00:00:00 · Jeffrey Ladish, Ruby Justice Thelot
[00:00:00] S0: It seems like we might be only, like, a year away from AI agents, which could, you know, take over all of the digital infrastructure in the world. There's a whole bunch of implications for AI development more broadly, but to me, the most important implications are how far are we away from AI agents that we we would lose control of entirely. Mhmm. How far are we away from AI agents, which, like, if they wanted to, could take over at least the digital world? [00:00:29] S1: We are back. We are live once again with Jeffrey Ladish, executive director of Palisade Research, which studies dangerous AI capabilities. Pretty topical at the moment. Jeffrey, welcome to MTS. Welcome back. [00:00:41] S0: Hey. Yeah. Great. Great to be back. [00:00:44] S1: So we were going to talk about Black October hacking, whatever. Mhmm. Now it seems like the big story of the day is this METR report, which says basically AI progress is quite a bit faster than previously expected, and this is largely due to AI driven acceleration. So what did you make of this report? Like, you know, there are a lot of people who were like, this this should should cause a substantial update towards, like, much shorter timelines. [00:01:19] S0: Yeah. I mean, I think it's really notable that this is a faster speed up than Daniel Kokotajlo's AI 2027 scenario predicts. [00:01:28] S1: Mhmm.
00:01:30 · Jeffrey Ladish
[00:01:30] S0: Don't know, you remember a year ago, a lot of people were down on Daniel Kokotajlo because he updated a little bit towards longer timelines than as predicted in AI 2027. And now he's saying, no, I think it actually might be shorter timelines than we predicted. And I don't know. The AI futures project guys are the people who I know who have the most rigorous models, and they're looking at all the parameters and trying to figure out what are the inputs, how will this impact the overall speed. And I'm just like, man, that's insanely fast progress for something that was already going really fast. So I mean, my initial reaction is... I don't know. I think I've been kind of skeptical overall about AI company claims about how close they are to recursive self improvement. Not maybe that skeptical. I think I'm like, well, I maybe expect it in 2028, fully automated AI R and D. Whereas I think I talk to a lot of people at Anthropic and they say, no, I think in early to mid twenty twenty seven, we'll probably be able to do this. Most of the AI development will be happening autonomously by AI agents. I think it's extremely aggressive in terms of that would transform the world so fast. But I've been a bit skeptical because I don't know. And and not because I think people are lying. I don't think they're lying. I think they're like, it's just easy to sort of be within a company and like, see a lot of progress and sort of drink your own Kool Aid and be like, yeah, think we can do it, know, whatever. But I'm like, man, I think METR is... I just take METR's reports really seriously. I think they have a really good track record at being extremely rigorous and, like, not bullshitting. And when they say, like, hey, here's the situation, it looks like a 2x speed up to us than previous trends based on AI uplift. I'm like, that's pretty good evidence. I will say, I think one thing important to note though is that METR said that they had to redact most of the best evidence for this in the report. So they're basically saying, this is our assessment, but we don't think, if you don't trust us, you shouldn't trust this assessment. We weren't able to share the data with you. Anthropic is not sharing the data with you. And I'm like, this seems not good. I trust METR, but I don't think that the world should have to trust METR on this one.
00:03:37 · Jeffrey Ladish, Ruby Justice Thelot
I think that Anthropic should release this data, unredact this data, so that whole world can assess this for themselves. Because I'm like, man, this has huge strategic implications for for national security, for US China stuff, for, like, you know, for all of this stuff. And, yeah, it it seems like we should know. You know? We should release the data. [00:03:57] S1: Look. I definitely don't believe METR is bullshitting, but, like, one thing one thing that stood out to me is that, like, the the main piece of evidence that people were looking at to confirm this, like, large speed up in AI capabilities was the benchmark chart, which shows a... You know, two different trend lines, the trend line from before October 2025 and then the trend line from after October 2025. [00:04:19] S0: Yep. [00:04:19] S1: But it seems to me that, like, the trend line of public models on this chart is not that different from the trend line of internal models. So why was today's report such a big update when we should have already seen the trend line of public models? It looks like the same line to me.
