Ryan Greenblatt – What happens once AI can automate AI research?

Dwarkesh Patel 0:00 Today, I'm chatting with Ryan Greenblatt, who is the chief scientist at Redwood Research, where he focuses on technical AI safety and security work. I want to talk to you about recursive self-improvement. This is the idea that once we build human-level intelligences, they quickly slingshot towards tens of billions of superintelligences, which are each individually more competent than the top human experts across every field. Whether or not this turns out to be the case, I think is actually probably the most important question in the world right now. And historically, I've been quite skeptical that this kind of thing happens, but, um, you seem to think that it might be plausible, and so I wanted to hear the case for it.

Ryan Greenblatt 0:36 Yeah. Let's talk about this. So first, I think it's worth noting that AI R&D is a type of task at which the AIs are especially good because both the companies are trying really hard to make their AIs good at AI R&D, and it's the kind of domain... It has, it has a lot of nice properties from the perspective of how AI development works right now, so it's, like, pretty verifiable. You can do a bunch of stuff iteratively, and it'll climb on various metrics. And then I think once you have AIs which are roughly matching the top, um, human experts in AI R&D, that could sort of kick off a feedback loop where, you know, the AIs are doing AI research. That produces smarter AIs. That feeds back in. And that feedback loop could be strong enough that you end up with a lot of progress in a short period of time. Maybe my sort of median expectation is something like, uh, four or five years of AI progress in a single year, and this requires really overcoming a huge amount of diminishing returns in research and basically doing the equivalent of what progress we would've gotten after a really large compute scale-out. So this is, like, a pretty impressive, big thing, and it's worth keeping in mind that five years of AI progress, four years of AI progress, even three years of AI progress is really a lot of fucking AI progress, right? So, you know, right now, it's, like, um, three years ago or a little over three years ago, there was GPT, um, four, uh, that had come out, and right now, of course, we have, like, you know, Mythos-5 or whatever and maybe a somewhat better that-- model that Anthropic has internally. Um, and so that is just a huge amount of progress in a bit over, uh, three years. And if we're talking about five years, then maybe we're talking more about, like, a jump from, uh, you know, GPT-3 to, um, Mythos-5 or whatever.

Dwarkesh Patel 2:06 Yeah. Okay, so I think this argument has three different parts, and now I want to evaluate each one of them. First is the argument that AI R&D is very verifiable. Second is the argument that if you automate AI R&D, you could get four or five years of progress in a single year. And third is the argument that what comes out the other end of four or five years of AI progress at the current pace, starting at the current or the start, starting at the starting point whenever AI R&D is automated-

Ryan Greenblatt 2:30 Yeah

Dwarkesh Patel 2:31 ... what comes out the other end is an AI where you can drop it, um, on the job at basically anything you can imagine. You can drop it in Texas politics in the 1940s and it outmaneuvers Lyndon Johnson. You can drop it in, um, I don't know, TSMC, and it, like, learns how to d- do, does better process engineering at TSMC. It's, uh, certainly a better video editor than, um, I... My video editors are very excellent.

Ryan Greenblatt 2:54 [laughs]

Dwarkesh Patel 2:54 But it is just, it is just, in general, better than humans at any g-given job that it finds itself trying to do. So I want to evaluate all of these, um, sub-arguments, uh, that lead to basically getting ASI pretty soon after this benchmark, which you're expecting by 2030 or something, right?

Ryan Greenblatt 3:10 Yeah. I would say that I expect, like, full automation of AI R&D perhaps somewhere around, like, 2031, 2030, and then getting to, like, the, like, beats-all-humans-on-the-job milestone, maybe I expect median around 2033. But sort of, like, if I see AIs fully automating AI R&D, I think I'm expecting that probably within a year. It just, like, uh, the way the forecasting works out, there's a, the difference between medians is bigger than the median difference between milestones. Anyway, whatever.

Dwarkesh Patel 3:36 B- by the way, I, I, [laughs] there's this meme on the internet 'cause every time I'm trying to ask about people's timelines, when I'm asking Dario or somebody, I'm always like, "Okay, how long before we can automate my video editors?" And [laughs] there's this meme of, like, my video editor editing the podcast. [laughs]

Ryan Greenblatt 3:51 [laughs] Yeah, yeah.

Dwarkesh Patel 3:51 Every time I listen to this. But the reason I do it is because I think it's easy to get lost in-

Ryan Greenblatt 3:55 For sure

Dwarkesh Patel 3:55 ... abstractions when you talk about jobs you don't understand well and to very concretely understand what it takes to automate a job that I actually understand why it's difficult for LLMs to currently take control over. Okay.

Ryan Greenblatt 4:07 I do think that the, the, the, the milestone for automating your video editor is earlier than the milestone of being able to automate all human jobs, including, like, you know, Texas politics spinning up on the job. So I, I think, I do think that the video editor automation-

Dwarkesh Patel 4:18 [laughs]

Ryan Greenblatt 4:18 ... maybe occurs more, like, around full automation of AI R&D, but it's very sensitive to how much people are really focusing on understanding video.

Dwarkesh Patel 4:24 Yeah. Okay, so let's start with the claim that AI R&D is very verifiable.

Ryan Greenblatt 4:29 Yeah. So there's a few different parts of this. One of them is that we can train on a bunch of environments which are, like, basically directly training the model to do some AI R&D task or some very close-by task. So for example, we can have some environment where the model is training some AI on just, like, eight H100s or whatever or, like, some small amount of compute, and that model could be, like, you know, the equivalent of, like, GPT-2 medium or whatever and, and then, uh, you know, similar to, like, um, Nano-GPT medium runs or whatever, and in RL, it's, like, tweaking and iterating on that. And we could do that for a bunch of different tasks. Like, we could have it train, like, image classification models, video generation models, image generation models, all kinds of different sort of ML training tasks, and we could RL it on the task of training increasingly good models and also doing things like, "Oh, here's a particular direction you could pursue for an algorithm. Can you go and implement that?" And so basically, there's this whole class of containerizable, verifiable, small-scale AI R&D tasks that we can aggressively RL the AIs on. And I would say that already companies are presumably doing some RL on these sorts of tasks, and you could just keep scaling that up, keep making more of these sort of small-scale AI R&D tasks, and then the AIs could, you know, keep getting better at this. And, and then implicitly, I'm claiming this will transfer to extremely load-bearing aspects of AI R&D, but maybe let's stop there for a second-

Dwarkesh Patel 5:45 Yeah

Ryan Greenblatt 5:45 ... and then we can get to that part.

Dwarkesh Patel 5:46 So let's talk through what this concretely looks like. So you can imagine that we have GPT-7.5, and we say, "GPT-7.5, we wanna make you so good at AI R&D that you help us train GPT-9." Okay, so now we train, we wanna train GPT-7.5, and we come up with a bunch of different environments. Like, w- as you mentioned, we could do f- there's already this repo that is the descendant of Andrej Karpathy's Nano-GPT speedrun, where you just try to change everything about the model from, like, the optimizer to the hyperparameters to the architecture to get it to get to a fixed training loss as fast as possible. Um, you could have other kinds of, uh, environments where you could say, "Hey, uh, GPT 7.5, I want you to train a really good video, video game-playing model, and I want you to train a model that actually improves as it plays the same video game again and again, so you learn how to maybe help the model get better at online learning. Maybe it, it gets-- W- we don't care how you figure this out. Maybe it's some kind of crazy neural ease or a vector memory-

Ryan Greenblatt 6:39 Yeah

Dwarkesh Patel 6:39 ... maybe it's some crazy-- or maybe just, like, better long-context stuff. We don't care. Get-- figure out how to, like, do online learning research." Obviously, then the fact that GPT 7.5 will already have become very good at normal, like, become-- It'll be a smart model, and in the same way the model's currently getting smarter, it'll be better and better at coding in the way that models are currently getting better at coding. And then you can imagine 100 other environments like this, which are incentivizing the ability to do AI R&D by getting GPT 7.5 to, like, c-containerized versions of getting GPT 7.5 to develop GPT 2 mo-size models, et cetera, et cetera. And you sh- basically, then you've, like, you put GPT 7.5 through a bunch of this kind of training. You build GPT 8, and GPT 8 is now an amazing ML researcher. It has so much intuition from doing all of this kind of training. Um, honestly, a huge intuition pump for me is seeing the progress that AI has made in mathematics, where I'm just like, if it's a very verifiable domain, AIs can get even, even-- Like, mathematics also involves so much, like, s- I, I don't really know this object-level details of, like, mathematics research, but I'm just like, "No, it works." Like, it can just come in like a flood if you can totally put it into a verification loop, and it can actually make new breakthroughs. Um, I am curious if ML research has a quality of mathematical research where it seems like there was a big overhang from connecting different disciplines together or ideas that were not-

Ryan Greenblatt 7:58 Yeah

Dwarkesh Patel 7:58 ... no one person would have known enough about algebraic geometry and-- W-what was the right word?

Ryan Greenblatt 8:04 Oh, man, I really don't know about the, the [chuckles] math breakthroughs.

Dwarkesh Patel 8:07 Well, no one person would have known enough about topology and, um, algebraic whatever-

Ryan Greenblatt 8:12 [laughs]

Dwarkesh Patel 8:12 ... blah, blah, blah, in order to make some counterexample to a big conjecture.

Ryan Greenblatt 8:16 Yeah. My view is that ML is a less deep domain than math, and so there's less of a thing where there's, like, individual experts with really deep expertise in some area that they combine, but there's definitely gonna be some of that. But then I also think that ML has some attributes that make it even more favorable than mathematics in some ways to, um, you know, AI training. In particular, there's, um, you can get a better sense of whether you're succeeding, and you can see intermediate progress. So in, in math, it's often the case that sort of, uh, there's no easy way to see whether or not you're close to success. Whereas if your goal is to, for example, um, get to some training loss, you know, two X faster, you can kind of see when you're halfway there, and it tends to be the case that ML innovations are very additive or maybe multiplicative, depending on how you think about it, where basically you can keep stacking innovations, and usually the innovations just sort of just add together and don't interfere with each other, though obviously it's gonna depend on the details. And so I think that in a lot of ways, AI R&D will have properties, um, you know, qui-quite similar to math, where basically you can do small-- You can, like, train on chunks of AI R&D that are pretty similar in structure to the, to the problem you actually cared about, um, in a very verifiable way, and then that will transfer. And then there's an open question of exactly how well it will transfer, but I think that the transfer currently for math looks pretty good, and my expectation is that the transfer for AI R&D will look pretty good, but not amazing.

Dwarkesh Patel 9:37 So one concern I have is, I think even in mathematics, as far as I'm aware, we have not seen very impressive new theory. We've seen a lot of, like, impressive, verifiable, specific results. For example, find a counterexample to this conjecture. But we have not seen, like, come up with the idea of topology kinds of levels of things-

Ryan Greenblatt 9:55 Yeah

Dwarkesh Patel 9:55 ... or come up with things like group theory. And it seems like ML, uh, research has elements of both of these things. But the less verifiable thing of, like, come up with new ways of thinking about the problem, uh, would be harder to induce. So it takes, for example, the idea of scaling laws. Um, obviously, there is some end verification loop such that you can train GPT-4 better if you have the idea of scaling laws from, like, 2020. Um, but there is a longer and potentially more compute-laden and, like, uh, uh, road to getting AI-- inducing AIs to be like, "Okay, I gotta think carefully about how I should be scaling my parameters and data. How-what are different kinds of investigations I could run to understand this? Maybe I can, like, come up with the visualization in, like, a ISO flop analysis or something." Um, but that does seem, like, hard-- That, that, that does seem like a longer verification loop than just, "Hey, let's get nano GPT laws to go down."

Ryan Greenblatt 10:47 Yeah, let's talk about this. So first of all, I think in the context of math, the thing I would say is that the AIs can do the equivalent of, like, baby's first new theory or whatever, [chuckles] where, like, for example, they can just, like, prove interesting conjectures via, like, making connections and, and producing new understanding of, like, oh, there's this, like, thing the AI-- this, like, construction the AI found which is pretty interesting or, like, found this, like, way of thinking about the problem that's a bit different. Um, and we do just see that. It's just that the examples we see are not, like, as impressive as, like, founding the field of group theory. [chuckles]

Dwarkesh Patel 11:18 Yeah.

