Podcasts
Paul, Weiss Waking Up With AI
Full Throttle at the Frontier: METR's 11 Weeks at Anthropic
In this episode, Katherine Forrest and Scott Caravello revisit recursive self-improvement in light of METR’s recent assessment of AI research and development at Anthropic, examining its key findings and caveats. They also explore open questions around compute, bottlenecks and the pace of future progress.
For the sources referenced in this episode, please see the links below:
Anthropic: Responsible Scaling Policy
Epoch AI: Epoch Capabilities Index
METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
Episode Speakers
Episode Transcript
Katherine Forrest: Welcome back to Paul, Weiss Waking Up With AI. I'm Katherine Forrest.
Scott Caravello: And I'm Scott Caravello. Katherine, how are the new digs?
Katherine Forrest: Well, you can see from the video that I am in our new space-age offices, and I just had a problem because my mic was plugged into some sort of space-age setup that has many, many wires in a very fancy wire squisher. I think that's the official word for it. It's called a wire squisher, and I could not get my mic to do its thing. But anyway, the office is absolutely gorgeous. It's absolutely gorgeous. The food at Paul, Weiss is still free, which I love. It's like, it's so good. It's so good. Like, you get to go down and have breakfast. You can come up and have lunch. It's like one of the few perks left in the world.
Scott Caravello: Is it true that there's a pizza oven? And if so…
Katherine Forrest: Yes!
Scott Caravello: …wood-fired or, like, brick oven?
Katherine Forrest: You know, OK, I went over and had a turkey and provolone cheese sandwich. So I did not, like, have, you know, a pizza, but I saw the pizza oven from afar. And so that was really exciting. And they do have, like, a barista station in various places. So, Scott, nirvana is waiting for you. It's just going to be nirvana.
Scott Caravello: Exactly. And for our listeners, I am moving into the office next week, so I am not just, you know, lazily hanging around at home.
Katherine Forrest: Oh, yes, you are.
Scott Caravello: Katherine is the advance party.
Katherine Forrest: Right. I am the advance party. We are moving in two phases, phase one and phase two. I have to get my George Washington big picture, my etching on the wall of George Washington saying goodbye to the troops. But after that, I'm totally good to go. So let's get on to our topic, though, because even though I'm distracted by all the newness that is around me, we have actually a really great topic for folks this week. And so let's start with a little bit of follow-up from something we talked about a couple of weeks ago, and then we'll jump ahead to sort of a new development, which is we were talking about this, you know, recursive self-improvement concept, which has become in some ways sort of AI, I wouldn't even say red line, because it's not a line that can't be crossed. It's more of a people's view of an ultimate warning sign of when AI might hit sort of a terminal velocity, when humans may or may not lose control over it. And recursive self-improvement is when AI improves itself and then improves itself and then improves itself upon itself without human direction, permission, or intervention. And so when you have recursive self-improvement, you can really have takeoff. And so the other week, we said that we were on the way towards recursive self-improvement because there has been, for instance, some of the highest-capability AI models right now are 80% coded by themselves. They are coding themselves, but there's still a human somewhat involved. But we're now seeing some of the results, we think, and that's the big story for today, some of the results of AI improving on AI. So this is really, folks, a moment to sort of stop and pause because, you know, we're going to be the bugs in the water with things sort of, like, bubbling up. It's not going to happen all at once. We're going to start to see things happening little by little. And this is one of those things.
Scott Caravello: 100%. And so, talking about looking at this evidence, it's actually seeing the results of this trend. We're going to talk about a report from METR, and that's M-E-T-R, which stands for Model Evaluation and Threat Research. And so that report came out on November 5th, and we're recording this on November 16th. And that report asks whether AI itself is actually speeding up AI development at Anthropic. And as you might have guessed, the short answer here is yes. Anthropic's models are now improving about twice as fast as they used to, and METR attributes most of that to the use of AI in the development process.
Katherine Forrest: Yeah. And let's talk about, like, you know, what "used to" is. We'll come back to this, I think, in a little bit, but "used to" is from '24 and '25. And so we're not talking about that long ago, and we're talking about a real speed-up in just that short period of time. Now, it's no surprise, but METR is an independent group that actually evaluates frontier AI models. And I first became aware of it with some of its red-teaming sort of capabilities and its sort of early work in that area, and it's an outside group. But it was granted actually some credentials, some limited credentials, for some of the METR folks to go inside the Anthropic models in using the METR lab. But it was really pretty unprecedented.
