00:00:00 · Parv Mahajan, Ruby Justice Thelot
[00:00:00] S0: If you use, their numbers, you find that it actually comes out to about 1.97 x depending on the exact measure you use, which is very close to two x, right, which is the threshold. And, if METR was running the report and especially if METR staff were examining all the evidence that Anthropic had access to, we would likely think that the threshold is very probable or even maybe likely that the threshold of two x had been crossed.: [00:00:23] S1: And we are live with Parv Mahajan, who is chief of staff to the president at METR, which is Model Evaluation and Threat Research, the nonprofit that just put out this report yesterday on assessing Anthropics AI R and D acceleration. Basically, are we closer to an intelligence explosion? Are we closer to full RSI than previously thought? A lot of people are now saying yes. So we'll talk about that. Parv, welcome to MTS. [00:00:51] S0: Yeah, thanks for having [00:00:52] S1: me. So for the audience, what exactly did this report find, and what did it not find?
00:00:52 · Ruby Justice Thelot, Parv Mahajan
[00:00:52] S1: me. So for the audience, what exactly did this report find, and what did it not find? [00:00:58] S0: Yeah. So first, for just some framing, this was a pilot high access, exercise in which we, tried to study what is the rate of AR and D acceleration inside Anthropic, and how much is AI contributing to that acceleration. And overall, what we found is that, we see this, like, sustained 2x acceleration, among historical trends. So since about late twenty twenty five up until now, we're seeing twice the rate of progress. So, like, two years of AI progress in, just one year. And we think that, currently, this is mostly due to AI. Although, again, the evidence is is is, very uncertain. And then additionally, we also find that there's no, like, observed bottlenecks, or, we also don't see very obvious causes of, like, strong diminishing returns that might imply, that we're seeing bottlenecks to acceleration or acceleration could slow down in the near future. So we were not able to rule out fast acceleration. Although, again, the evidence is uncertain, and we weren't ruling in very fast acceleration either. [00:02:01] S1: Right. So one seemingly very important, very crux y thing is like, this measures the rate of acceleration on well specified tasks. The kinds of tasks that end up in the Anthropic Epoch Capabilities Index, the ECI. But like, it seems like a very important thing is like, does this generalize to other tasks that are harder to specify? Like, is there any evidence on that one way or another?
00:02:01 · Ruby Justice Thelot, Parv Mahajan
[00:02:01] S1: Right. So one seemingly very important, very crux y thing is like, this measures the rate of acceleration on well specified tasks. The kinds of tasks that end up in the Anthropic Epoch Capabilities Index, the ECI. But like, it seems like a very important thing is like, does this generalize to other tasks that are harder to specify? Like, is there any evidence on that one way or another? [00:02:32] S0: Yeah. That's exactly right. So importantly, the main piece of evidence that we were able to publish without... Or or without many redactions was, ECI, which, like you said, concerns well specified tasks. Right now, I believe the public evidence is compatible with with, like, not really seeing that much acceleration in these fuzzy long horizon tasks, which you might think might be necessary for AR and D, acceleration or, like, massive RSI or something like that. However, as we say in the report, we do see signs of life of AI coming up with, like, kind of autonomous, you know, autonomous algorithmic improvements or, like, major improvements to parts of the pipeline that we would expect could be necessary for our side. Although right now, the evidence is very unclear. Right. [00:03:13] S1: Is there any evidence of, accelerating returns? Like how do we know that in order to achieve a self sustaining intelligence explosion, the returns on a given unit of AI research have to be positive. Like you have to get out more than you put in. So, like, what is the evidence on that question specifically?
00:03:13 · Ruby Justice Thelot, Parv Mahajan
[00:03:13] S1: Is there any evidence of, accelerating returns? Like how do we know that in order to achieve a self sustaining intelligence explosion, the returns on a given unit of AI research have to be positive. Like you have to get out more than you put in. So, like, what is the evidence on that question specifically? [00:03:31] S0: Yeah. Yeah. So the question is, like, if we, you know, if we advanced capabilities by 1%, does that make it easier to get the next 1%, or does it make it harder to get the next 1%? Like, are we only picking low hanging fruit or something up? So so right now, our methods were not... We we don't have evidence that we're seeing kind of this, like, hyperbolic growth. So more than exponential. Like, we're not seeing, like, that is getting easier and easier to get more algorithmic improvements. But, also, our methods wouldn't be able to find this, or our methods wouldn't be able to confidently find And in fact, several of the lines of evidence that, the report reviewed, find that it's possible we are seeing this kind of, like, rapid acceleration. But again, it's very hard to rule out at this point, but it's also very hard to be certain that we're seeing this kind of hyperbolic growth. [00:04:13] S1: What do you mean your methods would not be able to find it?
00:04:13 · Ruby Justice Thelot, Parv Mahajan
[00:04:13] S1: What do you mean your methods would not be able to find it? [00:04:16] S0: Yeah. Yeah. So so, basically, some several of the methods that we're using at the time, like, we we... There's not that much data, right? There's only so many internal models, and we only have about a year of data. Also, the methods that we... It... It's hard to talk about this because much of the evidence... Important evidence here was redacted. But, like, basically, the lines of evidence that we used, which included, like, interviews with researchers, self reported uplift, investigations of different parts of the training pipeline, and contributions of both AI and humans to different parts of the training pipeline, and Anthropic ECI, we're not confident that any of them would be able to show this kind of hyperbolic growth you're describing even if it did exist. So I think this is, again, like a method development question and, like, a open science question. [00:04:57] S1: Like, we just have no idea how to measure hyperbolic growth if it were occurring? [00:05:01] S0: Yeah. So there's, like, a couple things that would be, like, very obviously high... Like, if we saw hyperbolic growth, we would expect very obvious things. Like, for example, if, like, a 100% of the architectural innovations at Anthropic or another company were happening because of, you know, because of AI, then we could be like, oh, yeah. Hyperbolic growth is probably taking place. Currently, we don't have enough evidence, and there's especially not enough public evidence to rule this in or out. Another
00:05:01 · Parv Mahajan, Ruby Justice Thelot
[00:05:01] S0: Yeah. So there's, like, a couple things that would be, like, very obviously high... Like, if we saw hyperbolic growth, we would expect very obvious things. Like, for example, if, like, a 100% of the architectural innovations at Anthropic or another company were happening because of, you know, because of AI, then we could be like, oh, yeah. Hyperbolic growth is probably taking place. Currently, we don't have enough evidence, and there's especially not enough public evidence to rule this in or out. Another [00:05:25] S1: thing that I found kind of strange was like, a lot of people reacted very strongly to the graph that METR published of ECI accelerating after October 2025 or so. Except it looked to me trend line of the public models was basically the same as the trend line of the internal models, just shifted around a month later. So why do you think people reacted so strongly to this in particular?