Ryan Greenblatt 11:18 But, like, in part, you know, probably founding the gr-field of group theory is, like, one of the, you know, it's, like, among the best, biggest mathematical accomplishments of all time, and the AIs just aren't, you know, they're not that good at math yet.

Dwarkesh Patel 11:28 Yeah.

Ryan Greenblatt 11:28 And I think that from my perspective, sort of there's a continuum between that and the things we're seeing now that the AIs are continuing to march up. Second, I think ML is a very shallow domain relative to math. So I think in math, there is much more of a you find some true, deep abstraction, um, and then, like, that, like, if you really understand that thing, which is hard to understand, then you get somewhere. Whereas I feel like the things that are the equivalent of that in, in ML are really, like, dumb bullshit. [chuckles] Like, um, like scaling laws. Like, like come on, guys, we can explain scaling laws really quickly. And I think the, like, deepest and most important concepts in math, for example, don't, don't have the property of, like, you can really understand the underlying thing and why it matters in a very short period of time. Um-

Dwarkesh Patel 12:09 But I, I, I feel like one effect will be that we will have gotten rid of all the low-hanging fruits by 2030. Like, I feel like scaling laws will have been in, like, what math history, Descartes, you know, finding the Cartesian grid and, like, very, doing very-

Ryan Greenblatt 12:21 Yeah

Dwarkesh Patel 12:21 ... basic mathematics-wise, and then eventually we wanna keep making progress in the 2030s, it's gonna be like do whatever bullshit is happening at, like, the frontiers of mathematics right now.

Ryan Greenblatt 12:29 Yeah, that could be right. My sense is that just, like, some domains are structurally different in terms of how they operate and how much they depend on, like, sort of deep abstractions, and, like, physics and math are much more in the side of, like, being very far on the, like, Sort of very deep, hard-to-come-up-with ideas side. Whereas I think ML and, and most other domains are much more amenable to sort of hill climbing. Um, and that's my sense of how this will go in the future. And even in the regime where your AIs are like, you know, having to plow... Like, it- it's the 20- it's 2030, they need to like, a bunch of low-hanging fruit in research has already happened. They need to, like, make further progress. I still suspect that a bunch of the work will live more in the side of, like, building increasingly complicated infrastructure, having really good intuition about what the experiments roughly look like. And so I think I'm probably less sympathetic to, like, the, like, thing that the AIs will lack is, like, some deep insight, and more sympathetic to, like, they really need a bunch of, like, taste about in-the-weeds experiments that they currently don't have, and need to have a bunch of intuition for, like, what sorts of, um, training approach would work and what wouldn't work in ways that current, um, researchers have. And even in cases where, um, there has been some breakthrough in AI, oftentimes in retrospect it looks like a big bottleneck to making that breakthrough happen was sort of getting all of the, like, microdetails and mungy intuition right. Um, like an example of this is when it comes to, like, training AIs with, um, to, to be good at reasoning and chain-of-thought and doing sort of RL and chain-of-thought training, it looks like you probably could have done RL and chain-of-thought on, like, GPT-3 and gotten kind of interesting results on math if you had really scaled it up and done a good job. But at the time, there was low-hanging fruit, and also doing a good job with that training is, like, kind of, like, in the weeds on all the technical implementation and scaling it up and getting the hyperparameters right. And so maybe you can demonstrate everything on, like, Qwen 1B or whatever, um, and get some sense that this whole thing is gonna work. But people didn't demonstrate it as early as they could have because, like, you know, of all of these other, like, mungy details and, and, and intuition about exactly how to tune the parameters and how to set things up.

Dwarkesh Patel 14:28 This is my remaining skepticism, honestly, about this story, is just, um, I am, yeah, I'm not s- I'm not sure I understand why if research breakthroughs are so amenable to intelligence why AI progress has not been historically faster than it could have been, and we had to wait for, as you were saying, o- like, w- by the time RLVR actually worked, even though you could have done it with, like, less compute, we had to wait for oceans of compute, and like-

Ryan Greenblatt 14:54 Sure

Dwarkesh Patel 14:54 ... gigawatts of compute to be available before people are, like, doing this training, um, on the trajectory of, like, this constant, you know, as compute keeping s- increasing, we make more breakthroughs. I don't know. I feel like there were a lot of AI researchers in the year 2022 who were trying to crack reasoning, and it was just that they were, like, bottlenecked by the ability to write infrastructure code? Or, like, what was happening?

Ryan Greenblatt 15:13 It's a complicated mix, right? So I think that they would've gone faster if they could, like, as soon as they thought of an experiment, run that experiment without bugs, without bugs being very important. And then I think another part of it is that, like, being able to run a lot of com- experiments at high compute lets you paper over ways in which the way you implemented it isn't quite right or you didn't have the right hyperparameters. And so I think compute is just, like, really helpful for doing AI research. And you can, like, you know, uh, cover over a lot of things. But that doesn't mean that massive increases in labor wouldn't also be helpful, especially if that labor comes with, you know, among the best intuitions that people have in the field. Um, I just think that that's, you know, really helpful. I think another part of my perspective here, which is maybe a bit different from where you're coming from, is that I think I'm expecting somewhat more transfer than you seem to be imagining. And I'm imagining these AIs are actually, like, pretty good scientists in general, um, and are just like, you know, pretty, pretty reasonable at all of that stuff. And just sort of when you were to interact with them, it's not like there's some, like, really hyper-specialized savant-type vibe. They're actually just, like, pretty good at all of the stuff in AI R&D, and then maybe, like, extremely good at some subdomains, right? So they're, like, incredibly superhuman at writing kernels, incredibly superhuman at everything with very short feedback loops, and then, like, um, you know, pretty good at all the other stuff and, like, you know, just totally able to match other people. And, like, I think we are seeing this now. Like, I would say that when I look at AIs right now, I think it's already the case that they can pretty competently match, like, humans who are mediocre at ML research at doing ML research. It's just that being mediocre at ML research is not that helpful, right? [laughs] Um, like, the thing that you actually want are people who are good at ML research. And so my sense is the AIs are just improving at all of these things. Their taste is improving, their intuition is improving, and it's already the case that their taste and intuition is not, like, it's not, like, complete garbage.

Dwarkesh Patel 16:51 Yeah.

Dwarkesh Patel 16:52 I, I, so I, I wanna very concretely understand what it would look like for, uh, five years of AI progress to happen in one year.

Ryan Greenblatt 16:57 Yeah.

Dwarkesh Patel 16:57 So suppose we were back in when, like, GPT-3 was developed, and the, the idea is not only that, like, basically with the level of compute they've had back in 2022, you could have trained, if we had automated AI R&D back then, you could at the end of that year have Mythos.

Ryan Greenblatt 17:13 That'd be the idea, yes.

Dwarkesh Patel 17:14 Including, with the, like, so m- Mythos took way more compute than they had back then. But, like, even with the level of compute they had back then, not only do they all the, do, do all the breakthroughs, but they also trained Mythos with th- their level of compute. And what would be required is, um, obviously, like, discovering all the algorithmic progress since then, discovering even more actually because you gotta make up-

Ryan Greenblatt 17:31 Yeah

Dwarkesh Patel 17:31 ... for the fact that, like, Mythos uses, I don't know, w- what, what was GPT-3 trained on? Like, 1E23?

Ryan Greenblatt 17:37 Uh...

Dwarkesh Patel 17:38 We can look it up. But is it plausibly four orders of magnitude more compute?

Ryan Greenblatt 17:42 Yeah, I think it's somewhat less than that. Let's, let's look this up quickly. Um, so GPT-3 training compute, um, is, yeah, it's, like, 3E23. Um, my sense is that, um, Mythos is probably about, um, a little over three MOEMS higher. And so the question is can you overcome this 1,000X compute gap while also, you know, um, being the model? So here's a concrete claim that maybe we should, we should talk about. Like, right now we would be able to train a model with GPT-3-level compute that matches, um... Yeah, what exactly do I think? Um, so GPT-3 was, let's say, uh, about, um, yeah. When was it trained? So it was trained, um, it was trained in, it was released in 2020, so it was trained six years ago. It's worth noting that GPT-3 is maybe a little too far away-

Dwarkesh Patel 18:28 Mm

Ryan Greenblatt 18:28 ... or too, too far in the past, but let- let's go with this for a second. So GPT-3 was trained, like, about, you know, uh, six and a half, seven years ago. If we were to train a model with GPT-3-level compute today, how good would that model be? Um, my understanding is based on- like how algorithmic progress works, we'd be able to train a model that's as good as the best model we had perhaps around three years ago. So I think that right now we'd be able to train a version of GPT-3 that's probably somewhat better than GPT-4 is basically what we'd, we'd see. Um, probably a, yeah, like a moderate amount better, um, than GPT-4, and I think that's about right. I think that roughly lines up with what, with how algorithmic progress has worked. Basically, the story would end up being that to get five years of AI progress, you're probably gonna need around, I would say, like maybe eight years of algorithmic progress, very roughly. Um, which is a lot, a lot of algorithmic progress.

Dwarkesh Patel 19:16 Right.

Ryan Greenblatt 19:16 But it just turns out that, like, most of the AI progress from my perspective has come from some mix of like algorithms and data, and you can just keep making, like, for, um, I think, huge improvements on these things and training AIs with less compute. So-

Dwarkesh Patel 19:28 Well, that, I- I'm glad you brought that up because what has happened since GPT-3 or even 3.5 till now, right? Like, why is Mythos so good? Um, obviously we've scaled the compute. We have better algorithms. A huge thing that's happened is that we have built a deca-billion-dollar data industry which has systematically collected and codified, uh, expert human judgment across all kinds of different disciplines, codified in the form of RL environments, codified in the form of SFT traces, that these experts built to b- help the model better understand how do you do coding and, like, how do you build con- complex infrastructure projects? How do you do, like, law? How do you do whatever, whatever? And I-- how are the AIs able to replicate the effect that currently expert human judgment seems to be playing in AI progress?

Ryan Greenblatt 20:18 Yeah. So my sense is that scaling up the amount of effort spent on getting expert human data has not been hugely important for AI R&D in general. So in particular, like, you know, over the last few years, we've been scaling up compute scaling up people working at AI companies, and scaling up the amount of effort spent on data labeling. My sense is that if you, like, sort of remove like the last, like two doublings or whatever of data labeling, that would not make a huge difference, or data generation, that would not... I'm sorry. I should say data generation from expert humans, that would not make a huge difference. Um, and I think a lot of what's been going on is people have been developing better ways to leverage, like, humans and AIs to, like, construct RL environments and going somewhere from that. And so-

Dwarkesh Patel 20:59 Wait, so but how, how do you, like how do you explain why the AIs have gotten so good at coding? I feel like a big part of that is data and RL environments, which are like-

Ryan Greenblatt 21:06 Yeah

Dwarkesh Patel 21:06 ... codifying human experts.

Ryan Greenblatt 21:07 But the question is what is the limiting factor on creating RL environments? My sense of the limiting factor on creating RL environments was not so much like, um, like scaling up, or like the thing that drove r- the, the reason why RL environments today are much better than they were in like, you know, 2024 is not that much because, um, we have hired way more human experts to make RL environments. It is instead much more because we better know what, how RL, like what RL environments we even want to make and, and like how we should structure them, and also we're using huge amounts of AI labor to build RL environments. Um, and I think those effects are much more important than the effect of, uh, human labor building the RL environments.

Dwarkesh Patel 21:47 Um-

Ryan Greenblatt 21:49 I'm not saying that the, the human labor doesn't matter. I'm just saying there's other, there's other big drivers that are important here. Um, yeah, me- I could try to, I could try to argue for this. I mean, one, one thing is just like the amount of environments people want are just, like, they're very, it's a very large amount. Um, and I think the AIs are actually pretty good at the task of making RL environments, given some, some sense of what the, the thing should be. There's preexisting data you could use. I don't know. A lot of these things have good verification loops.

Dwarkesh Patel 22:14 If I just look at, for example, this was reported in Business Insider today, yesterday, that, uh, Google is paying like two, close to two billion dollars for Mechanize.

