal is like, already in common core,right? Like, for the tasks that you care about in pre-training, the problem is, there's not just like, getting signal at all.

It's like, filtering out all the noise that exists.

Yeah.

And that's quite an automated process. But like, as the models get better as we end into mid-training and post-training, the signal just like, doesn't exist anywhere in the original data we have. Like, no amount of filtering will like, get this, you know, there's no like, hidden proof of like, the Millennium Prize problem sitting in common core.

We can just like, filter until we see it,right? And so like, at that point, you have to get bits some other way, either from humans like, directly like, asking them to like, write out their reasoning or like, by like, creating environments where like, humans decide like, what environment should be created, what the objectives of these environments are, or like, you know, some kind of like, training on like, the human data that exists in deployment.

Like, you have to get the bits from somewhere.

Yeah, yeah. There's a question of how much of the progress in pre-training is being driven by data.

Yeah.

I did this investigation with Jerry Hahn, who's a student at Princeton, where we basically trained all the recipes from 2019 till now, pairwise with all the data sets from 2019 till now. So you're saying like, GPT-2 on the newest data set like, Ultra Fine Web, and you train Delphi, which is the newest training recipe or the open source training recipe on like, the Pile or some old data set.

And you do like, the whole grid and you see the getting to some level of capabilities, how much less compute does it take across this grid? And you see that the data seems to explain like, 9x of a compute efficiency gain, but the architecture improvements explains like, a 3x compute efficiency gain at a very small scale.

And so to the extent that that is true at large scale, that most of the pre-training compute efficiency gains are coming from better data, how much can that continue? Like, can you keep just filtering data more and more and building more and more synthetic data until...

Yeah, do you have a sense of how much this kind of pre-training progress can continue?

I think my prior is that like, again, the low-hanging fruit is like, somewhat exhausted with like, like, we got the internet as this big block and like, it's not like the internet is necessarily like, growing at the same like, rate.

All the useful stuff on the internet is growing at the same rate. So like, we've probably got like, a bunch of like, 0.1% loss drops to go, but like, not, definitely not as many as have currently occurred.

Yeah.

But like, that's also really interesting that like, you know, you find this like, what, cumulative like, 27 times improvement across both.

Yeah.

I think like, it was Epoch or someone who estimated like, three times a year since 2019, which would imply something like, you know, three to the seven, like, over 2000.

Yeah.

Like, times improvement. So like, where's that missing, you know, 100 times or whatever coming from? Like, that probably gives you a good signal of like, how much of this is like, post-training.

I think the explanation has to be that a lot of the compute efficiency gains are scale dependent and we're studying at extremely small scale. And that raises the question of, do the data compute efficiency gains or the algorithmic compute efficiency gains have more scale dependence?

I don't know if you guys have a prior on that. We just didn't have enough compute to investigate that question.

I mean, like, just naively,right? Like, the theoretically, the scale dependence of the architecture is like, fairly well known.

Yeah.

And like, you can fit a straight line to it. Whereas like, I would have no idea how to do that for like, combining pre-training plus post-training data and mid-training data.

I don't know.

I feel like data is actually more like, more important with scale. Like, I feel like architectures are kind of like, a one-time like, you know, an architect...

Yeah.

I think like, combining like, saying just like, an x percent efficiency gain is kind of misleading because like, what an architecture does is like, let you reach like, a qualitatively new regime which you couldn't reach with the old architecture.

And then within that regime, obviously the data is like, the prime we think determining it. But like, you know, if we say didn't have like, even like, GQA, we're doing like, full attentional day, we wouldn't be able to do like, a million...

it would be like, reduce expenses to a million context. And then like, because of that, we couldn't, we could never use the data which is like, actually at a million context. And so we couldn't get these capabilities even though like, if you just do a naive like, how much does this do at like, 2k context where the architecture isn't unlocking anything, then like, the data, you know, that will look much more important than in some sense it is,right?

It's unclear to me that these things are like, really just like, multiplicative gains in this way.

I see. Sorry, but then, what is the takeaway for the scale dependence of data?

So I mean, on scale dependence, I think like, a lot of the like, mid-training and post-training data we have now is like, actually gets better with scale. Because like, a lot of it like, the very long context horizon environment stuff really requires like, big models to be able to like, make use of them.

