RSI 之争、中美差距与锯齿状能力：Interconnects × Epoch AI 对谈全解读

I’m here with JS Denain, who is a senior researcher at Epoch AI. He leads the insights team. He is one of the people who I feel like I get the best feedback on my writing from, whether it’s from US-China AI capabilities, now RSI. And I just wanted to open this discussion and honestly go deeper with him, trying to understand how he thinks about these various things. And I think you have a very useful, moderate point of view, which I feel like you’re probably a step further into what would be called faster scenarios for AI progress. But let’s get into this, and it’s like, what measurements do you think OpenAI and Anthropic are seeing when we get all these proclamations on RSI happening very imminently?

Yeah. I think, so there’s the measurements they’ve published, right? So, OpenAI and Anthropic both had blog posts, I mean, Anthropic two at least, on the effect AI has on accelerating AI progress. I think at least the things they publish, I don’t think are super strong evidence of imminent self-sustaining acceleration AI capabilities, or full automation of the job of AI researcher. But I think the kinds of things we see are, I think probably the most striking thing I saw in the OpenAI blog post was increasing usage of AI systems in model deployment, like the increase in spending on Codex that we saw. And it’s kind of unclear how exactly to interpret this, because maybe it’s a measurement artifact where they’re only looking at Codex, but in fact, there was a bunch of ChatGPT usage before from the researchers. But overall, that plot, for example, just shows a 2X a month increase in Codex spending by researchers, and that does seem to me to be some evidence of they’re getting a lot of value out of this probably. I don’t think this is strong evidence that in six months we have a software intelligence explosion.

Do you think this is the same? So, what is the information they have internally relative to what we have? And this is obviously hypothetical. We don’t have this internal information. Because I get the sense that a lot of people are more scared in their updates from the labs than the information we have. And I try to take this very seriously of, what will they be seeing that is making the acceleration of risk comments go faster, and how much of this is material evidence versus how cultures evolve over time? And I’m much more interested in evidence.

Yeah. So two things. I think, first of all, I guess I don’t think, I don’t know, right? I don’t have full information here. I don’t currently think that either there’s some specific thing that people at OpenAI or Anthropic are seeing right now that we don’t have access to that warrants being way more freaked out about this. I also don’t think that... I think the public evidence we have right now, more general on AI progress and just a priori case for this being an important dynamic, I think is enough. I think to care about this particular dynamic of AI accelerating AI progress, that being a big deal and worth tracking. And then it’s kind of unclear what the urgency is of when the feedback loop really kicks in. So, basically on the what is there on the inside that people have access to, I could give examples of kinds of metrics, right, they could be looking at. It’s plausible that we have access to the capabilities of AI systems, but internal teams have their KPIs, and maybe they’re seeing compute multipliers in the pre-training team or other kinds of metrics that people are tracking going crazy. And then the combination of this plus some intuitions of how the different outputs of different teams combine yields a prediction on the trend in actual performance of the end AI systems. So that could be an early warning sign. It’s unclear to me that the recent discourse we’ve seen is evidence of things going crazy on those metrics.

And how do you think of the link between RSI and existential risk? So I would posit that you agree. I think that there are very real risks of AI, and I’m curious on how you think these, what I would describe as very, very early measurements change anything on the scope of risk. Because I don’t think if you had asked people six months ago, it would be as immediate to x-risk among people are very reasonable. I think there’s more people that are reasonable talking about x-risk again, which was a little surprising to me.

Yeah. So, okay, my sense is something like... So personally, I feel very uncertain about this, but I do feel, yeah, basically bought into there’s, I know Evan Hubinger was like, at least 10% of x-risk within I don’t know what timeframe. I think I’m like, yeah, I know, and this seems pretty reasonable over a decade-long timeframe. I just feel extremely uncertain about it, but I’m definitely very worried about this. Now, why am I worried about this, and where do I think the disagreements come from? And then how do I relate this to the early sense of RSI? My sense, I’m kind of a capabilities theory of everything person. I think, and I think some people disagree here, but I really think that principal component of disagreement between everyone is how huge do the capabilities get, how soon, of AI systems? And I sort of agree that there’s other factors that come in play for how big your economic growth gets, also depend on the diffusion you get. And you could possibly you could think that capabilities are going to get crazy, but the AI system’s going to be just aligned and benign and stuff, and so there’s no huge risk. But my sense is concretely, when I look at the main kinds of disagreements between people, most of the people who I see who are very skeptical of those most extreme scenarios... I think just expect capabilities to not be as huge or as I think folks who are—

What does being a capabilities maximalist look like in a few years? Because I think I’m probably on the skeptical side, so please continue.

I think it looks, for example, something like the AI 2027 scenario, right? I think it looks like the mechanism for this is AI is automating the AI research process, I think, and that’s a reason to pay attention to it. But in terms of effect on the real world, I think it’s like massive progress on robotics. I think a huge industrial explosion, AI systems are just managing factories. You have this kind of self-sustaining economy that just is able to make a large scientific progress much faster than you would have expected. And so I think concretely, the kinds of disagreements I would expect are on, yeah, if you have AI systems that are both very intelligent in the book smart sense, but also have been trained to have more affordances and use them astutely, have been trained to kind of manage projects in efficient ways and stuff like that. How big are the real-world bottlenecks to making very fast R&D progress or getting hard power over humans?

Yeah. Can we go into some of these in detail? Did you listen to the Dwarkesh podcast with Charlie, Beren, and John?

Yes. Yeah.

Yeah, because they had at the end, they had this section on various capability levels and timelines for getting them. And I feel like I agreed with most... I was very in agreement on the distribution they had up to this, and then was surprised by the timelines. And one of them was the 10X productivity for the AI researchers. And I think Beren and John were faster than I think. And my kind of statement is that I think the cycle from of having an idea and doing the experimentation to test it, I agree will be 10X faster very soon. But I don’t necessarily agree that I would say that AI researchers will be 10X more productive in net, which I would describe as the pace of the field’s complete understanding. And understanding is a different axis from just continuing to scale models. I think that’s one of my core confusions on the AI research side. So I’m just kind of curious how you think about this type of thing and how you might specify a 10X improvement in AI research into subcategories.