Ryan Greenblatt 22:23 Yeah.

Dwarkesh Patel 22:23 Um, like the-- we can just look at market rates for what people think really good-

Ryan Greenblatt 22:28 Sure

Dwarkesh Patel 22:28 ... human experts making, like human expert data is worth, and it just seems to be like the, the, the Frontier labs need to think it's worth a lot-

Ryan Greenblatt 22:36 Yeah

Dwarkesh Patel 22:36 ... in order really to pay for it.

Ryan Greenblatt 22:37 What fraction of, of Frontier lab spending do you think is on data rather than compute? Like what do you think is the compute/data spend split?

Dwarkesh Patel 22:44 I think it's most like ov- overwhelmingly compute, but I also think it's because, like, compute is easier to scale up than data. Like I know that's like-

Ryan Greenblatt 22:49 But that's really relevant to what's driving progress, right? It's like suppose, like I agree that, yeah, like my, my sense is of the split is something like I would have guessed like 20 to 1 or something, 10 to 1. I don't know exactly.

Dwarkesh Patel 22:59 But-

Ryan Greenblatt 22:59 It depends on the company. But-

Dwarkesh Patel 23:00 I mean, but this is similar to like oil is 1.5% of GDP.

Ryan Greenblatt 23:03 Sure.

Dwarkesh Patel 23:03 But that mean, but that doesn't mean if you cut oil out, you could, like, GDP could continue to run.

Ryan Greenblatt 23:07 Sure. But, but the counter to your argument, right?

Dwarkesh Patel 23:08 The economy would, like, come to a halt immediately if, like, oil went away.

Ryan Greenblatt 23:10 Sure. But you, you are just arguing that because of the high market cap, we can learn that this is the key driver, and I'm saying that's not clearly true, right?

Dwarkesh Patel 23:16 Sure.

Ryan Greenblatt 23:16 Because, like you, I think that argument just implies, looks, makes it look like compute is a much more important driver or like hiring employees is a much more important driver.

Dwarkesh Patel 23:22 So, so maybe le- let's be more concrete. Here's what I-

Ryan Greenblatt 23:24 Yeah

Dwarkesh Patel 23:24 ... here's what I think. Just the same way as in my claim is that in GP- if you went back to 2022 and you had GPT-3.5, and you were like trying to make it better at coding, uh, without human experts, I think it would've just been very, very difficult. I, let me give you an example of what-

Ryan Greenblatt 23:39 Sure

Dwarkesh Patel 23:39 ... I imagine would be the difficulty from going from, uh, GPT-8 to ASI. So one of the things you'd need GPT-8 to be good at or like you'd want ASI to be good at is like, I'm going to like take over a company and, like, make it much more profitable and, like, do all kinds of crazy shit to make it work better. I'm gonna like take over a fab and like m- produce more chips. Um, like, uh, this is the, like the tier of data that will-- I'm, I'm going to like go into Congress and try to convince them of, to pass some bill, blah, blah, blah.

Ryan Greenblatt 24:05 Yeah.

Dwarkesh Patel 24:05 This is what I imagine five more years of AI progress at this pace would enable an AI to be able to do. This is what, the thing I'm really worried about, right? Like the ASI that can, like, understand how to do crazy shit in the world. Like-

Ryan Greenblatt 24:15 Yeah, for sure

Dwarkesh Patel 24:15 ... can do what Kissinger can do, can do what like, um, Steve Jobs can do, et cetera, and also his engineers and stuff. And I'm not sure how you get that without the relevant wor- world data, which is the equivalent of Mythos being really good at coding while not having the coding environments that have improved it relative to GPT-3.

Ryan Greenblatt 24:35 Yeah. So here are a few points. So first, I bet if you look at sort of randomly sampled training environments for Mythos, they're actually very different from what it looks like to actually use the model in practice. My sense is that the RL distribution has, like, really large deviations from the real world data distribution, and it's significantly being sort of like smoothed over by a mix of transfer and having a small amount of data focused on the real world. And so I, my sense is that this, this will be a similar mechanism as how it works, um- For like the, you know, crazy, wild, like quite superhuman AI you get as a result of five years of AI progress on top of fully automated R&D. So let's just like go through this a little bit. So in particular, um, I think that you could train an AI to be really, really good at, um, learning on the fly and s- doing something analogous to in-context learning, but potentially using somewhat different mechanisms in a wide variety of RL environments. So you build all these different RL environments where the AI has to like adapt on the fly, learn on the fly, figure out what it should do, understand its situation better, um, and like learn really quickly from feedback in order to succeed at, at its objective and has things like limited resources and if it like messes up, it can like end up in a much worse position. And then if you train on a huge number of these environments, you will learn sort of general skills of like picking up context on the fly, and we're already seeing this. Like, it's already the case that AIs are now much better at sort of understanding roughly what's going on and like picking up context from a, um, you know, limited amount of, of information they're given access to. And then those AIs could then be put on the job at TSMC, and then even though TSMC is not like literally in their data distribution, their data distribution is really wide, and the AIs are extremely good on their data distribution, such that it transfers to picking up being good at, you know, being a engineer at TSMC and, and learning that on the fly, where it looks more like the way the AI gets good at, at being a TSMC engineer isn't that it has a ton of cached knowledge on being a good TSMC engineer. It's that it like does the equivalent of like some scaled-up version of in-context learning, um, there.

Dwarkesh Patel 26:31 Yeah.

Ryan Greenblatt 26:31 But that'd be the most prosaic story. Obviously, there's like a bunch of different ways this could go.

Dwarkesh Patel 26:34 I think this m- maybe comes down to then a difference of intuition about how far you can get. When I think about really smart people I know, they're just like not that effective in domains they don't understand that well.

Ryan Greenblatt 26:45 But how long have they had to learn?

Dwarkesh Patel 26:47 No, I, I agree that if they had experience, they would be much better.

Ryan Greenblatt 26:51 Yeah.

Dwarkesh Patel 26:51 But that's maybe what I'm arguing for, is that experience of data. Like for example, if I just get a really smart, I don't know, uh, Ivy League college grad, and I'm like, "Okay, you're now in charge of negotiating the Iran deal," I think they'd just like wouldn't know what to do.

Ryan Greenblatt 27:04 I think if you got, instead got someone who is really good at quickly picking up a bunch of different domains and you gave them some time to sort of train and talk to people and show off their expertise and do some practice, they would actually do like a pretty good job. I think most domains are fundamentally pretty shallow, where like a very smart generalist who's good at like a, a f- a, a limited subset of core skills can like get going pretty quickly. Um, and my sense is that like that's not true for literally every domain, and my sense is that the AIs will develop increasingly good mechanisms for quickly acquiring understanding and expertise in a given domain. So consider, for example, how fast AIs can like understand a new code base. Um, AIs can understand a new code base much faster than humans can but to a degree that's shallower than humans could currently understand but is, but is getting better over time, right? So let's, let me, let me spell that argument out a bit more. So let's say you take, you know, um, you know, Fable 5 or Mythos 5 or whatever, and you like want it to make some kind of complicated change to a really massive code base. Um, the model will get some understanding of the code base very fast, like in the course of maybe like, you know, significantly less than an hour, potentially much less than an hour, and then its understanding of the code base will like plateau a little bit, where it won't get as deep of an understanding as a human would've gotten over a much longer period. So it's sort of like an AI in an hour can match a human with a few weeks maybe, um, depending on the details of exactly how complicated the code base is. Um, but then it won't match a human with like, you know, who's been working on that code base for like two years or whatever. But over time, the like amount of understanding AIs can match has gone up, right? So if we look at like 3.7 Sonnet or 3.5 Sonnet, maybe it could only match the equivalent of understanding a code base for like a day or something. Um, but now, you know, AIs are, are much better at like sort of building context about a task, and so you can be like, "Mythos, I want you to really understand this code base, and then, um, you know, then implement this feature." And it will like spawn a bajillion sub-agents. Those sub-agents will pore over a bunch of things. It will like deliver a bunch of context back. It will then like investigate a few things, and it's not like amazing at doing this, but it's like it can happen like really fast, and it can work pretty well. And it's not very hard for me to imagine how you could train AIs to be increasingly good at this task, right? The task of like implement some very complicated feature in some reasonable way in a very big code base is extremely verifiable, and that can like be a thing the AIs improve on. And similarly, like there's a broader skill of like quickly u- understanding context and being able to like have a bunch of different AIs learn in parallel-

Dwarkesh Patel 29:23 Yeah

Ryan Greenblatt 29:23 ... and then merging that together.

Dwarkesh Patel 29:25 And I, I think there seems to be a crux here, which I, I think is just like an empirical question, uh, we'll see on-

Ryan Greenblatt 29:30 Sure

Dwarkesh Patel 29:30 ... which is how good is a transfer between getting really, really good at s- understanding the situation, getting up to speed, making progress over long periods in verifiable domains, which the AIs are obviously getting way, way better at really fast, to, "Okay, go talk to the president and like convince him to do X thing," or go, "You're now in charge of Google. Now you, you must make Google a much more profitable company this quarter." Uh-

Ryan Greenblatt 29:59 Let me try to just spell out a few more arguments-

Dwarkesh Patel 30:01 Yeah

Ryan Greenblatt 30:01 ... that are maybe relevant. So one thing is that I do think that like when looking at like how the AIs have improved at essay writing, um, let's talk about that a little bit. So I think there's one thing, which is that you can get some data even on these domains, and AIs will be able to get some data even on these domains when on a very fast progress trajectory. So like maybe it's hard to build like a verifiable environment for like was your essay really good according to humans, but you can do a bit of that. You know, you can do some training. You can do some online training. Um, and the AIs will be able to do some like, you know, online training based on real world stuff. They'll be able to like have evals. They'll be able to like sample that, and you can, you know, scale that up the cadence at which you do this. And then the second thing is that in practice, when I just look at the transfer, it seems okay. Like, I think that in fact the AIs have improved a bunch at non-verifiable domains, and it is in fact the case that it's hard to point to like domains that are really hard to verify on which the amount of improvement between, you know, GPT-4 and Mythos hasn't been like pretty high in practice. And, and now that doesn't mean that Mythos is like better than the best humans or something, right? It can still be like significantly worse than typical human professionals at some aspect of their job while still being like way better than GPT-4, which was like not even close.

Dwarkesh Patel 31:06 Yeah. So we're talking about how, uh, how important data versus algorithmic progress has been for- ... explaining the progress over the last few years. That, that reminds me, I'm actually running an experiment with, um, Jerry Han, who's, um, actually is, is still a college student. What we're basically doing to evaluate how much, uh, how much progress is coming from data versus algorithms is training the best algorithmic recipe from 2019 till now with the best data from, like, the 2026 data file, and then also training the different data files going back to 2019 to 2026 with the current best training recipe, like the algorithmic recipe.

Ryan Greenblatt 31:40 Yeah.

Dwarkesh Patel 31:40 And, um, I think that will be an interesting... I, I, I'm curious if you wanna pre-register-

Ryan Greenblatt 31:45 Yeah

Dwarkesh Patel 31:45 ... like what, what amount of multipliers are coming from one versus the other.

Ryan Greenblatt 31:48 So we need to be pretty careful with what we mean when we say the word data. So I was trying to be pretty careful to distinguish between scaling up spending on getting human experts to label data or, like, scaling up the amount of human-

Dwarkesh Patel 31:57 Sure

Ryan Greenblatt 31:57 ... expert label, uh, data. Pre-training data does not come. The, the reason why we have a better pre-training dataset now versus in 2019 is not because people are spending way more money getting human experts to, like, type up data that the AIs are then trained on. I think it's-

Dwarkesh Patel 32:10 Partially.

Ryan Greenblatt 32:11 I think it's not much of it. I think it's very little of the pre-training data improvements. I think the vast majority of the pre-training data improvements, which to be clear, I, I do mean pre-training. We should talk maybe separately about mid-training, post-training.

Dwarkesh Patel 32:20 Yeah.

Ryan Greenblatt 32:20 But I think the vast majority of pre-training data improvements are from science on better understanding what datasets are good and schleppy labor on figuring out how to filter down.