Yeah, yeah.

And like, this is not, you know, if you try and train like, your 100 million parameter model on like, Swebench traces, it's not going to get anywhere. And like, it's not going to show you the same kind of improvement that you would get if you train like, an actual sensible size model on it.

Yeah.

Yeah.

And like, it's hard as well now because so many of the architecture changes, like, you look at like, Kimi for instance, like, or DeepSeek, they're doing these architectural modifications with not just like, dropping the pre-training loss in mind, but like, for instance, how the models are going to be used in the real world.

So like, the inference efficiency, like, having some form of compressed attention in the DeepSeek models is not necessarily geared around, you know, this is fundamentally like, a further improvement. It's just like, okay, we're considering how the models are going to be used.

Right,right,right.

One question I'm curious about to understand the future is how parameter scaling will go as we're getting into more of an RL heavy regime. Like, I don't know. I don't know how fast historically. Yeah, you can look at sort of open source architectures and see how fast parameters have been scaling.

And maybe it's like, roughly 2x every year for frontier open source models. And to the extent that like, even frontier closed source models have like, 100b or 200b active parameters. Do you think that like, keeps 2xing year over year or now that we're in an RL regime where you also want to conserve compute on rollouts that...

And also, maybe there is like, a threshold effect where you have enough capacity and at that point, does increasing parameters arbitrarily doesn't matter as much. Do you guys have a sense of, in 2030, how many active parameters will a frontier model have?

Yeah, I think for the next few years, we're going to be like, like, because we're so focused on doing longer and longer horizon rollouts for RL where like, inference efficiency matters a lot.

Yeah.

It feels like the models aren't necessarily saturated on their ability to do that. The bottleneck is still the environments. And so we might see like, a little bit of a plateau. Like, I have a feeling that, you know, like, Mithos and the GPT models are much smaller than like, you know, the 10 trillion parameter range that people are talking about.

Even just naively comparing the open source models, you can probably back out that conclusion.

Yeah.

So yeah, probably for the next few years, I wouldn't imagine a huge growth in the number of parameters. But again, like, there's so many different things to trade off here. Like, you decide the size of your model based on like, how much pre-training data you have and then like, the difficulty of the RL environments that you've got to train on.

And you ideally want to like, get to the optimal point where, you know, you can get like, a decent pass at one or something on like, the hardest environments you have. And like, it wouldn't make sense to like, make a bigger model pass there because then you're just paying like, much more inference swaps than you need to.

So there's a lot of inputs to this. It like, depends on how quickly, you know, like, McCore and then in-house these guys can scale up the complexity of the RL environments they're training on.

I would expect the models to keep getting bigger just because people are scaling up compute and the GPUs are getting bigger. But I would say exactly how much they get bigger depends a bit on the scaling laws in non-obvious ways.

So one thing is that I think like, data efficiency is going to be a bigger driver than compute efficiency of like, the exact architectures people use. Now that we're getting to the regime where we're sort of running low on like, high quality pre-training data, so that might affect how sparse you want to make the model.

And then I also think we don't understand sparsity that well. And it's like, parameters are a different resource than active parameters, but it's... And like, sparsity has definitely increased a bit, but it's not clear that it's going to keep increasing without bound.

There might be some kind of sweet spot.

Yeah.

There's an argument that sparsity should make data efficiency worse because you might have to learn the same thing on multiple experts.

So that's debatable. So I think we don't... I don't think we have a good enough theory of scaling laws that we really understand why sparsity is helping and to what extent, how much it'll help and if that'll like, plateau at some point at a certain level of sparsity.

And sorry, can you spell out exactly what the implication of data efficiency would be on... So it sounds like you'd say, well, there should be less sparsity, but what are the other implications on parameter scaling?

I guess just that the scaling law, you're not necessarily looking for the most... You're not trying to optimize compute efficiency. So you have all your choices you can make on the architecture and each of these gives you a different scaling law.

And then like, traditionally you would look at some kind of envelope based on compute. So you would look at performance versus compute and take the envelope of like, the best models. But like, if we're making that decision based on data, so it's like, yeah, we're sort of assuming we can spend a lot of compute and like, we're sort of data is on our x-axis instead of compute, then we just get a different set of optima or a different set of models that are on that frontier.