Yeah. So maybe there’s a scale you could have here, which is the end thing that you might care about is how much faster is AI research overall? Or how much faster is Anthropic’s overall output? And then Anthropic as a company is producing some things, and it’s doing in one year what it would have taken it 10 years to do. And that’s pretty different from individual researcher productivities, where I think you could... So if most of what AI researchers right now are doing is this loop that you were describing, then it’s possible that you get a 10X productivity improvement for the median researcher based on the tasks they’re doing right now. But first of all, that doesn’t mean you get a 10X productivity improvement for all researchers. And even if you did, right, there’s other bottlenecks that hit such that that needn’t convert into a 10X productivity improvement for Anthropic as a whole, right? You could have all the researchers be 10X more productive, but because of compute or other things, the company itself still doesn’t move as fast.

I think an analogy I have is, I think that junior PhD students will be 10X as productive, but from the advisor’s perspective, their research agenda will not proceed 10X as fast. And it’s like to the extent that that contributes to Anthropic’s progress is another hard thing to jump on AI capabilities, where I think that listening to the Noam podcast. This is me, I’m just thinking, was thinking about this when writing about this, is there’s such an amount of inference compute coming online, and I think Dwarkesh highlights this very well, that it’s very hard for me to disambiguate massive speed-up in AI research from the fact that we have way more compute and can now much more effectively spend it on related problems. And I think we’re going to get all of this at once.

Yeah. So definitely I think this question of... So when you look at the OpenAI blog post they had on their acceleration, right, they do point out huge surge in Codex spending from researchers. Interestingly, actually, Codex spending in other parts of the company kind of had a huge surge in the spring and then kind of plateaued in the summer. But for researchers or the data team or engineers, it keeps growing, even accelerates sometimes. And so they point this out, and then they try to look at where there was an increase, which kinds of tasks had an increase in usage. And a lot of them are engineering tasks. There’s an increase in troubleshooting tasks. But they definitely point out that for a lot of the high-level strategic decision-making, they both anecdotally and also when they look at sessions, don’t seem to find a huge uplift in making better compute allocation decisions or deciding on research directions. And so to me, that’s pretty similar to the PI case. And so I think there’s a first question, which is, how much of an improvement... Imagine you just didn’t get that much AI uplift on that component of AI research, but the rest of AI research really went crazy. Then how much faster do things go? And the separate question is, I don’t know, how hard is this strategic decision-making? Can’t you just have a bit longer horizon RL? Or maybe you can bet on decent transfer from other fields where those kind of decisions are important, and then you do actually get that kind of uplift at the end.

I think part of my intuition is that science will look so fundamentally different that it’s almost hard to put a number on it. And it’s the pre and post-AI era, and we’re just in the rapid transition to what is a new method, new way of doing science, because I think all the conferences are ready to burn down and struggle through the next few years. I hope that they collectively figure out a way to like add AI oversight into reviewing and things that are scalable, because they have so many slop papers that they need this type of gate. So I don’t really know. And I think on the capability side, I’m like, what an AI progress is like clearly translated into new capabilities. There are two things. One is like pre-training scaling laws. Our loss is proportional to like an exponential increase in compute. And on the research side, what we are doing is we’re shifting the line so that it has a better offset and potentially a better slope. And then on the other side is the RL environments, and I think the RL environments we’re building now are very comparable to valuable work. So I expect the AI models to get much, much better at like knowledge work that can be scoped. But I don’t know if we have a good process for like churning out an order of magnitude harder environments, which would be closer to like cure cancer, solve these open math problems. I think math is a case that we could talk about. But that’s kind of like, I think there are unknowns on scaling the like raw intelligence more than efficiency. So I’m very optimistic in scaling efficiency.

Yeah. I agree that like in some sense, right, like inference efficiency is like a very like hill-climbing task. It’s like pretty well-scoped. And yeah, so I mean, it’s already something that has very, very fast trends, but I could imagine those trends. Yeah, I imagine those trends will go even faster. It seems really like the kind of... I mean, indeed, we have evidence from OpenAI, right? Like, saving on like serving costs, et cetera, through like building better kernels and like if you’re new to this. They don’t give that many details, but that’s already happening. One thing I’m curious about actually in your case is like, so here’s one way of defining like Anthropic, for example, like accelerates overall, which is you could look at the like ECI trend in like, the ECI of the best Anthropic quality point in time. And you can like look at the current trend line and you can ask the question, like over the next year, will we see a 5X acceleration? Like, will the slope be like 5X larger than it was, say, in like 2025? And I think it’s like pretty likely we see like a huge increase in this because it is like kind of a legible like KPI that like, I mean, it’s not literally KPI, but it is a KPI that like the company is aiming for, modulo like safety considerations, et cetera. And I’m curious about whether you think it’s very unlikely we get this or whether it’s more like we might get this, but like if we do get this, it’s mostly that like ECI has been Goodharted as a metric and like the implications for like real-world capabilities aren’t that huge.

I wouldn’t be surprised if we got this, but I think that it’s going to be like we’re on a slope and then we could get an uptick in slope of hill climbing, but then we like saturate what we know how to hill climb and then it goes to be lower. So it’s like all the things that we could measure I think are going to be getting pulled up very quickly by being measurable. And then we’re in the domain of like, how do we measure it? Because I think, like I talk to people that are trying... Like evals are so expensive to build now, and I do think that evaluations are going to be like how good and efficient it is at coding, how good and efficient it is at ML research, how good and efficient it is at knowledge work. But I don’t know how to like... Building those evals all seems tractable but hard. But then like how to make a breakthrough in fundamental chemistry seems really, really, really hard to measure. I was going to draw on like maybe frontier math as an example, but I think math is such an exception as like one of the most jagged pieces of AI. I think especially like open problems in mathematics are like the perfect target for rapidly improving AI because it’s like a falsifiable thing. And it’s like if we were to, say, see that in something that’s much more open-ended, I think I would update a lot. Or if the labs were like to come out and say, “Using Claude, we have a very, very big change in what our architecture of AI is,” to like there’s the famous like Jonathan Frankle–Sasha Rush bet, and it’s like, and the transformer is no longer like the lineage we are on. I think any of those things being very AI-driven would make me update a lot. But seeing more math, like I think I was surprised by the pace of math, but like not astonished.