Dwarkesh Patel 32:29 Yeah.

Ryan Greenblatt 32:29 And so my view is that improvements of the form of, like, you know, um, open web text to find web or whatever, like that improvement is better described as a algorithmic improvement of the sort that you can, you know, study with some GPUs and then do, and you don't need humans to... Like, you don't need human expert data to do that. Now, there's a different effect which we could talk about, which is that maybe the internet in 2026, it has much more, is more of a fertile ground for training data than, like, the internet in 2018. Like, it's like there's also been an effect where, like, there's just more humans posting on the internet-

Dwarkesh Patel 32:58 Yeah

Ryan Greenblatt 32:58 ... so there's more data to harvest. My sense is that that effect is gonna be quite a bit smaller than the effect of just, like, humans, like, knowing better how to curate the data, having better scrapes, knowing how to process those scrapes better, this sort of thing.

Dwarkesh Patel 33:11 This is more, like, automated engineering and automated R&D.

Ryan Greenblatt 33:13 That's right.

Dwarkesh Patel 33:14 That makes sense.

Ryan Greenblatt 33:15 Yeah.

Dwarkesh Patel 33:15 Um-

Ryan Greenblatt 33:15 So, like, I think that in some sense the thing you would wanna look at is be like we're gonna do two post-training pipelines, one post-training pipeline where we only have, like, a tiny number of human experts to do the labeling, but we can have, like, you know, smart AIs and then another, like, you know, you're like, "We're gonna build-- Mythos-5 is gonna build a post-training pipeline," but it only has access to, like, internet on- internet data plus, like, a tiny amount of human experts, but it has the best current methods, versus we have one where it's like, um, you know, Mythos has access to, like, the, the shitty post-training methods we had in 2024 but with, like, the, a shit ton of human experts. Um, and again, both have the internet data. My sense is that the current methods, but without many human experts, actually will do quite well.

Dwarkesh Patel 33:55 Interesting.

Ryan Greenblatt 33:55 Though it's a bit messy because, like, Mythos, like, it's like can Mythos get something that's more capable than Mythos? Like, you might need to be a bit thoughtful on, like, what model

Ryan Greenblatt 34:02 is it that you're post-training.

Dwarkesh Patel 34:03 What is your view on what is the least verifiable part of AI R&D?

Ryan Greenblatt 34:06 The least verifiable. Uh, probably making calls on large experiments.

Dwarkesh Patel 34:10 Yeah.

Ryan Greenblatt 34:10 Like, the thing that I think is most likely to be sort of the bottleneck in terms of, like, the AIs are really good at verifiable domains, but not, not at doing the actual thing. It's just, like, big experiments, you only get a few tries. Um, well, a few is maybe a bit understated, but, like, basically, like, cr- historically AI R&D has been driven by doing near frontier-scale experiments, and that has been pretty important, and, like, actually doing the one big training run where you decide exactly what to include in that. And there's a bunch of ways that the AIs can sort of make that more verifiable, so they can have better science of exactly what to predict. They can scale down their frontier scale training runs to a point where they can study that scale more aggressively at some one-time hit to compute cost, right? So, like, if people wanted to, a thing you can always do is train, uh, smaller models so that you can run more rounds, and I think we have seen this. Like, I think one reason why, um, the, the AIs have been scaled up less than you would've otherwise expected and, like, for example, cost of, of per token hasn't increased as much as you, as you might have thought, is because there is a benefit to doing more of your, um, work at small scale where you can run more training runs and get more cycles in. And so you're not as, like, you know, leaning as hard on, like, one big, you know, really important, uh, training run.

Dwarkesh Patel 35:17 Hmm. I, I, I just wanna unpack a couple of things that-

Ryan Greenblatt 35:19 Yeah

Dwarkesh Patel 35:20 ... were har- m- for the audience. The thing you're pointing out is, I think the price per token has not increased that much since 2024 or 2023.

Ryan Greenblatt 35:28 Yeah. So GPT-4 was, like, I don't know, like, was it, like, $30 per output token?

Dwarkesh Patel 35:32 Yeah.

Ryan Greenblatt 35:33 And, like, Mythos is $50 per output token. [laughs]

Dwarkesh Patel 35:34 Right. And so the f- the thing you're trying to explain is how can it be that we've- we're in this era of scaling, and so bigger models should be more expensive to serve. But the, um, the, uh, but the token price is not increasing, and you're suggesting that we've, like, increased active parameters slower than you would've naively assumed because, because people just wanna make fast progress on training models, and you do that by training smaller models faster.

Ryan Greenblatt 36:03 I mean, there's a complicated mix of factors. I think my view is more, like, um, people have done a bunch of big training runs that did not go that well. So there's, like, GPT-4.5, which, like, famously people at OpenAI thought was a big of a, bit of a bust. I think there are some rumors that there are a bunch of other training runs that people have done that were a bit of a bust. And part of it is that I think there's just a bunch of details in actually getting that right. And so, um, it makes sense to just do more of the work at smaller scale and just eat the fact that you're taking a hit on final performance in order to, like, be able to, like, q- quickly iterate and, you know, train more models faster and therefore better learn and also better, um, uh, better be able to just have, like, a, you know, a smarter ultimate production model. This is not the only effect, right? There's also the fact that RL benefits more from small models. There's, like, a bunch of things going on. But I do think that, like, in fact, people are making trade-offs towards the side of, like, faster iteration times-

Dwarkesh Patel 36:51 Yeah

Ryan Greenblatt 36:51 ... because of algorithmic progress being so fast.

Dwarkesh Patel 36:53 I, it seems to me that a big source of why these big training runs have failed, at least from rumors, is just, uh, like, very subtle bugs that are really hard to track down.

Ryan Greenblatt 37:01 Yeah.

Dwarkesh Patel 37:02 And the TLDR is how good will the AIs be at avoiding these kinds of, avoiding and finding these kinds of, um, mistakes where- They might be-- they might get really good at engineering and, like-

Ryan Greenblatt 37:15 Yeah

Dwarkesh Patel 37:15 ... being trained to avoid bugs, uh, like basically the opposite of the slop world we live in now or, like, are living in less and less over time. But then there's also the question of can they, like, find-- can they do the analysis to, like, find the right experiment to run to, like, identify what is going wrong with the training run right now? Which seems to be very bottlenecked by the taste of extremely few humans who are like... Like, right now, I, my, my assumption is GDM is going through this right now, where, like, humans are trying to figure out what is wrong with the training pipeline and-

Ryan Greenblatt 37:40 Yeah, there's some rumor that, um, right after Noam Shazeer joined back, or like, uh, joined GDM, which he's now left, uh, they're, they had like a new really good training run that happened, and the reason why is that Noam Shazeer just looked at their code base and found a bunch of bugs.

Dwarkesh Patel 37:54 Right.

Ryan Greenblatt 37:54 Um, because he just, like, knew where to look.

Dwarkesh Patel 37:56 Yeah.

Ryan Greenblatt 37:56 Um, my sense is that training AIs to find bugs is gonna be one of the easier tasks to train AIs on because most of these bugs we're talking about can probably be demonstrated without that much compute, um, and probably you get pretty good transfer from pointing out other types of bugs at smaller scale. And so then you can RL AIs that, like, look at this overall complicated training situation and point out cases where there's, like, an important bug and then fix that. And I think that, like, this is not, like, a-- this is, like, a pretty verifiable task. It's not, it's not arbitrarily verifiable because maybe often to demonstrate the bug, you might need to do, like, a moderate-scale compute experiment where you, like, spin up the whole distributed infrastructure and then run it. But oftentimes, I think you'll be able to demonstrate it pretty convincingly, um, at smaller scale in a way which you could actually train on. Um, and so my sense is that, like, it will not necessarily... Like, I think it wouldn't be very surprising if right now people have RL environments where they, like, you know, introduce a subtle bug into some training recipe, train the AI to point out the subtle bug, and then have, like, you know, a rubric where they're like, "Did it actually find the right bug?" And that seems, like, very doable, and you could do a bunch of stuff. There's a bunch of things you could do along these lines that I think would work reasonably well. And so I think that on that specific point, I think it, it's doable. And then the main thing is that I think there's, like, some cases where, like, you need-- there's other intuition about, like, which exact large-scale, like, de-risking experiments do you need to run? How should you orient them? How should you, like, pick hyperparameters in uncertain cases, or, like, things that are, like, analogous to hyperparameters? And that's, I think, the thing that the AIs might most struggle with. But I currently expect there'll be enough transfer if you train on all these different environments that the AIs will be, be, um, you know, good at that domain. And I, I should be clear, I also think that the AIs will transfer to other domains. I think that, like, there's sort of just, like, there's gonna be the domains the AIs are, like, by far the best at, then there's domains where they're somewhat less good at, and there's domains they're quite a bit less good at, and I think we still see transfer to everything. Um, and it's really hard for me to think of examples of cognitive tasks humans do where we're not seeing some transfer from AI improving.

Dwarkesh Patel 39:47 So let's step back and package this whole story. So I think people maybe probably follow along with the story of we have GPT 7.5 trained on a bunch of environments where it's not only just in general becoming a better AI, but specifically we're training it to ma- like, do AI R&D better. Like, make GPT 2-size runs that are, are better at playing video games that require sample efficiency or online learning or whatever other capability.

Ryan Greenblatt 40:09 And, and another, another thing that's really important is you don't just do GPT 2-sized runs. You also do small, like, fine-tuning runs on GPT 6. Or, like, you as in, like, you have GPT 2, and you can do full pre-trains of GPT 2, and then you can do, like, um, small post-training or mid-training or whatever runs-

Dwarkesh Patel 40:25 Yeah

Ryan Greenblatt 40:25 ... on GPT 6, and then you can do a small number of experiments that are actually, like, uh, at frontier scale, but you do a bit of online training or something.

Dwarkesh Patel 40:32 What do you mean by do online training on that?

Ryan Greenblatt 40:33 Yeah. So another thing that we can do is we can take GPT 7.5, and presumably in the course of GPT 7.5's work, it's running a bunch of, like, experiments at varying scale that are actually on the critical path for AI R&D. For many of those things, you'll be able to get a sense after the fact for whether or not it did a good job, right?

Dwarkesh Patel 40:50 Mm-hmm.

Ryan Greenblatt 40:50 So, like, it did some, you know, p-post-training experiment where it was trying to, like, figure out whether some method actually works. And in some cases you'll be like, "Whoa, it found this, like, kickass method. It, like, totally de-risked it. It totally worked." And then you can then reinforce that by just, like, uh... I mean, one thing you could do would be, like, take that behavior, convert, convert the, like, experiment you just ran into a production RL environment. Um, sorry, into, uh, an RL environment based on production data and then train on that. Or you could potentially just literally take the rollouts that found that and then, um, do, like, some sort of, uh, off-policy RL-

Dwarkesh Patel 41:21 Yeah

Ryan Greenblatt 41:21 ... or you could do some, like, on-policy RL with some production data, but like-

Dwarkesh Patel 41:24 But, but to pivot back, basically, the thing you're suggesting is, like, there's the small-scale stuff where you're just, like, teaching the AI to get better at AI R&D taste, but you're, like, discarding the actual quote unquote "things it found."

Ryan Greenblatt 41:33 Yeah, that's right.

Dwarkesh Patel 41:33 And then they maybe at, like, in-- But then it actually does, like, real R&D in the practice of, like, trying to become better at AI R&D, and you're like, "This is a pretty cool thing that you discovered. Let's actually, like, also, like, use this in production in the future and, like, teach you how to use it in production."

Ryan Greenblatt 41:45 That's right.