Scott Caravello: Right. And so this assessment touches on one of the questions that I had ended our last episode on recursive self-improvement with. And that's whether, you know, even if AI gets to this point where it can just rapidly and exponentially improve upon itself without human intervention, there are other bottlenecks that still might slow the process down, like the amount of compute that's actually available for the AI to do its thing, so on and so forth, right? If we don't actually have enough of that computing power, could that still have something of a slowing effect on AI development, even if all of the research and improvement is automated? So we're going to come back to that as well. But I want to front that because I think it's important. And this report at least attempts to give us some clarity on that question.
Katherine Forrest: Right. And actually, I want to add to something that I just said that is really important to how METR conducted this. So METR actually conducted the study independently, and it used its own lab, but it also had extensive access to Anthropic's internal research, its data, its personnel, obviously its models. And so it was sort of using all of that in combination. And the METR study was released, we should say if we haven't already, on November 5th. And they actually put some of the workers in place within Anthropic for 11 weeks. So the study went for these 11 weeks, from August 12th to October 29th. And they got a lot. They got benchmark scores, including for the models that Anthropic hasn't even released yet, or for those models they got, you know, computing data, staff surveys. They took 15 interviews, and they also themselves had access to the models.
Scott Caravello: Right. And so, I don't know, I mean, to me that kind of immediately raises the question, why would Anthropic open its doors like that? And at least part of it is likely due to what's in Anthropic's Responsible Scaling Policy, or RSP, which is the company's safety framework. And that lists AI research and development acceleration as a key risk indicator for AI. And so then the worry, as METR puts it, is, quote, "a sudden emergence of extreme and risky capabilities," end quote, if AI significantly speeds up development. And so, two other things worth knowing about how the assessment was conducted as we finally get towards talking about findings. One, METR didn't take payment. And second, this wasn't full employee-level access, right? So, you know, even though, as you mentioned, Katherine, they had access to a ton of internal information, it's not like METR could just go into the systems and browse documents on its own. It would ask Anthropic for data, and Anthropic could say yes or no to those requests.
Katherine Forrest: Right. And Anthropic had the right to redact anything that they viewed, that Anthropic viewed, as nonpublic. And it did. And actually, METR starts the report by talking about the redactions. And that's a very interesting and METR-like thing to do, right? METR, one of the things that METR does is it often lays out very carefully the methodology that it's undertaken, but also the limitations of what it's been able to do. And it's always done that. But here, with the redactions, METR does spend a fair amount of time on them because they say that Anthropic redacted the majority of evidence that METR found most useful to explain its findings. And that Anthropic told METR that removing confidential information was very important because otherwise it could help AI users, developers who might not have the best of intentions, move even faster. Or, I suppose, even if you had good intentions, they might move faster than some folks might deem safe. And so this sort of redaction piece is an important caveat because we're not able to tell you exactly how METR was able to discover all that they were able to discover. They have to sort of get at it in a bit of a roundabout way. And they say that the public has to rely upon METR's judgment. Now, I think METR's a very respected organization. That's not a lot of work in this area, so it comes to the table initially with a lot of credibility.
Scott Caravello: Right. And so then I'll just make one final note on this question about, you know, why Anthropic would provide this access to an outside group. And that, you know, for all we know, they may be using this assessment, including certain findings that were redacted from the public report, to inform its own view of the risk of AI R&D under its Responsible Scaling Policy. So with that, Katherine, let's really get into it now.
Katherine Forrest: Right, right. So what are the, you know, we've talked, talked about how METR comes out with this report. It talks about sort of AI creating AI model acceleration; it's creating additional AI model capabilities. And so the first finding really is what we've already mentioned in that sort of general sense, which is a double acceleration, a 2x acceleration. And they actually say it's between 1.8 and 3x, is sort of the range of acceleration. So it could be a little bit less, but it could be a little bit more. And to put this in sort of simple terms, METR is saying about Anthropic that Anthropic is now making about two years' worth of progress in one year. But it's really important to judge when that benchmark progress year is judged from, because it's not as if they're judging it from 2018. They're judging that benchmark year as 2024–2025. So it's about the speed of improvement.