That’s interesting to me. I definitely agree with this general sense. So like I think METR folks looking at like your nanoGPT results from autoresearch-style things compared to like what the humans were doing, it does seem like there’s this, I think Tom Cunningham calls this like the apple-picking model where AI is like much more efficient at the start, but then doesn’t actually like uncover as many new ideas. And you see this in this kind of optimizer research. Yeah, I mean, one thing I will say on this like verifiability point is like, I think a pretty common trend is like you’ll have some task that’s like not verifiable and you’re like, maybe you struggle to build an environment for it. But actually it’s like it’s a subset of a larger task that is itself like verifiable. It’s just like longer range. An example of this is like there are many like hard to verify tasks out of like companies. But in some sense, like revenue or like other like metrics, like valuations are like pretty legible. So that’s like one thing. I mean, the other thing is like expect things to be pretty jagged. But I think a big question is like, yeah, can you get, for a crazy world, can you get like a large, like self-sustaining industrial kind of explosion?

Yeah. Well, can we talk about robotics and industry? Because I have a background in physical robots and like I think the robotics trends will look much closer to self-driving cars than LLMs. And I think that a lot of the singularity arguments are based on robotics being able to look much closer to LLMs than the self-driving cars roll out. So, why would you disagree? Or, what is the argument that mass industrialization and robotic expansion is doable? Because my prior is so suspicious that I maybe even haven’t given it enough justice, but I’m very suspicious of this being a viability, and mostly in terms of being a relative timeline. I think it could happen over decades, but I don’t think it’s a two to five-year concern.

Two to five years seems rough, to be clear. I think I just don’t know as much about robotics here. I think is your main concern just reliability is really rough to get right in the same way that it was just a long tail of scenarios where things are, or was it more like a real-world thing where there’s much more regulation that comes up?

I think it’s building things is hard. I think that, let’s see. I’ll talk us through some of this. For example, I know places like Amazon, they build new factories to be robotic first, and those are more effective for them. And what this would take then is building a robotics factory. In the case of the US, it’s like you have to build a robotics factory that builds robots very efficiently in the US and then transition that or make a new one that is built by said robots. And I think the re-industrialization of the US is something that I think is like, there’s a lot of reasons why it is not happening. I think potentially in China it is more likely, but I also just haven’t been convinced by AI results on visual and action models that they’re progressing fast enough. I think I’ve had discussions with people in the multimodal field have described the techniques as being much more rudimentary and less developed than the text language models, and in need of much more fundamental innovation, where something like code plus RL is a very natural match that the hill climbing is very predictable. So—

So, it seems like there’s two things. There’s the trends in robot capabilities is not as fast as you would expect for LLMs, and also even if robot capabilities were huge, it takes a while to build factories. I think I’m sort of skeptical of the second one. I’m just like, if robot capabilities are sufficient, the total addressable market for this is massive. And if you look at data centers in the US, there has been extremely fast build-out. If you had robots that were just literally able to substitute for blue-collar human workers, I feel like the financial incentives would be huge. And I think a lot of the reason why in some cases, the US doesn’t have huge build-out is just a demand thing. I think that’s the case for power, for example. So I think in that case, I’m just like, yeah, I feel like we just, what is the Tyler Cowen thing? Don’t underestimate the elasticity of supply is the main thing I would point to. I think on the capabilities front, I’m more uncertain. In particular, I’m sort of still confused and haven’t really looked into the, how much do you get directly actually from LLMs and foundation models for robotic capabilities? In particular, the other uncertainty I have is, it’s not clear to me that extremely fine-grained, extremely dexterous capabilities are the main thing you need for massive industrial explosions. And so, this longer tail of the hardest part of robotics, I’m not sure if that’s the biggest blocker for massive industrial explosion. Overall, robotics is something I have less expertise in. I’m interested in how many of the scenarios for doom ultimately kind of route through hard power acquired through robotics. I think part of my uncertainty also comes from, is it plausible to me that the minimum abilities that you need to acquire a lot of hard power and pose pretty catastrophic possibly extinction risks is more like, have access to nuclear codes or something like that? I don’t feel like I have great thoughts on this. I’m interested in more threat modeling, but I think that’s part of the thing is, what are the capabilities trends is something that people have disagreements about, and so what’s the minimum capability that’s necessary to cause these extinction-level harms, or harms that are sufficiently catastrophic, they just permanently alter the human trajectory?

Yeah. The last point I would make on—

I think that’s the kind of questions I want to see a bit more thinking on, but yeah.

The last point I would make on robotics is that I think the robots will be very good in constrained and repetitive environments, like manufacturing robots, and I think much longer until there are robots walking around the street cohabitating with humans. And if I were to go deep on this, I would want studies and discussions with people that are building multiple different data centers to how much variety is in their job versus how much of it is you take box off of truck and you put box in location. And I don’t have any good signal on how clearly repetitive that is now versus more human dexterous and ingenuity in problem-solving.

Yeah. Although also currently, the environments have been designed around humans who are pretty flexible in those ways and have other constraints. You could imagine designing factories to be robot first is a thing that I think you’re mentioning Amazon, right? Sort of does that more now. And so you could also... Yeah, I think that’s the kind of stuff that jaggedness would get you, which is you might just have massive accelerations, including in the physical world of some industries, even if capabilities for fully being as dexterous or flexible as a human might not be there. Mostly, yeah. I think mostly, yeah, robot capabilities seem very important to track to me. We had some piece about this at Epoch, but we don’t claim robotics expertise.

Yeah. I agree. I think we could shift to another capabilities topic, which is how far behind do you think the top Chinese labs are of OpenAI and Anthropic? And you can define how you want to measure it, whether it’s like public models or like internal models. I think doing both is actually pretty interesting to think about. Like, how would you describe the gap? Like, choose your... This is one of the few things I’ll push you for a number on.