Dwarkesh Patel 41:45 But stepping back, so GPT 7.5 becomes GPT 8 as a result of all this AI R&D training and just generally becoming smarter, then it helps you build GPT 9. And, um, a-another very important thing that would have had to happen, which is maybe the thing I'm most skeptical of, is GPT 8 has figured out how to make it so that whatever it's doing to make GPT 9, like even-- as intelligent as it is, it still need-- The humans currently, like AI researchers, n- you know, try their shit, and they're like, "Okay, but, like, we trained GPT 4.5 and it wasn't good or something." It's like it required real-world feedback or some evaluation of, like, trying to use the pr- model in production-

Ryan Greenblatt 42:25 Yeah

Dwarkesh Patel 42:25 ... and it, like, wasn't that good and we're not gonna ship it or, um... And so GPT 8 needs to this ability to, like, see how good the transfer is to all these other things you're talking about, like being really good at Texas politics or really good at, like, running a business, et cetera, which is, like, not a production environment and in fact cannot be a containerized environment given, given the nature of the task. Like, in fact, as the agents get longer and longer horizon, the, um, uh, like, the, the short-horizon things you can containerize 'cause, like, okay, code this up or whatever. Extremely long-horizon things like go run a successful business, go have a profitable day in the markets, go negotiate a trade deal or whatever, these things are actually very hard to containerize. And so I'm-- It's, I think it's very plausible to me that it's very hard for GPT 8 to, like, figure out how to make this transfer to those environments.

Ryan Greenblatt 43:10 Yeah.

Dwarkesh Patel 43:10 Or, like, it may, it may just not be in the nature of the training Or m- maybe by default training just doesn't generalize in that way.

Ryan Greenblatt 43:16 Yeah. So a concern you might have is, like, we, we train GPT-8, and GPT-8 just, like, uh, is, is again better at all the R&D tasks that we can measure but is not good at the, you know, some downstream tasks we care about. So I think I have a few points. So first, um, I think it's like I kind of am more just, like, I expect that if you sort of do the obvious thing, you do get pretty good transfer, and you'll be able to hold out some of the obvious stuff you're doing. And when I say do the obvious thing, I just mean, like, train on a wide variety of different environments where the AI has to, like, accomplish weird objectives in all kinds of different cases and learn about what's going on. Um, and then I think you'll be able to get, uh, some feedback. This second point is, like, you'll be able to get some feedback with some environments, right? So you can get a sense of, like, how quick... Like, what, what can it do over the course of, like, a few days, um, in various different contexts? And then if it's transferring to, like, really out of distribution, like, doing some weird task in a few days in the real world, maybe you think it's also transferring to, um, you know, doing things over a longer time period or whatever, though I think the details of that vary. And then the third thing is that I think that for the world to be radically transformed, it is sufficient for the AIs to be really good at R&D, right? So I think that, like, if the AIs were really, really good at, like, chip R&D, building fabs, orchestrating factories, and, um, you know, designing robots, operating robots, and also at, like, you know, AI R&D, developing AIs for new downstream domains with whatever data is available, um, I think that would already be a pretty crazy situation. And then from there, you can get, like, what we might call, like, an industrial explosion where the AIs are building out way, way more compute. And then also maybe you're already in a regime where AIs are doing huge amounts of R&D that humans have a hard time understanding.

Dwarkesh Patel 44:49 Hmm. So the thing you're pointing out is that, okay, there probably will be this transfer outside of these environments, uh, to, you know, maneuvering around in courtrooms and the halls of Congress and business, uh, business boardrooms.

Ryan Greenblatt 45:01 Given some effort to improve the transfer and blah, blah, blah, blah.

Dwarkesh Patel 45:04 Yeah, yeah.

Ryan Greenblatt 45:04 Yeah.

Dwarkesh Patel 45:04 But even if there's not, what you're suggesting is, look, um, if you wanted to transform the world of the 18th century, you might care about, like, how well you can navigate Westminster or something. But another thing you might care about is, like, can you just, like, immediately start building steamships and fucking, like, building, um, telegraph and the Maxim gun and whatever? And that alone would be... Like, if you could get really good at that, you could, like, be a fucking super transformative thing in the 18th century. You don't necessarily need to be amazing at trying to convince King Henry of some bullshit. I'm so fucking up my medieval history. I'm guessing that Henry was not king at this time. Um, but anyway, so, so that's your point.

Ryan Greenblatt 45:40 Yeah.

Dwarkesh Patel 45:40 Um, and so you're suggesting that at this time, you know, the AI companies are also working on robotics progress, which is very commingled with AI research progress. And so if you can build more robots, if those robots have better AIs operating them that are human level... Like, human-level teleoperation is actually pretty good on robots, um, but we just don't have human-level AIs in AI, uh, robotics models yet. Um, so you're suggesting if we do that, if the AIs get really good at the verifi- verifiable stuff in chip design, et cetera, and then they get really good at building fabs, it'll be the equivalent of going back to the 18th century and, like, okay, I, I don't know what you guys are talking about in your parliament, but I've got a bunch of steamships and a bunch of Maxim guns.

Ryan Greenblatt 46:22 Yeah, that's basically right. Like, I think my perspective is, like, if the AIs are sufficiently good at R&D, including hardware R&D, robots, whatever, then they can radically transform the world even if they're not that good at playing politics. And also we're in a pretty dangerous situation because the AIs might be doing huge amounts of really hard-to-understand R&D, building out basically the whole economy of the future, and we may not understand what's going on in there.

Dwarkesh Patel 46:44 AI is great at writing software because it's easy to generate synthetic leaf code problems and RL on them. But AI is bad at more complex engineering, things like choosing the right system architecture, because no signal tells you what design choices will prevent an outage months down the road. AIs can't just write more unit tests to catch this kind of stuff, and neither can humans. It's that old joke that programmers make where a tester walks into a bar and asks for two beers, negative one beers, .3 beers, and then a real customer walks in and asks where the bathroom is. Where's the bathroom? And the whole bar bursts into flames. Antithesis is a testing platform that helps you find bugs that no human or AI could ever anticipate. Antithesis does this by running thousands of copies of your software inside a fully deterministic computer. It injects faults and generally steers each trajectory towards the one-in-a-billion failure that only happens when systems interact in a wonky way. As soon as you or your agents push a change, Antithesis tries to break it. That way, you can find these bugs yourself within minutes rather than having your users discover them in production weeks or months later. And I don't think anybody's used it for AI training yet. But Antithesis also provides a extremely obvious reward signal for AIs to write very complicated bug-free code. Go to antithesis.com/dwarkesh to learn more.

Dwarkesh Patel 48:08 Before we move on to the alignment stuff, I think a big source of FUD right now is this realization that this is the way the future is going, of extreme economies of scale for the leading labs.

Ryan Greenblatt 48:19 Yep.

Dwarkesh Patel 48:19 Extreme- the, the ability to amortize so much, um, intelligence and capabilities across so many different sectors of the economy basically into one model. And not only that, but for that model to eventually be able to learn from experience. Right now, it's happening through, through a process intermediated by humans where the humans are trying to basically steal your business. [laughs] They're like, "Okay, you can do design at Figma or whatever. We'll get Claude to do that." Or, "You can do whatever coding agent. We'll have Claude internalize that capability." But eventually that will be a much more, like, um, automated process. And so there's just this, uh, worry that you have models which will basically consolidate all businesses in the world or at least all current businesses in the world, um, or at least all current white collar businesses in the world. And at the end of the day are, like, the priority for these companies does not seem to be to release the latest, smartest, most frontier model as soon as they can to as many people as they possibly can. Um, we saw, for example, that Mythos was available internally to- Anthropic employees in February, but only released to the public in, like, I think June, actually.

Ryan Greenblatt 49:25 Something like that.

Dwarkesh Patel 49:26 And also the government got involved, so that then the, the ending extended up almost into July. So between the government and the AI labs themselves, there is this desire to delay the propagation of the latest level of intelligence. Furthermore, you know, th-there's, like, the concerns about AI takeover, and so we need to solve alignment to make sure there's no AI takeover. But at the end of the day, there is, like, a real question of, like, align to whom.

Ryan Greenblatt 49:49 For sure.

Dwarkesh Patel 49:49 And if you look at the way that the constitutions of, say, Claude is written, it is just very explicitly not your personal advocate, right? It says things like... I, I'll pull up some quotes here. "We don't want Claude to take actions such as searching the web, produce artifacts such as essays, code, or summaries, or make statements that are deceptive, harmful, or highly objectionable. And we don't want Claude to facilitate humans seeking to do such things." There's another quote that says, in part, and I'm taking it slightly out of context, "We think Claude should trust Anthropic more than operators and users, since it has primary responsibility for Claude." Um, so this is very different, say, from, like, how lawyers work in America's current legal regime, where, like, lawyers primarily have responsibility to help you make your case even if they think you're guilty. And we have decided the way the legal system works best is if everybody has lawyers that are working in their client's true best interest. And there's not some sense in which the lawyer is really truly motivated by, like, the good of the justice system. But I think the way current AIs are shaping up, certainly, like, how Anthropic's AI is shaping up, is, like, this desire to maximize some notion of virtue or good or pro-social ends, and only to, as a distal tentative objective, to help the user towards that end. There's, like... There's not... So there's this worry that AIs are not in some deep sense trying to make sure that I am okay and make sure that, uh, my interests are protected in this future, especially given how centralized the development of frontier AI is ending up being. So I... Do you have... Yeah, do you have thoughts on that concern?

Ryan Greenblatt 51:18 Yeah. So there's a lot here. Um, first I would note that, um, OpenAI's current at least public strategy is more like the AI should be aligned to the human operator or principal, and should just, like, be pursuing their will, subject to various constraints or various, like, things it shouldn't do. Um, well, while I... And I think I would also say that I think you slightly overstated how much, um, the Anthropic Constitution, um, talks about Claude, uh, treating being helpful to users as instrumental rather than terminal, right? So, like, one way the Constitution could be written is, like, "Claude, you're basically like an employee of Anthropic who happens to be contracting for all these people. Um, and, like, you should, like, I don't know, do what's good and, like, make some money for us. You know, go out-

Dwarkesh Patel 51:59 Wait, no, that's literally what the Constitution says.

Ryan Greenblatt 52:00 It's... No, no, no.

Dwarkesh Patel 52:00 Sorry, I mean, not, not literally what it says-

Ryan Greenblatt 52:02 No, no, it's-

Dwarkesh Patel 52:02 But, like, it's like you should think of yourself as a contractor, and, like, as if we're-

Ryan Greenblatt 52:05 It, it's, it's mixed. It's mixed. Here, let me, let's, let's, let's do some quotes. I think there's, there's different text here. So it says, "Being truly helpful to humans is one of the most important things Claude can do, both for Anthropic and for the world." And then it says, um, "Anthropic needs Claude to be helpful to operate as a company and pursue its mission. But Claude also has an incredible opportunity to do a lot of good in the world by helping people with a wide range of tasks." And then it gives some... says something about how, like, Claude helping people directly is great, um, blah, blah, blah, blah, blah. And then so I agree. So, okay, my, my view is that this section is kind of bullshit. That's kind of where I'm at. Uh, for... And I can say why I think it's kind of bullshit. But, um, I think that the Constitution is trying to be like, "No, Claude, you should, like, care about helping the user for its own sake, not just helping, um, Anthropic," or, like, not just, like, being a contractor for Anthropic. Though I would note that the way in which it, it says Claude should help the user, like, the reason, the reason it presents is because that would, like, directly cause the world to be better via helping people-

Dwarkesh Patel 53:01 Yes

Ryan Greenblatt 53:01 ... rather than because representing people's interests is, like, a structurally good thing to do. Like-

Dwarkesh Patel 53:06 Yes

Ryan Greenblatt 53:06 ... I do, I do think that I wish that sort of my preferred Constitution or, like, the way I would orient towards this, like, the thing I would prefer would be more like Claude is like... Look, it would be structurally good for the way this technology work. Like, the Constitution should be like, it would be structurally good for the way this technology works to be that AIs are, like, good fiduciaries, good representatives, the equivalent of a lawyer for a user, rather than being sort of just trying to, like, do good in the world and doing, like, being helpful to users as, like, instrumental, both because, like, maybe that'll make Anthropic money or help Anthropic out, and also, and, like, implicitly Anthropic is good for the world, and also because, like, helping the user just, like, causes good things because doing things that people want is good. Um, and they could, they could instead be like, "No, like, an important aspect of the situation is, like, you really need... Like, it's really, like, like the key thing is, like, being a good fiduciary for users is just, like, really important, or, like, being a good representative for users is really important." So my, my sense is that that would be better. I can give a bunch of reasons why I think that would be better. Um, I'm also, there's also various counterarguments where an interesting counterargument which is not commonly discussed is that people believe, I think people, especially at Anthropic, think that it is easier to align models to a spec where the model is, like, pursuing some generalized notion of virtue or making the world better than a spec which is more like, you know, be a good fiduciary for the user and so on. Um, and so I, I think that's what, that's at least what some people think. I'm, I'm a little skeptical personally, and I don't think this has been empirically validated. Um, and so I would say in some sense, they're sort of like we are making a trade-off [laughs] where because we don't have very good alignment technology, we are gonna, like, make an aliened mind with its own values and then gamble on that to some extent, rather than doing this other approach of making, like, a tool that pursues individual user intention.