Yeah.

And I also don't think that we've necessarily like, you know, doubled like, the size of the models every year for the last few years. Like, the people have been training like, one trillion parameter models for at least a few years.

Like, there was even an open source one called Falcon, but like, Liam from Periodic Labs, like, I think posted yesterday on Twitter about how like, an early experiment at OpenAI was like, training a one trillion parameter model that was very, very sparse.

Oh yeah, that was what they did before OpenAI. That was like, at Google, the switch transformer.

Yeah. So like, it was like, you know, very, very good at like, knowledge, but terrible at reasoning because it was so sparse. And so like, you know, it feels like we've been playing in this like, 100 billion to, you know, up to 2 trillion parameter range for like, at least a little bit.

And like, it certainly hasn't been. This is like, nice linear increase. Yeah.

I mean, I feel like there's two things. So as Charlie was saying, like, inference efficiency is super important for RL rollouts. And so like, this will really push down active parameters quite a lot. And then I think the total parameters really depends a lot on the hardware as well.

So like, you really need to get like, very high memory bandwidth and like, VRAM size to like, actually be able to serve like, multi-trillion parameter models. And so like, you know,right now, you know, people still using a lot of like, H100s and stuff.

And so as everyone moves to GPUs and then their rubens will get like, more actual like, the ability to scale and like, actually serve and like, do like, large RL inference at like, different, at larger scales. The data question I think is interesting because naively, like, larger models are much more sample efficient in like, the actual data points.

And so like, even if you're like, not saturating the model, it's still better to go bigger because like, the models will, larger models generalize better and like, get to a better loss for the same amount of data. And soright now, I think we kind of have a lot of data and like, that's not the constraint rather than compute.

And so we're having like, smaller models which are like, very inference efficient. But if compute is no longer the bottleneck, it might come back to larger models which are like, sort of undersaturated, but like, they have this generalization ability because they're much larger.

If you just look at like, the basic Chinchilla scaling law and you just maximize out parameters.

Yeah.

It actually decreases the amount of data you need to get to the same loss very little.

Yes.

If you go to infinity on parameters, the amount of data you need, I think, goes down less than 10x just because of the nature of like, the power law.

Yeah. We're now on the way too much data side of the Chinchilla laws,right? Soright now we overtrain models according to Chinchilla. And so we could easily get back to a point to which, as we're running out of data, we move back to like, the Chinchilla optimum point.

We're even like, a bit on the overtraining, like, you know, under-training model side.

But surely like, even with these new chips that come online and stuff, like, we're just going to be so compute bottlenecked for the next few years that that won't necessarily be the case.

I mean, this could well happen. Yeah, this depends on like, the ratio you have like, training and inference compute really. It's like, if you're super bottlenecked on data, not on compute, you should go bigger. If you're super bottlenecked on compute, you should always go smaller.

And then like, yeah. But you can also use compute to generate synthetic data. So it's like, one of these very hard things to predict.

Yeah, I think part of the reason it took people so long to figure out the scaling laws in the first place was that if you don't get all these thingsright, then you don't get such a clean relationship. And like, the beautiful stray lines on graphs like, hide a lot of complexity on how you have to make sure like, to scale every hyperparameter theright way or like, parameterize your optimizer in a way that scales and where you don't have to change your hyperparameters as you change the model size.

And bugs have their own clean scaling laws as well,right? Like, you know, like with Kaplan for getting the cosine annealing thing or like, even just like, not considering embedding parameters, I think. And so that messed up the estimate at smaller models because embedding parameters are a decent size of the model.

A bit on RL. So I feel like a year ago, a lot of people were making this argument that RL will not be super successful at scaling for models. I think John, you wrote a research paper where you were pointing out that models learn one bit per episode.

When you RL, they basically learn, did I get the answerright or did I get it wrong? And I wrote some blog posts earlier this year. I was like, it's even worse than that because when the pass rate is low and the model is very unlikely to get the answerright, it learns almost nothing at all from an RL episode.

But I look at the models today and they seem pretty smart and it seems to be the result of scaling up RL. Beren, you had a post, I think, a few weeks ago where you were trying to explain what's going on.

But why has RL been more successful than one would have naively thought?