00:04:38 · Jeffrey Ladish
[00:04:38] S0: I mean, I I think... Yeah. You're you're saying, like, this is, how the overall capabilities seemed... Seems to have changed. Yeah. That's a good question. I'm looking at that tweet. Yeah. I don't don't have a very good answer to that. I'm like... My... I I feel like it's hard to look at your subjective sense of exactly how good the models are. I don't know. I think for me, the biggest the biggest jump was like, the difference between pre Fable and post Fable to me was huge. Like, I just went from, like, using agents sometimes to using agents every single day. And, like, that was, like, the really big jump. And now it's a little bit like... I don't know. I think I'm like, I really want Anthropic to release, us to Anthropic. Show us what is it like being an Anthropic engineer? How many agents are they using every day? What does that actually look like? And then concretely, what does that lead to? In some sense, I'm like, yeah, we can look at these benchmarks. That is, I think, best thing we have for measuring AI capabilities. But I remember back in the day when the way we measured AI capabilities was mostly METR bringing in software engineers and having them do these long time horizon tasks. And then it was not like, you have to look at these 40 benchmarks. It was like, look at just like, can the AIs autonomously do the work? That is the thing I most care about. And so I'm like, yeah, I don't know. I think this is a fair criticism of like, well, these benchmarks aren't very good, but I'm like, how do we get back to the world where the benchmark is mostly like, let's take real economically valuable work and see if the AIs can do it autonomously? I think that's what I would prefer to be the measure.
00:06:17 · Ruby Justice Thelot, Jeffrey Ladish, Ruby Justice Thelot
[00:06:17] S2: Yeah. [00:06:18] S0: But but I'm, like, a little bit, like, I'm not sure what METR is supposed to do. Right? Like, it's, like, a tricky it's a tricky problem. [00:06:25] S1: I mean, plausibly, like, if if METR really believed that they had access to information that was, like, world historically important that could prevent humanity's extinction or disempowerment Yeah. Presumably, they would outweigh that in their consideration, or or or rather that in their consideration of whether to release that information would outweigh, like, Anthropic's IP concerns or whatever. And potentially, whistleblowers within Anthropic would make the same calculus. [00:06:54] S0: Well well, I do I do not think that METR should release information they have agreed with Anthropic Not's release. Like, I think that, like, the trust there is really important. I'm like, the fact that they can play the role that they're currently playing really relies on trust. Now I think there are potentially cases for Anthropic whistleblowers to come forward there. My hope is that anthropic leadership themselves will be like, hey. Or like, if I was at anthropic, the first thing I would do is I'd go to leadership and I'd say, this is of world historic importance. You guys agree that this could, like, result in human extinction. We need to be much more transparent with this information. This is really decision relevant information that the world should know.
00:06:54 · Jeffrey Ladish, S3
[00:06:54] S0: Well well, I do I do not think that METR should release information they have agreed with Anthropic Not's release. Like, I think that, like, the trust there is really important. I'm like, the fact that they can play the role that they're currently playing really relies on trust. Now I think there are potentially cases for Anthropic whistleblowers to come forward there. My hope is that anthropic leadership themselves will be like, hey. Or like, if I was at anthropic, the first thing I would do is I'd go to leadership and I'd say, this is of world historic importance. You guys agree that this could, like, result in human extinction. We need to be much more transparent with this information. This is really decision relevant information that the world should know. Ad[00:07:31] S3: I wanna shout out our sponsor, Lovable. You know the app you've been meaning to build? The internal tool, side project, or product you'd otherwise lose a weekend to? With Lovable, you get software you can ship today. That includes an editable codebase you can inspect and change, plus two way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more so you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable, lovable. Dev. Now onto the episode.