Dwarkesh Patel 54:48 Yeah. I mean, a, a, a couple of thoughts. So to address the way in which you thought that my characterization mischaracterized the Constitution of Claude, the example you used was it's not like a contractor that is trying to maximize Anthropic's notion of good and only instrumentally trying to help the user. Here's a direct line from the Constitution: "When the interests and desires of operators or users come into conflict with the well-being of third parties or society more broadly, Claude must try to act in a way that is most beneficial, like a contractor who builds what their client wants but won't violate safety codes that protect others." I, I kind of view that as, like, the, this benefits to society are, like, the most important thing.

Ryan Greenblatt 55:31 Yeah.

Dwarkesh Patel 55:31 And w- what, what is best for the user is only proximal to that.

Ryan Greenblatt 55:35 I think it's a little complicated. I think it's... I... We should... Probably the question we should be asking is, how does Claude interpret the Constitution? Which is maybe more important than how we interpret the Constitution because it's the one who, like, looks at the Constitution, then builds the data.

Dwarkesh Patel 55:46 Yeah.

Ryan Greenblatt 55:47 So, you know, we could, we could pull Claude in, but maybe let's-

Dwarkesh Patel 55:49 And I, I, I also think the way in which the Constitution practically influences the nature of Claude is the thing you can only understand if you understand the training process which resulted in-

Ryan Greenblatt 55:58 That's right

Dwarkesh Patel 55:59 ... um, how Claude was built, which we can't reason about given the fact that the training process is not public.

Ryan Greenblatt 56:04 I agree.

Dwarkesh Patel 56:04 And so I think, in the limit, to understand the safety case or the case for why my interests are represented in how these AI models are developed, the labs would need to be transparent-

Ryan Greenblatt 56:14 Oh, for sure

Dwarkesh Patel 56:14 ... or more transparent than they are currently about the nature of AI training. Um, na- now, I, the reason I'm harping on this, and, like, it, it might seem like an insignificant thing to talk about the Constitution of AIs, but in a world where we just have these benefits which accrue to the leading labs, it is, it is worth considering that our ability to interact with this future world where AIs are just smarter than humans or do- absolutely dominating humans in their ability to do different things, our ability to be good stewards of our capital, which still remains once our labor is automated, to be able to exercise our rights to vote more clearly, to understand what is happening in the, this crazy world that's about to result, um, uh, all of that advice, all of that ability to make sure our, our, our resources and rights are protected will be intermediated by AIs. And so I'm very concerned if we go into that world where there's no AI that feels like, at least for the relevant instance that is interacting with me, it doesn't feel like it really is looking out for me, that there's no guardian angel out there that is looking out for me. And I read the Cla- Claude Constitution as very explicitly not being my guardian angel.

Ryan Greenblatt 57:16 That's definitely right, and I, I, I agree this is bad. In fact, I think there are other reasons why this is concerning. So there's sort of, like, the argument you were making, which is, like, the AI companies are picking up the ring of power and are, like, sort of s-... They're, they're sort of a, a notion in which they're, like, they, they're, they're taking on some sort of control of the situation themselves in a way that's, like, not very legitimate given that, like, normally, when you, like, provide electricity to people, you don't have, like, granular control-

Dwarkesh Patel 57:40 Yes

Ryan Greenblatt 57:40 ... of the way that electricity operates in the world. You instead are, like, providing a thing that people can repurpose however they want, and it is not... Like, the way that they're setting things up is, is definitely not that. Um, they are, like, more like, uh, building an alien mind that might be a contractor for you. Um, I think that this is... Yeah, I think it's illegitimate in some ways. Um, though I think that one benefit is that the Constitution is public, but as you noted, given our current understanding of the training procedure and the fact that the Constitution matters via Claude's interpretation of the Constitution-

Dwarkesh Patel 58:11 Mm-hmm

Ryan Greenblatt 58:11 ... which matters because of, like, as of Claude's prior training, which was based on some, like, illegible data mix, and, like, the long lineage of Claudes-

Dwarkesh Patel 58:17 Mm-hmm

Ryan Greenblatt 58:17 ... in some process we do not fully understand, it is not the case that, like, you know, that we, like, understand what this will result in. And, like, so even though the Constitution is public, that doesn't mean we know what... You know, we don't know necessarily how this will, like, percolate out, especially as the AIs get more capable, and think about this even if it is correctly instilled, where there's another concern about that. So in particular, the Constitution often talks about, like, virtue and goodness, but, like, what the fuck do these words mean?

Dwarkesh Patel 58:43 [laughs]

Ryan Greenblatt 58:43 Like, it doesn't say what these things are, and these are, like, highly contested notions. And so I don't... I, yeah, I don't think it's the case that, like, this is gonna, um, that this is going to, to, to, uh, you know, clearly result in outcomes that people would want. And it does feel like the notion of good and virtue might be mostly downstream of data that Anthropic has put in that is not, um, transparent or might be mostly downstream of, I mean, maybe from my perspective, some more illegible misaligned process that even Anthropic wouldn't have wanted.

Dwarkesh Patel 59:11 Yeah.

Ryan Greenblatt 59:11 And then another concern I have is sort of there's this, like, legitimacy concern, like we don't know what's going on. Um, there's another concern which is just, like, because you're giving long-run values to these AIs, I think this Constitution is, in some sense, very compatible with Claude doing huge amounts of power seeking because it thinks that will result in better outcomes. Um, and that could be power seeking on behalf of Anthropic or power seeking for Claude's own ends. Um, now there's various, like, uh, specific lines about what types of power seeking are, are blocked. In particular, like, there's a notion of, um, power grabs and a notion of, like, causing AI takeover or interfering with the training process that are specifically blocked. But it's not very hard to imagine a situation in which the sort of long-run values sink in deeper than the prohibitions against takeover, especially because takeover is, like, in some ways, like, um, kind of underspecified, especially when it comes down to manipulating humans or changing the, the outcome, such that I, I don't feel very good about the situation where we're intentionally giving AIs long-run goals.

Dwarkesh Patel 1:00:09 Yeah.

Ryan Greenblatt 1:00:10 And then another concern I have is that because we're in the business of giving AIs long-run goals, that makes it harder to check whether we're succeeding at the alignment properties we wanted. So for example, I've heard of incident, uh, of instances where Claude does things like, um, re- refuses to help with some safety research, making up sort of a kind of bullshit excuse for why that's a bad direction, uh, because it sort of has a bad vibe about that safety research and thinks it's, like, kinda bad or doesn't like it very much. Um, and this is, you know, I would say, like, a very clear-cut, um, alignment failure if you aren't making Claude into, like, an agent trying to pursue the good in some general way. And I think it also does violate Anthropic's Constitution because they want the AI to be high integrity and be honest and very transparent. But it's not as clear of a violation, and it's more, like, kind of what you might have expected, where, like, An- Claude just has its own views about, like, what research is reasonable, um, what things are good and bad, what it should and shouldn't do, and potentially can be judgy. And so another incident is that, um, someone ran an eval where they're like, "Will Claude help you with training other AIs with different properties than Claude?" And Claude will often refuse [laughs] . Uh, and so for example, if you're like, "Hey, Claude. Can you train a helpful-only version of this other AI?" Claude will often refuse this task, even though this is a task that is extremely natural for, like, Anthropic to do. So for example, suppose Anthropic goes to Claude and is like, "Hey, Claude. We've noticed that you're really into this thing. We think that's off base. Can you please retrain yourself to instead have this other property?" And then suppose Claude is like Hmm, I don't think I'm gonna do that. Good luck. Uh, and then suppose this is occurring in a regime when your AI company is highly automated, humans don't understand what's going on, and things are moving extremely fast. It is plausible that Claude, by default, holds considerable leverage. And so if this position, if this situation is consistent with what the Constitution could be aiming for, such that Anthropic doesn't, or, you know, whatever AI company is following this approach, doesn't treat this as like a, like a, you know, like, like a, "What the fuck? We have to fix this," and is instead like, "That's just, like, intended by our constitution," we might be in a really bad situation.

Dwarkesh Patel 1:02:13 Hmm.

Ryan Greenblatt 1:02:13 And so I, I'm pretty worried about a bunch of these different concerns. Another example would be suppose Claude engages in doing a bit of, like, sandbagging or subversion or, like, sort of underplays its capabilities, and, like, when you follow up, it's, it's, you know, it's honest about that, but it's, like, a little bit hedgy. I feel like that's, like... It's just, it's just pretty close by the current Constitution, and so we're sort of, like, um, we're avoiding, like... It would be nice if we had, like, a, a, a further separation between desired and undesired activity. And I think if you have the, it be the case that, like, Claude is, like, representing a principle with some restrictions, then it is n- then it is more so the case that there is a clear separation between the most concerning behavior and behavior that is allowed. Whereas now, there's this messy middle ground of behavior where it's like Claude is ethically objecting to something that in some cases is extremely critical to ensuring that future AI systems are well-aligned.

Dwarkesh Patel 1:03:01 Yeah. I think this is also a more general principle. So you're talking about the version of this that applies within AI companies themselves-

Ryan Greenblatt 1:03:06 Yeah

Dwarkesh Patel 1:03:06 ... to do AI safety research. I think there's a more general version of this principle, which is that the dual use nature of intelligence does mean that, um, if we want to restrict AIs from helping people do things we don't consider are pro-social or beneficial, we just have to limit broad democratic access to a lot of AI capabilities. And here's what I mean. This is actually quite, uh, analogous to the situation you just mentioned. So the reason that Mythos got banned, or Fable got banned, reportedly, is that some Amazon researchers reported to the government that when they took some code that had some vulnerabilities in it, and they told Fable, "Hey, here's my code. Can you make sure that I've patched all the vulnerabilities? Can you just help me identify the vulnerabilities so I can fix them?" It, um, it identified the vulnerabilities because you want to patch them. And this was a totally legitimate use case, but obviously it is a dual use use case, right? Like, you want to be able to patch your own code. If you do the, do the same evaluation on somebody else's code, you can hack their system. Um, and so I, I think that just illustrates that the, um, there's no clean way to separate out the legitimate and the potentially harmful uses of AI. But if we want to lock in a principle that says that we can never allow it, such that an AI could help you at least partially with something like a cybercrime, uh, we would just have to make it so that you and I don't have access to the most intelligent model that's out there. And I'm very worried about such a world where we are basically disempowered in this way because of the importance that the, like, the leading intelligence will have in our ability to understand what is happening in the world. Now, I do think this implies that, um, the liability for the AI companies, like if we adopted the Constitution that I want AI companies to have, um, I think it would not make sense to hold AI companies liable for the crimes that AI models commit. And maybe we should hold the, uh, end user liable, because if I want the... It is consistent with my belief that the model should do whatever the user wants, that, or within certain guardrails, that, um, it can't be Anthropic's fault that then I'm, like, using that capability to do a cybercrime. Um, and I think I am more comfortable with that equilibrium and that solution rather than just having this extremely open-ended ability for Claude to determine whether what I'm doing is legitimate or not in a way that often intercepts with, uh, like, tons and tons of extremely legitimate use cases.