I mean, so I think the success of RL comes down to a bunch of different things. So first, I think what is slightly underestimated is actually the mid-training. So an awful lot of like, what we see as successes of RL actually comes from like, very, very good mid-training data, which is basically where we're like, essentially doing pre-training but on like, synthetic reasoning data and like, the kind of environments that like, get the model warm started for RL.

And so this actually takes the model like, almost like, 80% of the way to like, the final RL checkpoint often. And then what RL does on top of that is it does like, a lot of, you know, essentially tweaking to the policy.

And so this is one of the reasons why it doesn't need like, as many bits as you would naively think. It doesn't have to learn all of these behaviors from scratch. It needs just like, a few bits from these episodes, which you do get.

And then the other thing that I really point out in my blog is that these bits are actually extremely high signal compared to like, regular like, pre-training, which is why you need RL at all versus just like, SFTing on like, successful reasoning traces.

Maybe because it's exactly the bits about how to get the answerright.

Well, there's two things. So yes, one, it's exactly the bits about how to get the answerright. But like, this is not exactly how you think of it because in SFT, you have a trace,right? You have like, a bunch of math reasoning and then the answer at the end.

The bit is still there. Like, you still SFT on the answer token. So that bit is still there. What's important is that the objective ignores all the other bits. So in SFT, you like, have like, you know, to try and match like, the exact reasoning tokens that the model produces.

So you're essentially getting like, too many bits about like, the exact way this other model you're training on reasons. For RL, you only get the one bit. And that means that like, that signal is not drowned out in the noise of like, all the other bits the model has.

Yeah, yeah, yeah.

And so that's what really like, it's really a super dramatic like, increase into the signal to noise ratio during training, which is why like, RL is like, so dramatically efficient in terms of steps.

I don't know if you guys have thoughts on that.

Yeah, I like, there's been so much debate about like, what RL does to the model versus like, you know, mid-training or SFT or whatever. And like, you know, everyone talks about how, you know, pass at one will go up, but pass at 256 will go down.

Like, very rare correct reasoning traces will be like, downweighted and kind of like, outweighed by a gradient signal from like, easier kind of reasoning traces. And I think the simple like, way to view RL now is that if you have a large enough like, a large enough amount of compute to sample a large enough group size such that your probability of getting a bunch of correct answers is like, pass some like, not insignificant probability, then like, it will be upweighted.

And like, to Beren's point, like, basically mid-training and, you know, more pre-training, like, the pass at one, the starting point for RL like, scales in a log number of pre-training tokens.

You can answer very basic questions. I guess that answer makes sense and maybe there's empirical research which shows that this is what's happening. But then I just look at the models themselves and I don't know what's happened. Like, yeah, maybe you can give me a sense of what is the basis of the AI progress over the last year.

But if it's, yeah, maybe it's just upweighting the policies which were going to do the correct thinking anyways. But it just seems like qualitatively, the models have gotten so much more capable. And anyways, maybe there's nothing to, there's no inherent contradiction there.

But how do we square like, the relatively small impact this take would imply that RL would have from the actual qualitative capabilities the model seemed to be gaining?

So like, one thing I want to point out here is that like, it doesn't necessarily imply that RL has a small like, effect,right? Even if you have a few bits and like, you only change the parameters a small amount, like, the actual impact on like, function space, the model and it's like, the input to output mapping can still be like, super dramatic.

You know, even if it's like, even like, one bit can change like, your function space a lot and it can like, rule out like, half the hypothesis space, which is huge. So like, I don't think it's necessarily the case that like, small amounts of bits, small amounts of RL, once you're starting from a really good point, means that like, you don't have dramatic impacts in behavior, at least like, not necessarily.

I think it comes down to two things. I think the first thing is that everyone was hoping that RL would like, generalize this reasoning across like, all these different domains. And I don't think we necessarily got this like, horizontal generalization.

Like, just training on math doesn't necessarily make you the greatest coder. Like, you do have to do RL on code environments. I think what we did get though is like, horizon generalization. Like, the models just learned how to use more tokens for longer and still make progress on some sort of task.

And so like, you can train on environments where they get longer and longer and longer and then put them into a completely new environment. And yes, like, they may not have generalized the reasoning patterns which allow them to do well in that environment, but they've at least generalized the ability to like, continue on that task for longer, which is correlated with like, success.