Yeah. I think I would go with like, I don’t know, six to eight months or something, roughly.

From public to public?

Yeah. Like, release date to release dates. I don’t have a great catch numbers on like how long the internal to public deployment. Yeah, I’m sorry if it increases a bit if you’re counting when the model was built, because I would expect the delay to be larger for OpenAI and Anthropic than it is for Chinese labs. But yeah, I don’t know, just looking at ECI and there’s a few different methods you can do there, and there’s a good amount of noise between. Those methods are kind of reasonable, and my sense is they give you something on the order of like six to eight months-ish.

I would say that like Kimi K3 and GLM 5.2 are closer to like two to three or four. So do you think those are anomalies down to measurement or overfitting? So I have this discussion a lot with Florian. And Florian, who helps me with Interconnects, is constantly badgering me down. And I think I intuitively have landed something closer to you. And I think those two models, and in particular Artificial Analysis, were very close. I think Kimi might have been even under two in the Artificial Analysis Index at the time. And I’m just putting this out there, and I go back and forth all the time on it.

So I don’t remember the especially most up-to-date details on the Artificial Analysis Index. My sense is that it might understate the gap for curation reasons or something, but I don’t want to be too confident there because also they update methodology and stuff like that. I think even in ECI, though, there are some cases where it was more like four months. Yeah, I don’t know. Four to eight months seems reasonable to me. I think like two months, I would be like, “Nah, that seems like a bit more overfitting-y.” We did internally look into a bit how much overfitting explains the gap. Like, just doing the kind of like, what’s the ECI lag analysis for, if you just take an ECI based on private benchmarks versus not, or private benchmarks plus results were from models that were from benchmarks that were published after a model was released or something like that, where there’s no risk of contamination. I think overall, that effect was basically not statistically significant, which was interesting. I think if you did rely just results in model cards as opposed to the results being put in ECI, we probably get some contamination or overfitting effects. I still think there’s some amount of, yeah, something like ECI might underestimate the gap just because I do think OpenAI and Anthropic are probably like... There’s probably other capabilities just would appear less in benchmarks and where there’s been more optimization or serving a broader set of users. But yeah, I don’t know. That’s my overall number.

If we were to ban distillation effectively, not even just ban, but if distillation were effectively be stopped, where do you think the number would be in six to 12 months? I think without RSI being super crazy. I think RSI going super crazy, the labs pull ahead by a lot more.

Yeah. On current trends, I think I’ve actually updated towards distillation is actually a really big factor. We can talk about the different factors. And so, okay, so currently if I’m saying six months gap, then in six months, it’s probably not literally 12 months, right? So, my guess is, yeah, you are at like eight or nine, something like that. You go from like six to nine maybe.

Why have you updated on distillation being effective?

Yeah. So I think it’s like, yeah, there’s two reasons. I think, a lame reason is, I don’t know, more people seem to be saying it’s this.

They have been. I’m surprised by it.

Yeah. The Stolen Thoughts paper did suggest it was actually relatively easy to get the stuff. Yeah, random gossip. I think also a big one is, we can talk about this, thinking about the other explanations for the lag and them not seeming as compelling as when I first thought of them. I do think the Claude routers thing also just seems like a pretty big deal. And initially when I thought of distillation, I hadn’t considered that. But I do think Claude routers giving the kind of right prompt distribution for realistic usages, use cases and stuff like that just seems quite useful.

But it’s also like these companies have their own usage at this point. And it’s also like we have the benchmarks. So the benchmark distribution we already have. So I don’t necessarily think distillation is helping with... I think the routers could be very helpful, but I was confused by that prompt distribution thing because once you have the benchmark distribution, you just make similar examples or find similar examples.

Yeah, so I agree. So I think the Claude routers just give you a bunch of useful data to train on, and that will generally improve capabilities. I think that’s one explanation. I agree that I think giving you the right prompt distribution is, I think, going to be useful for real-world capabilities. I agree that purely if I’m discussing this ECI lag or something, then yeah, you can just use the prompt distribution for benchmarks. And so I think it’s not going to be that good of an explanation there. Yeah. I think just... Yeah. Also, so as an example, right, people saying that mid-training is a really important thing or something. I think actually in that conversation, right, Beren was like, “Mid-training is actually getting you 80% of the way there.” And more claims of this form and generally of RL being very useful but not doing that much of the exploration work, I think is also evidence that doing SFT on language trajectories is really useful.

Yeah. I want to go through—

I think he mentioned on a post on this is, it’s a good initialization, but I guess I’m updated that this initialization is really, really important. And yeah.

Yeah, because I’m still on the side that I think doing RL well and doing RL faster so you can go bigger is the way people are getting capabilities. But I have been hearing more about this repeated... The framework would be it’s a very repeated cycle between mid-training SFT and RL, and then your peak RL helps you feed into mid-training and less of a very big RL run, which I never thought... The RL runs are a week probably. I don’t think they’re insane, especially like Zhipu or Kimi. And that’s... So I could see this, and I was like, “I don’t really know.” I generally thought that because the models are getting so useful, I don’t know how... It just seems like the scale of distillation would need to be so big to really do that for mid-training. And the counterargument is Kimi and GLM 5.3 are also strong models, where it’s like, if the gap is that small, it just doesn’t make sense to me that the help would be so big. That’s why I’m still on the not convinced, but looking for more evidence. And then some of the other re... I have this blog post that I wrote in front of me. If you want to talk about distillation more, you can get that comment in before I switch to other topics.

I guess I’m pretty interested in talking about the alternative explanations or something for... I think there’s this general phenomenon of Chinese labs have way less capital and particularly way less compute than US labs. And that delay is huge. And then this, whether it’s four months or eight months, it’s a much shorter delay in capabilities than in capital. And so, this requires some explanation.