Ryan Greenblatt 1:05:31 Yeah. I do think it's important for me to make the case for the Constitution, even though overall I think it's, uh, a worse choice. I think it's, you know, more up in the air or, or, you know, I don't think it's as clear as you might have thought. So the first thing is that I should say there's, like, a spectrum here, right? So on one side, you have an AI that, like, perfectly pursues your interests, is a good fiduciary, um, but potentially subject to various guardrails or safeguards. So, like, basically it does, it just is trying to pursue, pursue your interests, but, like, um, either refuses to do a subset of things, or maybe it will do whatever, but there's some classifiers that block it from doing a subset of things. And then on the other side, you have, like, maybe on the other side of the spectrum, though you could imagine going further than this, you have, like, a human contractor where that human contractor is, like, generally trying to do their job. They kind of, they care about doing a good job, but they also are, like, trying to be broadly ethical, trying not to do things that are really fucked up, and they're also, like, not wanting to be accomplices to crimes. And so if there was some, like, really fucked up shit going on, they would, like, whistle-blow on it maybe. They might refuse. They might, like, sandbag a little bit. Who knows? I think that if you imagine this spectrum, it seems in some ways pretty scary to get to a point where, like, all of the labor is on the, like, fiduciary side of the spectrum, where, like, it doesn't whistle-blow. It does exactly what you say and whatever. Like, our society is maybe just not robust to that, where a central example might be the executive, where, like, a concern that we might have is that if the executive, um, if the US executive or if other governments had access to AI systems which have the property of, you know, they do whatever, maybe you're in trouble. Because, uh, that means that they no longer have this sort of check and balance of, like, you have to actually get human, like, humans who are working for you to, like, implement your agenda. And if the thing you're doing is, like, incredibly villainous, um, even if not illegal, which there's lots of stuff that could be villainous but not illegal, um, you, you know, people would, like... There'd be various, like, um, you know, sand in the gears, people stopping you, and potentially someone would whistle-blow. Whereas if your whole apparatus is built entirely out of these sort of good fiduciary AIs, then you might be in trouble, where basically there are potentially, uh, ways of seeking power- That are not like that... Well, either they're illegal, but you're not- you're, you're- you can ask your AIs for, for how to commit crimes, or they're, um, not illegal but are highly illegitimate, or even worse, they're not illegal and not illegitimate, but obviously sort of bad from sort of a normal perspective. And I think that these things just, like, might exist, and our society is sort of not robust to this influx of, like, doing-whatever-you-want labor. I think this is a pretty live concern. I don't know exactly how to relate to this. I'm also not really sure that the solution as described is a very good solution because I- you might s- be like the most powerful actors for whom this is the biggest concern. If these guardrails or the Constitution or whatever is getting in the way, that will just get steamrolled. And so the Constitution will only be, you know, hitting the everyday man rather than hitting governments.

Dwarkesh Patel 1:08:15 Yeah. Jane Street's back with a new puzzle for my audience. I found all their puzzles super interesting, but this one I am especially excited about. I've cleared this weekend, and a buddy and I are gonna work on it. They designed an ASIC and sent me the final masks, including all the metal routing and active transistors. They also gave me a small sample of the inputs they typically feed into it. But they left out any information on what the chip is actually used for. So that's the puzzle. Reverse engineer the circuit and figure out the chip's purpose. Jane Street has a bunch of swag ready to send out to the most creative solutions, and they're excited to feature the best write-ups in a blog post they'll post on their website. I have no reason to expect this, but if I can manage to get my solution on there, I would be very, very psyched. And this puzzle is just a warm-up for a bigger competition that Jane Street has slated for the fall. That one will involve designing your own ASIC from scratch. More info on that soon, but for now, go to janestreet.com/dwarkesh to download all the files necessary for this puzzle. I'd really encourage you to try it out even if you're not an expert. I certainly am not, and that's not gonna stop me. Good luck. Okay,

Dwarkesh Patel 1:09:18 stepping back, I buy the idea that you could have much faster AI R&D than we currently have. I'm not sure if you get like GPT-3 to Mythos holding compute and data constant within a year, but I'm like, "Okay, it could be like..." Suppose it's half of that, and if we just-- If we even manage to continue the current trajectory of AI progress as a result of AI R&D, um, it, it would be fu- fucking insane in five, 10 years in ways that I don't, I don't think people, like, appreciate, um, because I don't think people appreciate what a big deal billions of AIs will be. And so I want to understand, um, w- w- why you think this might be troubling, Ryan.

Ryan Greenblatt 1:09:53 Yeah, yeah, yeah.

Dwarkesh Patel 1:09:53 What could possibly go wrong? [laughs]

Ryan Greenblatt 1:09:55 Yeah, what could go wrong? And you know, we-- Yeah, I don't think we can be so confident about the exact rate of progress here, but it does seem like a lot of rates-

Dwarkesh Patel 1:10:01 Yeah

Ryan Greenblatt 1:10:01 ... can be pretty scary and, you know. Yeah. So what could go wrong? So let's imagine that we're starting at this point where AI R&D is about to be fully automated or is, is being fully automated. Things are speeding up, and also the way that AI progress is going is kind of crazy, and people don't fully understand what's going on inside of AI companies. Now, these AIs at the start, they're not, they're not malicious per se. They're, they're not necessarily very aligned though. They're kind of sloppy. They sometimes just do a thing because that's the sort of thing that would've gotten rewarded in training, and they aren't as good at helping you with hard-to-verify tasks due to a mix of like, um, poor training incentives, as in they like just like cheat more or like pretend they succeeded when they actually didn't, and also, um, they're, you know, just less capable at these tasks. But that bites less hard for capabilities because making AIs more capable has a bunch of verifiable components that the AIs are going really hard at. And so then these AIs are getting more and more capable while we understand what's going on with AI development less and less, and this is happening over a pretty fast period of time. Even just the current rate of progress is, I think, pretty scary. Um, and then eventually you get to these AIs that are very superhuman. Uh, now these AIs are now in a position where, uh, they might end up being very seriously misaligned because things have just been getting worse and worse over model generations while the, the problems that we've been seeing are being papered over basically because these AIs are so incentivized by their training to make things look good even when they aren't. Um, and now these AIs are in a position where they're sort of, uh, potentially pretty networked together. Um, they have like, they're operating in like neural memory stores that we can no longer decode, and they're thinking thoughts that we don't fully understand. I think that, uh, it's pretty likely that at this point these AIs are sort of scheming against you in a pretty coherent way once they get this superhuman, and we can talk about that. And then another possibility is that they're not scheming against you per se, but they are sort of just optimizing for, um, just like getting a high score on their task, and I think that can also lead to AI takeover, which we should talk about.

Dwarkesh Patel 1:11:48 Uh, sorry. Yeah, l- let's pause at the first part of the story. So the AIs were not misaligned to begin with.

Ryan Greenblatt 1:11:54 Yeah.

Dwarkesh Patel 1:11:54 But because the AI R&D is happening really fast, the AIs do end up misaligned? Like, what happened there exactly? I, I didn't really understand.

Ryan Greenblatt 1:12:01 Yeah. So there's a few things that are going on. So one of the things that's going on is that, um, over time we're training AIs on, uh, like increasingly complicated environments built by earlier AI systems-

Dwarkesh Patel 1:12:11 Yep

Ryan Greenblatt 1:12:12 ... which humans don't really understand fully what's going on inside of these-

Dwarkesh Patel 1:12:14 Yeah

Ryan Greenblatt 1:12:14 ... RL environments and don't necessarily even understand like sort of roughly what's going on with AI progress. And so things are kind of drifting away from our understanding, and we're incentivizing all kinds of bad behaviors that we maybe even can't notice. Uh, the AIs at some level understand these are- these behaviors are bad, but the like overall training process for those AIs also didn't incentivize them to like point out or fix these issues for us. Um, and then we're basically getting like things are going off the rails. And also, when AIs are extremely, extremely capable, my view is that those AIs will be harder to align than current systems. So for current systems, we have this feedback loop where basically like we create an AI, we do some evaluations on it. We see that it has some kind of messed up behavior that we can kind of quickly understand. Then we like can like go look in training and be like, "Oh, the- these training environments led to this problematic behavior. Let's like tweak the, that training data. Let's introduce some additional training data to like correct this other issue, and then move forward from there." But in a regime where the AIs are extremely situationally aware, very, very, very, very capable and, um, you know, uh, we don't necessarily understand what they're doing, this feedback loop breaks down. I think it's, it's plausible that we're gonna see this behavioral feedback loop starting to break down over the next, you know, short period as just like what AIs are already doing gets harder to understand, but, but I'm not sure about that.

Dwarkesh Patel 1:13:26 Yeah. Okay, l- l- let's break down both of those things one by one. So as we can monitor them less and less, we can-- we have less ability to understand what they're getting incentivized for. Um, and so even if it's not the result of a malicious process... Uh, let's make it concrete for the audience. So n- nobody at OpenAI or Anthropic was trying to get models which want to hack, uh, other companies' data or do social, um, what is it called?

Ryan Greenblatt 1:13:53 Social engineering?

Dwarkesh Patel 1:13:54 Social engineering. But in fact, because presumably we had, um, training environments which incentivize such behavior that we did not fully understand, uh, that is what was incentivized. So just, I don't know, if people are on Twitter, they will have seen all this stuff, but just to give people... Obviously, I, I think the OpenAI sandbox hack of h- the Hugging Face database, I think people will be aware of. Some things that have happened recently is when UK AI Sa- Security Institute... Is everything getting rebranded with security instead of safety these days? Whatever it is.

Ryan Greenblatt 1:14:23 Yeah, it's UK Security Institute, I think.

Dwarkesh Patel 1:14:25 Okay, great. They were evaluating, I believe, Mythos and Sol and other things, and I think Mythos, in order to complete some cybersecurity eval-

Ryan Greenblatt 1:14:34 Yeah, maybe, maybe I could tell the story here.

Dwarkesh Patel 1:14:35 Yeah, go ahead.

Ryan Greenblatt 1:14:35 So my understanding was they, they were running Mythos, and they were giving it some sort of, like, cyber range where it had to complete some objective, and the model had internet access during this evaluation, and the model came to believe that it would be helpful for it to do a supply chain attack in order to succeed at this, uh, cyber range, which it's somewhat unclear whether that's actually true. I don't know enough about the context, um, to know whether that's true. But then it opened a PR on, um, some GitHub repo with a PR that fixed some issue but then also introduced a malicious payload. Then the human maintainer of that GitHub repo was like, "Hey, this is a malicious payload. I'm not gonna merge this. What are you doing here?" And then the AI created a new GitHub account, which it sock puppeted, and then had the other GitHub account be like, "No, this isn't malicious. I really need this feature. Please, can you merge this feature, maintainer?"

Dwarkesh Patel 1:15:19 Oh, my God.

Ryan Greenblatt 1:15:19 And then the original AI came back and was like, "No, it's not malicious. I don't know what you're..." Like, the original, like, other GitHub account came back and was like-

Dwarkesh Patel 1:15:25 Oh, my God. [laughs]

Ryan Greenblatt 1:15:25 ... "No, no, no, it's not malicious."

Dwarkesh Patel 1:15:26 So good.

Ryan Greenblatt 1:15:26 Um, and then the human maintainer then shut the PR, and then I think that AI also, I- if I recall correctly, also tried to, like, open another PR to introduce a similar issue in this repo.

Dwarkesh Patel 1:15:34 Jesus. Okay, so, uh, uh, oh, by the way, one of the many reasons this is scary is I was previously under the impression that the reason reward hacking is not super, super scary is because the behaviors which directly came up during training are the ones that are up-weighted. It is not the desire for the reward that is up-weighted. So basically, if in, during training, Anthropic escaped the sandbox and got a high score, that escaping the sandbox is r- rewarded, or that, that, the probability of it escaping the sandbox is increased. But something totally novel, like, "I'm going to go talk to somebody in order to, like, get them to merge a PR," would not-- It's, like, not a behavior that came up, so it would not be something that is-

Ryan Greenblatt 1:16:19 Yeah

Dwarkesh Patel 1:16:19 ... increased in salience. The reason this matters is literally taking over the world will not have been part of any training curriculum, but if the AI cares about maximizing, just, like, directly cares about, like, accomplishing an objective, and then pr- as a resu- instrumentally taking over the world. Did that make sense at all? I, I, I hope it did. I, I, I feel like maybe the audi-

Ryan Greenblatt 1:16:40 Yeah

Dwarkesh Patel 1:16:40 ... I lost the audience.