I think there's a paper called Edge Bench, which showed that the rate at which models can work for longer is like, doubling every three months. And so that's a clear evidence of generalization. And I think like, the final way to think about it is like, in pre-training there's this idea of like, quanta.

So you have this very smooth like, pre-training loss curve. And when you actually look at what's happening in the model, like, the model is learning all these like, very discrete like, tasks and there's like, all these like, emergent points where there's like, kind of a phase transition.

Like, it didn't have induction heads, now it has induction heads. And there's like, tens of thousands, millions, probably like, hundreds of millions of these things. And you average them all together and you get this very like, smooth loss curve.

I think like, to an extent, like, a similar thing is happening for RL. Like, there is this very slow outer loop, as Beren mentioned, of, you know, we will train a model and then RL it and then like, the next kind of model iteration of training, we will dump a bunch of these synthetic reasoning traces into the mid-training data.

Like, we're kind of hitting all these quanta for all these different tasks. And like, on an individual task level, it may look like a phase transition and like, you're suddenly going from like, a 0.5% pass rate to a 90% pass rate on like, a particular like, finance task or Excel task or whatever.

But you average all these things together and plus the horizon generalization, you kind of will go out, wow, we've got like, qualitatively better models.

Yeah. I mean, I think a lot of this as well is just like, I think RL does generalize a bit. Like, suddenly you get like, some transfer between like, math and code or like, puzzles and math and this kind of stuff.

Also just like, the share amount of environments I think the people are targeting is just like, vastly greater. So like, you know, before, when you tried to do, you know, some task which like, you do in your daily life, like, two years ago, like, the labs wouldn't really care about this.

They wouldn't like, train the model for it. And now like, it's just so much broader. They have a lot of environments targeting this specific thing.

Earlier in the conversation, we were talking about RL in the context of causing this entropy collapse or just, you know, concentrating probability on solutions the base model were already done and causing relatively sparse updates in the policy. But when I think like, when I think about, I think there's also another story about RL, which is going back to the Atari games and then AlphaGo coming up with move 37, the super creative move that because it was never initialized on human data, it can like, think in ways that humans are not even thinking and come up with extremely creative solutions.

Yeah, do you have a sense on when we should expect or if we should expect RL on LLMs to result in things like move 37, just extreme creativity, even beyond human creativity because like, there's just de novo, de novo initialization of intelligence?

I mean, so a couple of things here. Like, first off, I think that the AlphaGo is using MCTS, which obviously does like, more exploration and like, stuff than regular policy gradients. But I kind of also think that like, RL doesn't necessarily like, reduce the creativity.

And like, I mean, even if we look, I think, you know, this is obviously qualitative, but if we look at like, the, you know, the OpenAI Hugging Face incident, like, these models were coming up with like, multiple zero days at a time to like, break out of the sandbox.

And like, this is clearly like, some level of like, move 37 creativity, I think, already, which we just get from just like, the general generalization properties of the LLMs. Like, I don't think it's definitely not the case of like, RL is like, totally destroying their like, entropy.

Yeah.

Especially on long horizons.

Yeah, I mean, one thing that people call creativity is just solving hard search problems. So, and so that's like, like, move 37 is obviously an example of that or like, writing some kind of poem that satisfies a ton of different constraints.

So that's something AI is obviously going to be extremely good at if trained for it. Then there's another way in which the models, like, the diversity of their outputs is a lot lower after RL and they sort of develop these tics.

And like, even though the models seem like they're good at writing, when you do some kind of like, distributional analysis, you find that like, they're reusing certain themes like, all the time and they're using the same character names all the time.

So there's actually, it's not like you're getting the same kind of diversity that you get when you, like, from human authors. You're sort of getting one really good like, style. So I think that like, that kind of diversity has definitely been like, cut down by RL a lot.

And in fact, oh yeah, since we were talking about distillation earlier, that's sort of something, yeah, one thing that's happening is that so many people are distilling mostly from Claude that like, like, all the open weight models write the same way as Claude and use the same, like, have the same tics.

So this seems kind of concerning to me that we're having this like, this monoculture emerge.

Yeah.