Yeah. Do we think the Chinese labs are still releasing meaningfully faster from time that RL is done to public gets the API and evaluation scores? In the past, I thought that the time to release was much faster, and I think the labs are releasing intermediate versions faster now, so I don’t know if that explains as much of it. But I would say a year ago, I would put a lot more to Chinese models get model out within days, American labs could take months, and that would artificially squeeze the gap very substantially.

Why would it squeeze the gap that much?

Because you have the trajectory of capabilities, and higher one is OpenAI and Anthropic, and they stop training here in time, and then the Chinese labs are here. But then it’s just like the model is stagnant before it gets released.

Yeah, I think this depends on the method that... This both depends on the method that you use for computing the lag or something, right? So my sense, I think when I was saying four to eight months, I think a lot of that uncertainty interval also comes from, are you measuring the method by looking forward or backward or something? Or how are you resolving the uncertainty? Are you taking this kind of staircase or this kind of staircase, or interpolating or something like that? So I think that’s kind of still accounted for in the uncertainty. So my guess is that’s not a huge... Or that’s still, even with this, I think there’s still a large effect to explain. And I think, yeah, these are other potential explanations, right? So one is maybe there’s a big delay in compute, but in labor or potentially data, there’s less of a delay, and those are important factors. And so even if you’re two years behind in compute, if you’re six months behind in data and in fact ahead in number of researchers or something, or not that far behind in number of researchers, maybe that helps catch up, and that’s not even a spillover effect, right? That’s just a reason. Yeah, there’s distillation. Other explanations people have are just ideas might leak, and I don’t suspect that’s a huge effect. People can play with a model, and through playing with the model, they kind of infer maybe what it was trained on or what’s useful to push on is another explanation. Those things are not quite like distillation, but using Claude as your reward model or using Claude to clean data or something, maybe have some spillovers.

Let’s see. I’m trying to quantify some. I would say that I think the Chinese labs care about benchmarks a bit more for financial... Some of them are public companies, and showing close benchmarks helps them a lot. So I do think they care about it more. I don’t know if that’s going to give you a month or two. I don’t know what are you going to give back, like a month or two. I think there’s a good chance that the Chinese labs are better at organizing talent on just doing really mundane hill-climbing data work. I don’t think that’s a huge effect, but it’s just like they have a ton of talent, and culturally through the way that... I think that there’s a good chance that it’s just more grind. They might actually grind more effectively than the American labs, which is... I don’t know how. I wouldn’t put a lot to this because I think it’s fairly close, but I think that’s a chance. But that’s a very marginal... That’s not a gigantic lead cause. I would put more of it to compute difference than to distillation. But maybe distillation is very related as a compute difference because it’s a way to turn... It is a definition of way to turn money into very high-quality data, which is something you would otherwise need compute for.

Yeah. Another theory that I’ve vetted about, and I’m not sure how big of a deal it is, is there’s this data market in the US, and maybe one thing that happens is the best RL environments or evals, it kind of takes a lot of compute actually to figure out which ones they are. And so maybe there’s kind of a collaboration between data providers and AI labs to actually test and figure out what RL environment really worked. And then maybe that data provider, once they’ve done this iteration, which implied a lot of R&D compute spending from a frontier lab, then they’ll build a bunch of similar RL environments because they’ve learned the lesson, and then they’ll sell them with some delay, but still sell them to other people, including in China.

I’ve heard this from people.

And so there’s kind of this implicit R&D compute that was spent and that it gets saved. I’m not sure how big of a deal that is.

I’ve heard this multiple times, including from people in China that have... This is like multi-hop type of rumor that I’ve heard a few times is like, we just have to wait a certain amount of time and then we buy the RL environments that Anthropic bought for a 10th of the price. And I think this could contribute a lot to capital efficiency, but also as somebody who buys into human factors being very important in model progress, which is just like competition, I think the analog to the competition that OpenAI and Anthropic feel right now is proof of concept as a way to make something way easier. And I do think the OpenAIs and Anthropics of the world are much more likely to innovate in open-ended things like Navier-Stokes, multi-agent mega-scaling than the Chinese labs. And having talked to many of the labs, I don’t think they would contest this, but they probably won’t put it on the record. But they’re just like, trying to keep up is so much easier.

Yeah. So that’s more like the... Yeah, so the way I describe this was like four-minute mile style effects of like, you show that something is possible at all and then people have the conviction to go down that route, don’t waste a ton of resources exploring a bunch of other things, is the thing you’re pointing at here.

Yeah. And I think this—

I guess I’m looking for examples where you think this... What are examples? Because for example, reasoning models is not this, right? Like, DeepSeek Math or whatever came way before o1. Yeah, I’m kind of curious if... Yeah, I agree this seems kind of plausible, and I’m not sure I can come up with examples of big innovations that a frontier lab has—

I think it’s more in org structure. So it’s like there’s R&D and modeling compute at the labs, which I think kind of get lumped together. And realistically, make-model-better compute of the next generation model versus the two or three generations down the line is very different. And I would think that the Chinese labs are just spending on a shorter time horizon. And then if there’s a major innovation in architecture or something that gets bubbled around the US ecosystem, I would assume the Chinese labs will figure out as well. I don’t think that an architecture innovation is kept very well between Anthropic, OpenAI, Meta, and Google. I think that these are all leaked very quickly through personnel turnover.

I see. So you think that actually has a big effect. So basically, I think it seems like you are just like, you think that with the other explanations, you can just get pretty far, such that you don’t really need distillation to explain a large fraction of the gap.

Yeah. So, I think without distillation, the gap would be fairly similar. It’s like also just building these models, there’s so much low-hanging fruit, and I don’t think distillation plugs that much of the low-hanging fruit that you are doing. It’s very unclear. It’s like once you establish your reasoning trace as being... This isn’t proven, but I think in mid-training and SFT, you need to have a pretty distinctive style in your reasoning trace to kind of stack it and have a defined reasoning strategy for the model. And Kimi and GLM have this, and I think they could modify it from where they are. And it’s just like, how does that compound over time if Anthropic and OpenAI change way faster, maybe it matters more. I don’t know. Doesn’t sound like we’re getting somewhere. I did like the post you had on understanding Chinese job postings, which I thought was a very clever way to try to peek behind the curtain of the Chinese AI industry. So I don’t know if you have anything you thought was particularly fun about this. It’s like, it was so hard to understand the industry that is going on there from outside.