Scott Caravello: And so to put that in really plain terms, right? This doubling, it's like a car going from 30 miles an hour to 60, and you're covering the same amount of ground in half the time. And so the score that METR's actually using in coming to that 2x acceleration figure is something called AECI. And that's actually an Anthropic measurement, and that's based off of the Epoch Capabilities Index, which is a score that a research group, Epoch AI, uses. And so what that does is it takes the result–
Katherine Forrest: You don't pronounce it ep-ic? It's e-pock? It's not epic? I've always called it epic.
Scott Caravello: I assumed it was “epoch”, like age, you know, like…
Katherine Forrest: Isn't that epic? Doesn't that mean ep-ic? Is it epoch?
Scott Caravello: I thought it was…I don't know. I mean…
Katherine Forrest: All right, all right, continue, continue.
Scott Caravello: Wow, just, you know…
Katherine Forrest: OK, Epoch or Epic Capabilities Index. OK, go ahead.
Scott Caravello: E-P-O-C-H for anyone who wants to email us and settle this debate.
Katherine Forrest: People can vote. People can vote. Go for it. OK, the Epoch Capabilities Index.
Scott Caravello: So the way that score works is that it takes the results from a lot of different benchmark tests, which we've talked about so many times, that are used to measure capabilities on different tasks, and then rolls them into a single number to talk about capabilities.
Katherine Forrest: Right. And METR's finding is that, based upon what it calls, quote, "well-specified tasks," end quote, Anthropic's models are improving at about twice the rate that they did from March '24 to October 2025. So that's what I mean by that benchmark year. Boy, twice as, improving at about twice the rate. Incredible.
Scott Caravello: So American English, it's "epic," and then British English is "epoch."
Katherine Forrest: OK, and so does the Scott Caravello from Italy, that gets you to the English? When do you get to take advantage of the English pronunciation? Did you go to school in England?
Scott Caravello: No, but I'm going to London next week, so maybe, for Thanksgiving. So maybe that's it. I don't know.
Katherine Forrest: Maybe you're sort of, you're just anticipating it. I just want to know whether or not you have any kind of English connection that would allow you to call it "epoch" as opposed to "epic."
Scott Caravello: No, and I have not been to the UK since I was five years old, so really, there is like…
Katherine Forrest: There we go. OK, now back to, so they're improving at about twice the rate as they did between, you know, March 2024 and October 2025. We've solved the ep-ic sort of dilemma of pronunciation, but go ahead and put some color on that.
Scott Caravello: Yeah, exactly. And so I think this is really, really important to understand, because the tasks that are being measured to come up with this capability score and then get at the pace of development are tasks that have clear answers and that are easy to grade. And so METR actually flags that it's less sure how this speed-up that it's measuring, that 2x, would carry over to broader, messier tasks. And that's one of the caveats in this report that really stuck out to me, because as these models advance and we're trying to understand how they're advancing, what else they can do, the growth that matters might be on complex problems that don't have a clear answer. And that may be what we actually want to use AI on in the real world.
Katherine Forrest: Right. METR gives a range on that improvement, and while it leads with a doubling of the pace of progress, as we've said, you know, it does talk about the range being somewhat flexible. They've picked sort of a midpoint because it can be somewhere in the ones, up to three times the old pace. But how do we even know that the cause of this improvement is about AI's role in accelerating the R&D or the actual capability improvements that we're seeing? Well, Anthropic has made headlines for its compute deals, right? We've seen a lot of that. It's gotten a lot of compute. It's built compute; it's leased compute. So it's not necessarily only about that, but METR in fact even says it can actually rule out compute as being the primary causal factor in this. You know, what they spend time doing is looking for, for another reason.
Scott Caravello: Right. And even though they say that they can confidently rule out the fact this faster growth is coming from increased compute, the details behind that were redacted, so we don't really know exactly how it arrived at that conclusion.
Katherine Forrest: It's actually interesting, because normally you would think of putting in more compute as at least having some causal relationship to it. But anyway, we're going to take METR's word for it. They've got a lot of credibility, as we've said, going in, but their rough guess is that most of the current speed-up comes from Anthropic using AI to do its own AI research, though it calls that a guess with substantial uncertainty, which is also interesting. So they've eliminated compute for most, if not all, of it. They then say that AI doing its own AI research, they think of it as that being causal, but they also say that that's a guess with substantial uncertainty. So it thinks that non-AI factors also play a large role, but that AI has become also relatively more important. So there's a lot of unknowns. But METR does name some other things, like data improvements. You know, we've seen a lot of data improvements, both the quality of the data and how the data is prepared for training. But many of the details on that were also redacted. And then there's also some disagreement within Anthropic about itself, what's contributing to the pace. So you remember how they did 15 interviews. Some of that may have come from that. We're not really quite sure, because a lot of that's redacted as well, but one researcher pointed to hiring as being a very large driver.