Again, I don't think this is like, fundamental to RL as like, a method though. And same with distillation. Like, even with distillation, like, you're just training on the data. It's like, just because your data is not like, super broad, that doesn't mean like, the training method itself is somehow wrong.

It's like, a problem with the data. And I think a lot of, for instance, like, the RL like, entropy collapse is basically due to like, exploitation of fairly simple like, verifiers when you don't have like, a huge diversity of environments.

Because like, for instance, like, the writing, I think the writing is presumed to be graded by some judge and like, the judge has some specific tics and like, the model is learning to award hack the judge and that's why like, it collapses.

But like, this is really a problem with the judge. It's not a problem with like, RL in general.

Okay, super rapid fire predictions about the future. So I want timelines on the following couple questions. By when do we have models which you can, here's what it feels like to a user. You basically hire them as a drop-in remote worker for all kinds of white-collar work, not just coding, but I don't know, video editing, law, paralegal, et cetera.

Like, it's like, literally an actual remote worker with like, full computer use with like, literally a month of seamless learning and operation and executing on like, complex projects and it required interacting with other people, et cetera, et cetera.

It's like everything a human worker could do over a month.

If you like, mandate it to use like, a browser or whatever rather than like, these, again, the firm setting up the information to be like, programmatically accessible, like, maybe a couple of years. But if it's not like, browser-based, like, it can send Slack messages, it can do all this stuff.

I'd still probably say around a year.

Yeah.

Yeah, I mean, I would say maybe like, for the like, full generality, maybe like, three years. But I think to Charlie's point, we will end up with like, a lot of people like, making their organizations easier for the AIs to use.

And so you get like, 80, 90% of the way there before that.

So but the thing that's, the death between one year and three years there is just literally like.

Like, I think there's going to be like, a long tail of like, miscellaneous stuff which like, some humans can do which like, will take the models like, quite a while to do.

Yeah, like, you mean everything of sort of computer use stuff or like, basic cognitive capabilities?

I mean, I think this really comes down to a question of like, how quickly can we solve this kind of like, online learning and like, whether we can like, get like, 80, 90% of the way there with like, compaction and like, writing files to yourself and stuff.

And like, that's my big uncertainty. I really don't know.

And another, like, maybe an example of something that it wouldn't be good at is like, you know, if I have to like, yell at someone to get something it would like, really push someone to get something done. Like, the model isn't just going to do that.

It's just going to be too nice.

Yeah.

Yeah.

I'd say there's a wide variation in quality of human remote workers. So if you try to hire someone like, off of Upwork to do a software engineering project, there's going to be a huge variation. It's like, often quite hard to get them to do like, to do a good job or like, pay attention to all the feedback you're getting.

And like, I would guess that in some cases it like, it'll be worse. Like, the pre-AI version of this was worse than what you can get now from existing AI. So I think it might end up being a little complicated because maybe to some extent we already have this like, for some like, not so high quality of work.

But then like, then it's obviously like, we're not, yeah, we're not matching human level in certain like, higher quality like, forms of work. So, but I basically agree with Charlie and Beren that maybe, yeah, we'll, yeah, we'll have some version of this in a year or so that's like, okay.

And it will be able to do, maybe we'll have that form factor and it'll be able to do some things really well, some things not so well. And things will be improving from there.

Like, we shift the goalposts based on the very long tail all the time. Like, I think, I feel like you've used this example before of like, doing your taxes or something. Like, this year I literally just like, told Codex to like, go get everything I needed to do and send it to the accountant.

And like, there was this massive list of stuff it had to use computers to click through and like, download some stuff. And I don't know, it was like, fine. It was perfect. So like, I don't know, a lot of this stuff it can already do.

Yeah. Okay. Give you 10x total productivity uplift. Basically, if you, if it takes you a year to make a breakthrough now, you make a breakthrough every month.

I think I would just refuse to give you a scaler on this. Like, like, we might already be past that in some like, types of work. Like, let's say you're just trying to prove, you're trying to do like, certain types of math.

Oh, sorry, sorry. But for you as AI researchers trying to make like, advance, you know, the state of AI research. Because how much are like, AI researchers sped up?

Yeah.

Or uplifted?

Somewhere between five and ten years.

Oh, really? Okay, that's far away.

Really? You think it's longer than like, full general remote worker?

Yeah.

Interesting.