Yeah. So this is work by Cheryl Wu and I. I think it’s, as a source of evidence, it’s more like you can get interesting anecdotes or interesting facts. I think it’s not so much the kind of Epoch-y thing where you can get a trend line that you really follow really well. I think some job posting trends are good for this, but I think especially for China, it is a bit more like, “Oh, well, you can see they’re hiring in this region for data center roles, and it seems like that suggests that at least some of these labs are basically constructing their own data centers as opposed to renting compute from Alibaba or whatever.” So you get a few kind of nice facts like this. I think you also do understand a bit more what kinds of product strategies the different companies have and what they’re aiming for. So I think it’s pretty interesting for that. This post was now published in June, so things move fast and things may have changed. But yeah, I think there are interesting takeaways on company strategies, on how they—

Can you say more about company strategies? Because I think that a lot has been cast on, say, like... The different company strategies in China. It’s like, what is DeepSeek’s strategy? I think MiniMax is one of the clear ones, and then Zhipu. And Zhipu has some on-premises deployment stuff, but at the high level, I think Zhipu can be much closer to OpenAI and Anthropic. But do you agree with this, or do you think any of them are more defined or have more interesting different things?

Yeah. I mean, one thing is just, I think we have this plot of how many job postings there are for B2B sales. As an example, it really does seem like Z.ai, like Zhipu, has a lot more of them than MiniMax or Moonshot. I think also, they have different strategies in terms of how much they want to expand internationally. I think MiniMax is a decent amount of that, more so than Z.ai, I think was the thing we found. I think there’s a few things like this. I think for DeepSeek, we didn’t see as much of that, for example.

Having done this analysis, do you think you could hand an AI model your taxonomy and have it redo it every three months? Do you think this is now a thing that Epoch could almost automate?

So I think we do have some upcoming project on job postings. I think specifically for this kind of thing, I feel like the takeaways are kind of ad hoc or something. And so I think it relies a bit more on you have a bunch of this data, what actually seems interesting? And I feel pretty often that LLM’s assessment of what are the top five most interesting things is not that great. But one example of a thing where I think... So I actually have this really crappy vibe-coded job listings tracker, which looks a bit more like the quality’s better for an OpenAI Anthropic. And I think Epoch might have something on this in a bit. But one example where AI is pretty useful is just, I don’t know, you just want to reasonably classify jobs as are they basically research jobs or are they basically engineering jobs? Or maybe research engineering is actually hard, but are they research jobs or are they GTM jobs? And the structure of the teams are going to be not coherent across different companies, but you can use AI to automatically track this over time and have some coherent categories. So then you do get trends like, “Oh, it seems like there’s been a surge in hiring from this frontier AI lab, and it’s basically mostly been sales, or it’s been mostly hardware infrastructure.” I think that’s pretty cool, interesting, and automatable.

Yeah. I think if you do this again, I’m interested in if you find things on data market expansion, which would be like your job is buy data. Because I’ve heard from many people that the Chinese data market is exploding, and I have poked various, whether it’s people I know at labs or people who I know follow Chinese industry closely. And it’s like, they’re like, “Yeah, but I can’t get anyone on the phone to know what this manifests as and what their budgets might be,” because Anthropic famously has billion-dollar budgets. And it’s just like I think it very quickly can become very important, like compute export control discussions stuff. By the end of the year, I suspect there will be some article talking about this and going deep reporting it. But from my perspective, getting that reporting information is near impossible.

Yeah. I think data is just really rough. I think for Epoch, data has been a tricky thing. It’s obviously an enormously important factor, but it’s harder to quantify or commoditize, and it’s also quite secretive. And so it’s like, I think a lot of the information is gossip. It’s hard to know how much to trust. My sense on Chinese, I think that article actually has something on this. Just it does seem like there’s a lot more in-housing in general by Chinese companies, and I think this also applies to data. And kind of relatedly, actually, there’s a lot of hiring for intern roles. I think if you go to the jobs listings pages for a lot of these companies, there’s specifically two tabs. They have summer internship tab or the internship tab and the normal hiring tab. And there’s a lot of, I think, hiring fairly junior people, and some of that might be for data purposes. But you did also see a surge in in-housing for data in the frontier AI companies. I’m not sure if this is still the case, but xAI was the only company that had this, but it was kind of nice because if you went to the careers page, there was actually the human data category, and you could just see which data roles they were hiring for. So there’s stuff to do on job postings on the data front as well, both xAI. I did this with Mercor and a few others previously. But I think it’s similarly, I think the most interesting things have been like, “Oh, well, that’s a crazy-ass role.” xAI is hiring for meme specialists or whatever, and it’s more ad hoc fun things than great trends, at least from what I’ve seen.

Yeah. I think that this is not going to change, unfortunately. It seems very much like individual teams at labs have contacts at data companies, and all of this is done privately, and it’s not... I guess even tracking compute is fairly hard, but now data centers are a manifestation that is somewhat trackable. And NVIDIA is a public company, and that is what is done. If the data companies were public, it would probably be easier to tease some things apart.

Agree.

Okay, we have another timely topic. How do you feel about frontier model safeguards? Given that in the last 24 hours, we learned that three people used public Claude models to hack OpenAI and get internal access, which is just crazy to me. Did you see this? It’s genuinely surprising to me that this is the leaky world we are in. At the same time, we are discussing RSI things much more seriously.