Scott Caravello: And so now, taking all that and connecting it back to recursive self-improvement that we, you know, kicked off the episode with, where AI is improving AI on its own. And like you mentioned, Katherine, this is not a statement that RSI, that strong, true version of RSI, is here, where we're talking about fully autonomous AI improvement. But these are results, right, where we're seeing the use of AI in the model development pipeline to really drive a speed-up in progress. And so what does that actually look like on the ground in the context of this assessment, before we're actually getting to this point where AI is actually automating the entire thing from front to end? Well, METR was told that it's mainly AI assisting researchers, so writing code, running experiments. But, and I think this is a big "but," if not the biggest one in this entire report, researchers described, quote, "signs of life for fully autonomous contributions," which could come soon. And I know, Katherine, right, you said that we, you know, have one foot on the Yellow Brick Road on the way to superintelligence, to RSI. That is, like, a really significant statement as we're talking about moving along that pathway, because there are significant things coming, or so it seems.
Katherine Forrest: Right. And so the signs of life that they're talking about are not signs of, like, a living thing, but what they're saying is that they're seeing the possibility that AI is making autonomous contributions. They're really seeing that. And so that is an incredibly important finding. And there's a quote that I want to read from an Anthropic researcher who actually isn't named in the report. And they say, quote, "Now only bottlenecked on ideas and compute, no longer bottlenecked on code." So, that's actually really interesting. OK, I'll read it again. "Now only bottlenecked on ideas and compute, no longer bottlenecked on code." You know, I'm just, actually, I'm pausing there for one second, though. So, Scott, because, you know, one thing that's really interesting about that quote is that METR had said that compute is really not the driver of this. And so they're saying they are only bottlenecked on ideas and compute. So that would suggest that compute actually is playing more of a role in it than, you know, than not. But anyway, we're not going to solve that today.
Scott Caravello: Well, that's what I was going to get at. And because, you know, I think one of the most interesting things about actually reading through this assessment from the first page to the end is seeing where those kind of potential contradictions are and the different views between METR and Anthropic. And obviously understanding that we can't necessarily get to the bottom of it and sort of form our own conclusions about who may be right in that debate, just because so much of the underlying evidence is redacted. And so that's just another important caveat to put out there one more time.
Katherine Forrest: Right, you know, and let's just sort of pause before we go into the next piece of this, which is, one thing that we do know is that the Anthropic models, like the other frontier models, have gone through these massive capability improvements, right? Just massive. So when we're talking with this METR report about how quickly those improvements are coming and what's causing those improvements, you know that it's incredibly important, but it's also demonstrably true that these models are improving at incredible velocity. So let's go back to this bottleneck question. So METR itself didn't find any convincing bottlenecks that they say could get in the way of further acceleration. So let's just sort of pause on that. So one thing that you would wonder about is whether or not, OK, you've got all of this acceleration, you've got things sort of now at, say, 2x, twice the speed of improvement as they were, you know, in 2024–2025. And then you've got METR saying, you know, not only that, but we don't even see any bottlenecks that could get in the way of further acceleration over the next 12 months. And so they're saying things could get even faster. And, you know, METR says that it looked at the algorithmic progress, the training compute, the compute for experiments, the data. And the worry, or its worry, is that if nothing is holding acceleration back, then the speed-up that they found could actually compound, and better models could speed up the creation of yet better models, which could create yet better models even faster. So the concern, which is really also the concern with recursive self-improvement, or RSI, generally, is that the speed-up of the pace of development becomes so fast that we humans are barely able to keep up with it. And that we're also not as involved with it as we used to be. And that ultimately, you know, there's a concern of loss of control. Now, that said, METR is also candid in its views that whether or not, not there are any potential bottlenecks changed as they were conducting the assessment; it sort of went back and forth. And so this was a point that was debated by the METR folks.