I think you're probably, I think I'm realizing you probably have very different definitions of fully general remote worker. I could have specified that.

Yeah, probably. Yeah, this is true. Because I mean, like, yeah, because obviously like, an AI researcher can be a remote worker and so like.

Yeah, no, I'm picturing like, you know, normal white-collar work over the period of a month.

Yeah.

I think it starts to diverge a little bit past a month.

Like, a very competent white-collar worker.

Yeah.

But not necessarily like, a super creative researcher.

Yeah.

I would say like, two years.

Two years?

Yeah.

10x? Okay. How about you, Beren?

I can kind of see that actually. Because like, it really is just like,right now it's already like, definitely more than 10x of like, coding stuff. And so it's like, if it can do even like, one or two loops of like, experimental feedback, that would actually be massive already.

So 10x uplift of AI researchers within two years. If you just plug it into like, a very naive model of like, AI progress and how much is coming from AI researchers and they're like, there's like, a 10x increase in their productivity.

Yeah, you have like, radically accelerated pace of AI progress starting two years from now.

Yeah, I mean, I think like, this will mean that AI progress doesn't get bottlenecked on like, AI researchers' ability to run like, small experiments that gets bottlenecked on other things.

Of course, of course. But it just like, happens 10x faster.

For sure, yeah.

Which is a huge deal. And that also like, helps the next thing, which makes, gives you 100x speed up happen sooner, et cetera.

Yeah, I'm happy to stick with longer on that one.

And what's like, the crux?

Like, my capacity to absorb information and make the like, Bayesian optimal decision on the next experiment.

Yeah, makes sense.

Yeah, I mean, I'm assuming that like, you can delegate some of this to the AI. So like, the AI is becoming decent at like, deciding, you know, it's run this experiment, it's got this result, it runs like, the next experiment.

And then if it can run like, two or three experiments in a row without like, crashing, then like, that is actually a big update in like, uplift.

And okay, final question. An AI which is, which dominates top human experts across every single field of work that can be done over a computer. So not only AI research, but all cognitive work. And not just like, short horizon work, but like, literally if it takes like, three years or something, the AI will still do better than humans.

This is basically just like, ASI?

Yeah.

Okay.

I would say like, three or four years.

The fuck?

I mean, that doesn't seem wrong.

I mean, I would say like, like, AI is obviously being more, getting more attention. So it's like, one of the harder things, but it's like, a lot of energy is being put into it. And it's also like, not one of the hardest things for AI because it's like, involves a lot of code and math, which models are really good at.

Maybe for things that involve like, 3D and like, spatial stuff and physical stuff, I think that will take a little longer. So, especially if it's not like, yeah, if it's like, mechanical engineering or something and it's not getting like, the most attentionright now, that might take a little longer.

But it also has to include fields where there is relatively little data because of the nature of the field. And it has to like, learn that data on the fly. So, for example, it has to become superhuman at like, being an engineer at TSMC or something.

Oh yeah, so you would have to assume that like, the on, like, the onboarding, yeah, you can give the AI the same onboarding material. And oh yeah, then there's some like, something has to be solved about like, sort of longer horizon learning or yeah.

I'd say five to ten.

So basically, yeah, you think automating AI research is like, ASI complete or something.

Yeah, I think so. Yeah, I think there's so many things in the world which like, even if you have some sort of memory system external to the model, and even if like, context length grows a little bit, like, there are just fundamentally things like, even if you could research the information or write notes to yourself, like, you'd need more than a million tokens of context to be able to do.

Yeah. Yeah, I mean, I kind of agree in like, the five-year range, at least for like, the stuff that like, labs are focusing on. But I think like, there's going to be a long tail of stuff which like, the AI could theoretically go out and learn about, but like, no one has bothered to do it.

And like, the compute hasn't been allocated to that. So that might take longer for like, literally every single human expert.

Yeah. And sorry, but by this, I also included like, the ability to learn as fast as a human, a new domain.

I mean, I think that's not necessarily necessary, actually. Because like, the AI will have vastly greater experience than like, any human.

Right. Thanks so much for doing this, guys. I feel like this was a great format for getting different experts to disagree and debate and discuss things together. It was very productive.

Cool. Thanks for having us.

Thanks.

Absolutely.

Thanks for having us.