Yeah. I briefly saw this incident. I guess I’m not sure how much of an update it should be specifically on like model safeguards versus like you know security quality of the companies, or probably both. I think it’s a big question, right? Like, how good are the model safeguards? Seems like quite important, especially when people are talking about open models where it’s clearly open model safeguards are worse because there’s way more affordances that someone can have, but also open models are less powerful. And so just in terms of the you know misuse implications, it seems there’s just like slightly safeguarded, very powerful model versus like not at all safeguarded you know less powerful model. And I feel it was a priori kind of unclear to me, like which one is going to have the largest effect. Yeah, I don’t know. I think this is an interesting question. I think I feel also like people feel differently about this or something. It seems like some people feel pretty strongly that it’s hard to at least persistently do really bad things using the frontier models. But I’m like, yeah, there’s definitely clearly examples like this one, or really the Stolen Thoughts paper or something. It just seems like the safeguards weren’t good enough to prevent that from happening, even though there’s strong incentives for the companies to prevent it.

Yeah, the Stolen Thoughts—

I’m kind of confused about this.

Stolen Thoughts I put into like a serving issue, which is just like you were able to kind of jailbreak the API functionality to get the reasoning traces out in a reliable manner. Which I think the Chinese labs almost certainly had been doing for some time. And that potentially has a way to show more about what the effect of distillation was. I would guess that there are other ways that they get those tokens out, unfortunately, if there was one such prominent one. And then there’s the models themselves, which is the safeguards they have on the models, which is like the classifiers and prompt distributions and things. And I would expect the models to be somewhat permanently leaky. I just think that it’s hard to imagine. If you have a less aligned model and the task is, “Help me figure out how to get around said aligned model’s safeguards,” I would guess that that’s in scope of a strong cyber model’s abilities to figure out some sort of jailbreak for another model. It seems hard to permanently squash these things with how leaky they are now, despite the supposed launch, like higher false positive rate and being so general, and people get upset when they launch and they’re getting downgraded all the time, and people’s accounts get banned for doing bio research and things. So it’s just like the whole thing seems like a very big mess. I would be on the opinion of like, I guess maybe this is somewhat why the whole Fable freakout went down, which is like Anthropic was like, “Mythos is a powerful weapon.” And then somebody told the government and was like, “It’s actually easy to get around it,” and then they freaked out. But I think of it as like some of the distillation story has been around distillation lets them scale, take dangerous capabilities. But I think if you’re hosting an API at all, you have to assume in the current and near future that you can extract the capabilities out of that model, even if it’s slightly safeguarded.

Yeah. I guess one thing is I’m not sure how they prioritize between different threat models, but they sure seem to have some incentive to prevent distillation. It’s also hard to know what the counterfactual is, right? We have these examples of people doing those bad things. There’s a lot of usage of those models. It’s also compatible with, in most cases, it is actually really hard, and those people figured it out, or there was some random blip that allowed them to do it. So I feel I don’t really know how to update on that evidence. But yeah, at least in some cases, people just do manage to get around it. Maybe one distinction is also like, how long do you have to do this for? So there’s the model’s alignment that’s in the weights of the model, and then there’s your synchronous monitoring that you can do in the moment. And surely that’s going to be way less effective than how good you’re going to be at detecting things where asynchronously able to spend a ton of compute to parse through trajectories and analyze group trajectories together and stuff like that. But that takes a longer time, and so if someone needs to do an attack very quickly, then it’s just too hard for someone to detect it in a moment. But then if someone wants to pursue a very long campaign, maybe it’s harder. I’m not sure how long those campaigns take or how effective those KYC things are. But that also seems relevant for other cases that are not about leaking capabilities, but about other cases of misuse.

I guess to make it more specific, how much of, say, risk proliferating from open models do you think is down to the fact that people will serve the open models or can serve themselves without the additional oversight and classifiers, versus what I think was talked about more, which is just like fine-tune the safety away, which I think very few people are going to fine-tune Kimi K3. But if these classifiers that squash 99% of people at inference time just kind of blanket a lot of the risk, I think that’s almost more plausible. It’s a bit of a sequence to get there. You have to assume people could get around model safeguards, assume a lot of usage.

Yeah. I guess the fine-tuning thing right now is not a big deal because not that many people have the combination of motivation to do it and competence and willingness to do it. The bar for that also falls as AI research becomes more accessible. But yeah, I think right now, the fact that there’s basically no safeguards, or my sense is the inference providers aren’t trying as hard as Anthropic or something at this. I could be wrong about that, but that’s my impression. And also you can pick whichever one is least secured. So I think that’s a bigger deal, at least currently, yeah, I would expect. But yeah, I think it also depends. I think the main thing that I care about for this is actual threat modeling. I just feel kind of confused about which kind of group I should have in mind as the most likely person to do a bad thing with AI. And then I feel like if someone gives you an actual scenario, then I feel like you and I will have decent takes about will it be easier for them to use an open model or to try to jailbreak a closed model? And I just don’t currently feel like I have a good sense of this. But I think—

When this is so ambiguous, I think it makes me feel a bit better about the near term, which is just like these safety tools are, it’s to some capacity inadequate now, and there are strong open models, and it’s not like all hell is breaking loose. And you have to keep reevaluating this as things change. But I think it doesn’t seem like things are going to change. I don’t think there’s that much of step function changes. I personally need to learn more about biorisk that people very often talk about, and I think need to... My question is, what is the actual manufacturing pathway and where is the model intelligence at now? Where I think cybersecurity, we’re in it. We’re in what it’s going to be like for a bit, and it’s not good, but it’s not like nothing horrible has happened yet, and it’s okay. I don’t think there’s going to be more.

I don’t know what the sense with the evidence is, right? Companies don’t love to admit. I think it’s famously kind of hard to estimate cybersecurity costs and stuff. I know we want to look at insurance costs. I don’t think they’ve changed that much, but it would be interesting to look at right now. Like cybersecurity inference prices or something is like what percentages you can look at, but companies don’t love to disclose they’ve been hacked, and so they sometimes just pay. But yeah, definitely, yeah.

That’s a good point as well.

I don’t feel like I have a ton of expertise there, but I’m pretty interested in those questions.

I’m glad that this was a kind of like, “We don’t know, it could be worse,” and not a, “This is the list of very bad things.” Which is, it’s hard to motivate people when it’s like, “I think we need to keep preparing, and you need to assume that the Hugging Face thing might happen to you.” But it is not a super, it’s not an avalanche of compounding problems right now.