Scott Caravello: Yeah. And, you know, not only do they say that their understanding of whether there were bottlenecks changed several times as it received more information throughout the assessment, but ultimately, METR said that it's not even confident in where it landed on that bottleneck question.
Katherine Forrest: Right. And so let's just cover one more finding that gets also to the heart of the assessment, and it connects to something else that we've already covered. Now, a while ago, we talked about Mythos preview in Fable 5. And by the way, those are now, they've gone up a level, and the Mythos 5 release and all the circumstances around that, that was the first generation of Anthropic's now most powerful tools. And some of Anthropic's earlier system cards had floated the idea that the Mythos preview step change in capabilities was actually a one-time, time jump. So I encourage people to go back and to read those system cards. But METR says that Anthropic's newer internal models make that less likely. So it looks like there could be a lasting change in pace, and we might not be anywhere near the end of step changes.
Scott Caravello: Yeah, exactly. And so METR had access to an internal model that it calls Model 4, right, an Anthropic unreleased model. And it said, for example, that it scored over 50% on every benchmark it was tested on. And meanwhile, nine months before the assessment, Opus 4.6 was below 50% on many of them. So it seems like it's getting at a real, durable improvement in capabilities, which is just going to continue to drive this acceleration of progress.
Katherine Forrest: And this just reminds me, we've got to have a whole episode, Scott, on benchmarks, because, you know, certain benchmarks now, it's so hard for anybody to tell what these benchmarks mean, because on some benchmarks, these models are scoring at, like, 90% and 100%, and some benchmarks are still, you know, you talk about over 50%, and it seems like it's incredible. And, you know, it all depends on what the benchmark is. But just trust us that, you know, when you're comparing like to like, and what METR is saying is, hey, on Opus 4.6, an incredibly powerful model, the models were scoring below 50% on one particular METR-administered benchmark, and now they're well over 50%. That's an important distinction, because we're talking about sort of a jump, a capability jump, and again, you know, the potential for acceleration spurring further acceleration, and yada, yada, yada.
Scott Caravello: And to add on to that, METR actually thinks that the doubling in how fast these models are improving might not even be the whole story.
Katherine Forrest: Right. And that's the redacted part, or part of the redacted part, because there's actually a fair number of things that are redacted, as METR explained. So METR said that it saw evidence that it can't describe that made it take that doubling even more seriously, and that hinted that the pace of change is still increasing. All right, so get that. So it can't tell us what it saw or why it thinks this, but we've got this incredibly, you know, sophisticated organization that says that acceleration of model capabilities may still be increasing beyond the two times that it found. In the report, though, METR said actually that it would be surprised if the acceleration were significantly higher than the overall range, which, again, goes from the ones up to three times. But if the growth is hyperbolic, which it could be, then, you know, it could be substantially understating acceleration in the near future. So we just don't entirely know where this is going. But again, no reason to think that there's a bottleneck.
Scott Caravello: Yep. And exactly on that point, right. You know, I mentioned internal Model 4 that they were looking at and seeing how the benchmarks improved, but that was a model that was already trained. It was already ready for them to run the benchmarks. And so METR is flagging that their results might actually be, frankly, behind, because they only see the results after the model's trained. So whatever's happening in the AI-assisted research today is going to show up later, when the next generation of models are trained. So again, it could be higher. This is not, not, frankly, like, extremely certain, as hopefully all of these caveats have made clear, but it's really, really interesting and shows where the trend lines are going. But so, I know we're almost out of time, Katherine. So where should we leave folks?
Katherine Forrest: Well, you know, what it says is we've seen a lot of, of change with model capabilities in the last six months, last couple of months, even this fall. You know, we talked about them, we announced them, we talked about the new Google Gemini Carbon, and we've talked about models all the time, the new ones that are coming out. And so I think we're going to be seeing more and more and more of it. And we're also, at the same time, seeing, you know, the Millennium math problems be solved. We're seeing proofs of the Millennium math problems that have been solved actually come out from other organizations that are validating them. We're seeing now other models solving their own problems. We just had one that was announced over the weekend with 28,000 Google Gemini agents working on various scientific problems. So we're going to see continued acceleration of change. So I think, Scott, that's all we've got time for today. Next week, we'll talk about our continued acceleration of change.
Scott Caravello: Cannot wait.
Katherine Forrest: All right, Katherine Forrest signing off.
Scott Caravello: I'm Scott Caravello. Don't forget to like and subscribe.