Yeah. I do expect Hugging Face-style things to happen kind of in the wild regularly and outside of OpenAI. And if it’s just that, then that’s not massively bad. I mean, it’s still a kind of screwed-up world. But I do expect more of this just happening with a variety of models and neoclouds and stuff over the next year. Just a bunch of open source agents that people can’t control.

Yeah. I do think your threat model thing is a good understanding, and it’s like, who are the people that would try to take down an energy grid in a domestic area? And I think the list is pretty small. And it’s like, what are they doing? And I’m not an expert in this at all.

Yeah. I feel like the biggest safeguard is just there’s not that many people who want to do horrible things. Fingers crossed.

Yeah. Okay, I have a fun question. I think it’s like Epoch AI creates so much of this kind of area setting information for different parts of the AI discourse. One thing is, is there a fun process on how you choose problems that you think are important? Are you guys constantly underwater and wanting to do different things, or is sometimes it’s like, “Now’s our moment?” Just behind the curtain of Epoch and what you guys do culturally is fun to me.

I think what determines project selection, I do think it is very what seem like the important questions to us are quite curiosity-driven overall. I think that’s probably the main thing is just what seems like a huge deal, I think has been the main determinant of where efforts go. More specifically, yeah, another framing is the Epoch was created as this thing where it seemed like neither the industry nor academia would be able to provide a good public resource is also well-maintained on the most important trends in AI. So it seemed like there was something missing. Another framing is, I know, what do people have misconceptions about, is sometimes, “Oh, people are wrong on the internet,” kind of things. Some articles can be explained by this. But yeah, I think it’s very curiosity-driven. And also it’s just like, oh, what’s evidence that’s leaking information that’s laying out there that we can use? Like, the job postings is this. Data center tracking is very much this as well. It’s like, man, these are big objects. Surely you can see them and track them.

Do you think the information needed to understand the AI industry is stable in your access, like getting less accessible, getting more accessible over time?

I mean, it really depends, right? I think data center is, so far at least, it’s pretty good, and they’re getting bigger, which makes it easier. I could imagine at some point it becomes a really strategic asset, and there’s way more adversarial pressure to hide them or at least hide key information. For now, it’s pretty good, right? In contrast, Epoch is an effort to track model training compute, and at the time, you could just look at the paper. It’s much harder to do this with Astra. So yeah, I think it depends on the verticals. One thing that I mentioned before is, yeah, data is pretty rough. We’ve had some stuff on this, but it’s much tougher than compute, and it’s pretty secretive. So yeah, I think it depends on the vertical. Oh, yeah. One thing is, as the economy grows and companies go public, that’s better information, right? So that’s true for compute companies. As IPOs happen, that also helps get more data.

Yeah. I find it very interesting. It’s like I’ve long wanted to know what a post-training... Like, what is a flowchart of a post-training recipe at a frontier lab? And I think it would be wild, and I think those flowcharts exist. And even some old ones, I think, would give people a lot of thought, because I do not think it looks remotely like the OLMo three-stage approximation type thing. And we have some from like Nemotron and some of the Chinese labs release stuff on it. But when you hear John Schulman talk about post-training and other people, it seems like the total Wild West and just stacking everything together. That framing for like that framing for what is the complexity of putting a model together, I think would be very useful for people as a mental model for what is being scaled and what are you scaling through. Like, yes, you are scaling, but if you’re scaling on top of this kind of house of cards thing, it’s not perfect.

Yeah. I’m actually curious. Yeah. I have this like... Okay, can I test my current mental model of this, which is pretty uncertain? My sense is this, okay, there’s a big post-training team, and it just has sub-teams that are kind of specialized. Like, some people are going to be doing math, and some people are going to be doing biology or finance or something. And the job of those teams is to ask for data and curate that data and make the RL environments good, and also figure out what kinds of hyperparameters work well for their own domain. And they do this by buying very high-quality evals that they use as a north star to iterate on and buying those RL environments. And their main output is basically like, “Here’s a curated set of environments, and here’s the hyperparameters that we want.” And then my sense is indeed now things are basically all those teams bid to put some fraction of their data into a final big RL training run. And they also bid to maybe for the hyperparameters of that final RL run to be close to what they want, question mark. That’s at least for the RL process. And then there’s in fact a big RL run. Another alternative would be, of course, they all do their experts, they distill them back into... Does that roughly what things are like? This looks—

I think the org side, I agree a lot. I think the org side, there’s definitely per domain expertise, and it’s a lot of why I didn’t want to go get a frontier lab job, because most likely I would just be given, like, “You’re going to work on this data and you turn the crank.” I don’t know if the complexity is a lot higher on how you put it together. I think it’s interesting that the open weight models are split between multi-teacher on-policy distillation, and some are just like do sequential RL. I think the diversity there would be very, very interesting to the industry of whether you have multiple RL runs, how do you decide which environment to do and which of the stages? How do you manage mid-training and SFT? I think mid-training and SFT, I think, are one stage in the frontier lab, and the base model is just before mid-training, and then you do a whole bunch of stuff at the end there. And I just think there’s a lot of things you can do, and I feel like if OpenAI and Anthropic’s recipes were closer to mid-training and then big RL run, if that’s actually getting them to where they are, I think I would expect progress to go faster than if it’s mid-training and fork and model merge and bunch of experts and MOPD and then more RL. I think that just is a bit of a bottleneck on progress, and it would be very useful to have more information on what does model training look like.

My vague sense, but I don’t have super hard evidence on this, is that it has gone more in the direction of, yeah, away from MOPD with multiple experts and a ton of checkpoints, and a bit more in the direction of like, okay, there’s in fact one big run towards the end, but of course, much federation before, but I don’t know that.

Yeah, maybe the rumors we’re hearing about mid-training being important and distillation is from that style of loop rather than the expert style loop. I don’t think we’ll know. I think we could continue that another time. This was fun. I think you should keep giving your incisive responses when I write something. I think this will be very useful to people and good to finally chat a bit longer.

Likewise, yeah. Thanks a lot.

See you soon. Bye.