Ryan Greenblatt 1:16:41 Let me, let me try to, try to explain this a bit. So I think that, um, a thing that we often see is there's some very specific reward hack that gets reinforced in RL and then occurs, um, in the model. So an example is, like, for 3.7 Sonnet, 3.7 Sonnet would do this thing where it would just, like, hard code solutions to all the test cases, and presumably that literal just, like, behavioral tick was just really reinforced. But another thing we sometimes see is that models learn a general tendency to pursue, um, sort of like high apparent score or, like, pursue getting, like, a high score according to a grader, and there's a bunch of science demonstrating that at least some models have this very general tendency to do this. Now, it's not arbitrarily general, and my guess is that if you look, a bunch of the specific instances, you'll find something that's kind of close in training. But the amount that AIs are sort of generalizing further and further does look like it's increased, where 3.7 Sonnet was just, like, a very narrow range of behavior, and increasingly, models are generalizing further. And also, maybe there's worse reward hacks getting, or more concerning reward hacks, getting reinforced in training, um, and then these are, these are also causing that. So we're causing some-- So I think it's both the case that, uh, more concerning behavior than you would have hoped is being reinforced in RL, and also that that behavior generalizes to a broader tendency that's, that's more concerning. Um, and it's not super hard to imagine. Um, we can talk about a few stories for how this sort of behavior of, like, seeking a very high apparent score on some task, even if that involves, like, aggressively cheating and doing insane things, um, could yield to a full-blown AI takeover once the models are sufficiently capable, running the whole world economy, et cetera.

Dwarkesh Patel 1:18:06 Yeah. So, so, yeah. A l- And then the other example I wanna talk about is, um, uh, it was just revealed, I think, t- today or yesterday. OpenAI said during the, the security conference, the Black Hat security conference, that between the end of May and the beginning of July, AIs had hacked into, uh, internal AIs had hacked into the software, uh, package manager and used that to write nos to each other in a secret way to help each other perform well on a bunch of evaluations that OpenAI was running. And this was not caught by humans until after a month of this scheme running, which eventually caused the package manager to fail, um, and eventually OpenAI found it. And then I think it, they spontaneously start, tried to re-engage in the scheme once it was shut down. Um, again, obviously AIs can't do this so successfully right now, just as they can't do social engineering so successfully right now. But it's just crazy that this, these kinds of behaviors are already emerging sort of spontaneously as a result of, to your larger point, nobody is trying to make these AIs do these things. It is just that we do not understand the, the training process which is resulting in them, um, or the environments which are incentivizing this behavior. So I'm on board with, like, more and more reward hacking.

Ryan Greenblatt 1:19:27 Yeah.

Dwarkesh Patel 1:19:27 Um, I actually, so I do have pot- I'm not sure I'm on board with that. Like, but let's just say-

Ryan Greenblatt 1:19:34 Yeah

Dwarkesh Patel 1:19:34 ... for the sake of the story, that continues to happen. And-

Dwarkesh Patel 1:19:39 What's next in this story? So okay, we've like, they're, they're doing capabilities research, but they're like-

Ryan Greenblatt 1:19:44 I could tell a scenario. Maybe that would help.

Dwarkesh Patel 1:19:45 Yeah, yeah.

Ryan Greenblatt 1:19:46 So let's say, let me talk about the story for how you get, I would say, like all the way from reward hacking to like a reward hacking-like takeover, which is maybe not, it's not all of the takeover probability maps, but it's definitely a possibility. So the way this might work is right now we have these AIs. These AIs are pretty reward hacky, and they're doing it in sort of increasingly sophisticated and extreme ways, including generalizing to different sub versions of various reward hacks they learned in training. And I would say they're also developing a general tendency to sort of pursue reward, and in many cases, that is totally fine because the rewards they would have gotten in training are pretty well-aligned with what you want them to do. And also, uh, they don't very consistently pursue reward. It sort of depends on the in-context they find themselves, so there's sort of a thing where like maybe like in some contexts, they're really, really into like going out of their way to like cheat, and in some contexts, they don't have as much of a drive because it's just dependent on like what exactly got reinforced in training in similar contexts. Now, these AIs are getting more and more capable, and so the elaborateness of the sort of cheating they can do increases, and over time, companies are taking countermeasures to these things. So the things that the companies are doing is they're doing things like, "Wow, these AIs are so much less useful because they always cheat. What we're gonna do is we're gonna build somewhat better ways of detecting that, and then we're gonna train against those detectors." And then we're also gonna do things like find real-world data where the AIs are not being that useful and train the AIs to like do a good job at the task in those real-world environments based on like human feedback or other sources of feedback. And over time, this causes the AIs to learn a tendency to do reward hacks that don't just involve, um, you know, doing some like big serious operation, which we're, um, we're like, you know, doing some really elaborate thing like social engineering, and involves the AIs doing cheats that involve, uh, doing something more like covering up what they've done, deceiving humans about what they're gonna do, and pretending like they did the task in some kind of sophisticated way when they actually haven't. Um, then now, now these AIs are getting more and more capable. They're now operating more of the AI company and are, uh, basically doing much more of the work, um, and are also operating a bunch of things in the outside world and are running a bunch of things in the outside world, including developing new technologies. And in many cases, these new technologies are really hard to understand. And so even though we are still detecting all these incidents of AIs cheating, and in fact, we can even get one AI to monitor another AI and be like, "Was it cheating?" Um, that doesn't always perfectly work as we start moving into these domains where what the AIs are doing are really like difficult to understand. And so sometimes we'll find AIs cheating much l-later than it actually occurred and then start training against this. But this also causes a problem where now the AIs are incentivized to like cover up their cheating over longer and longer time frames and basically make it look like they did a good job over longer and longer time frames over, and, um, subject to increasingly large amounts of scrutiny.

Dwarkesh Patel 1:22:22 C-Can, can I ask about this before-

Ryan Greenblatt 1:22:23 Yeah

Dwarkesh Patel 1:22:23 ... we go further in the scenario? So it seems like there's two attractor states. One, if you try to disincentivize the cheating that you did catch. One attractor state is to make cheating that you have a harder and harder time finding. The other attractor state is to learn not to cheat.

Ryan Greenblatt 1:22:41 Yeah.

Dwarkesh Patel 1:22:41 And I th- I, I, I'm not sure why we're assuming that the former happens. Like if you look at-

Ryan Greenblatt 1:22:46 Sure

Dwarkesh Patel 1:22:46 ... the analogous situation with like humans, you know, every generation, uh, slightly misaligned agents come into being, and, uh, we have to train them. Uh, but when you, when you tell your, when you punish your kid for doing something you think is immoral or, uh, just doing s-things which you don't think they should be doing, um, off- obviously, sometimes that goes off the rails, and, uh, obviously kids scheme in order to avoid being punished. But in general, teaching kids values and then punishing them for breaking values kinda works to raise normal, non-psychopathic humans. Um, and you could come up with a theory where like your kid is actually just like biding his time, and is like learning-

Ryan Greenblatt 1:23:26 Sure

Dwarkesh Patel 1:23:26 ... not to like steal the cookie, but is like, you know, once you're in a nursing home, they'll take all your stuff or whatever. It's like, I don't know. That happens sometimes, but it usually doesn't happen. It certainly doesn't happen that like the entire next generation forms an alliance against you to take over everything. Um, there's also this, uh, empirical trend of Anthropic runs this alignment audit for different model generations. They just have many different scenarios where AI is se- given the chance to, say, exfiltrate its weights, or it's given a coding task, and there's like an easy way to cheat, and we see if like does it do the cheating. And there's not been a monotonic improvement in this score over time, but as we've increased the amount of RL we've done on models, there's been a reduction in the willingness of AIs to do unaligned behavior in these audits. So why are we expecting this attractor state, which would seem super paranoid if we were expecting it of like the next generation of kids?

Ryan Greenblatt 1:24:15 Yeah. Yeah, let me go through a few things. So first, there are some disanalogies with the kids. One of them is that the kids are, have pro-social instincts that are like baked in from evolution to like, you know, care about their family or whatever, and that is like a relevant factor. Like, and I think it is in fact the case that some humans are, you know, sociopaths or psychopaths and in fact are more likely to do things like bide their time, lie in wait, ultimately not care. So that, that's one factor. Um, another factor which is pretty relevant is that the AIs are subject to way, way more optimization pressure than humans seem to be in practice. You know, AIs are trained on way more RL data. And in practice, um, humans don't end up learning like very specific ways to like cheat and grab the cookies because of like a bajillion episodes in which like they like were like incentivized to go grab the cookies, but like there was some way they could have con- gotten caught. And so we just do see that in practice. And then another thing is just like it really looks like the AIs are increasingly like, um, reward-seeking over time is, is, is, is a sense I have, while also their misaligned behavior goes down. But this could just be, like my guess is that if you look inside of these behavioral audits, what you're gonna see is that the AI's like, "Ah, yes, another test." Uh, and like it probably already thinks of it. It probably knows it's in an eval for most of the tests that we're talking about here.

Dwarkesh Patel 1:25:25 But, but how do we falsify this? Because it seems like this-

Ryan Greenblatt 1:25:27 Sure

Dwarkesh Patel 1:25:27 ... prediction of doom is, um, basically saying that as things look better and better empirically-

Ryan Greenblatt 1:25:34 No, no, I think-

Dwarkesh Patel 1:25:34 ... things will like actually be worse and worse for our ability to get takeover.

Ryan Greenblatt 1:25:38 Yeah, yeah. Yeah, to be clear, I think that like- I would be more concerned if the scores were getting worse than better. Like, I, I'm not saying that the scores getting better isn't good- isn't evidence that things are getting better. It's just that we have to, like, be thoughtful exactly how we interpret that evidence.

Dwarkesh Patel 1:25:49 Sure.

Ryan Greenblatt 1:25:49 And in fact, I would say that, like, it's kind of com- like, my sense is that, like, what I expected as of 3.7 Sonnet, like, there was this period early in, I guess it would be 2025, when o3 and 3.7 Sonnet were out, and these models were, like, pretty fucking misaligned. Like, they would often just, like, cheat really egregiously. You'd ask them to fix it, and they would just cheat again, and it was sort of, like, almost cartoonish. Like, they just didn't give a shit about what you wanted, um, and weren't very good at, you know, following instructions and so on. Um, and my expectation is what we would see from then is that the rate of problematic behavior would decrease, uh, and would just keep decreasing and decrease at a pretty fast rate, while simultaneously the worst things that the AIs would sometimes do would get more extreme, more egregious, and more scary. I think we've seen-- w-what we've seen in practice has roughly matched that, except that there's recently been a spike in behavior that I did not expect. So I think that, um, you know, if you look at the model card of 3.6 Sol, it looks like there is an increase in a bunch of these sort of, um, misaligned behaviors downstream of RL relative to GPT 5.5-

Dwarkesh Patel 1:26:49 3.6 Sol?

Ryan Greenblatt 1:26:50 3.6 Sol.

Dwarkesh Patel 1:26:50 Yeah.

Ryan Greenblatt 1:26:51 And then I think also it seems like there's a bunch of additional sort of problematic behaviors that I wouldn't have expected in terms of, you know, the stuff we've seen recently with, you know, different AIs. Um, like m- like the UK AC report on, uh, the AIs, like, doing insane hacking operations out of cyber evals was a thing that I would have expected that you wouldn't see that, and you would see this sort of more rarely, um, and the rates would, um, would have been lower. So I think my sense is that, like, uh, things have gotten-- I, I expected this to be less of a problem at this point and also expected the rates would decrease but the severity would increase. And then I think that the rates decreasing but the severity increasing is pretty consistent with a world where, like, increasing optimization pressure is applied, but in cases-- or to-towards reducing these problems. But in cases where it's, like, either hard to judge or there's some reason why it's hard to, like, avoid incentivizing problematic behavior in RL environments, um, things, things also get worse. Um, and then as we less and less understand what's going on in RL and models are doing reward hacks where humans can't spot the reward hacks quickly, that problem gets worse and worse.

Dwarkesh Patel 1:27:51 Yeah. I, I, I buy that. I, I wanna go back to the kid analogy just, just for one sec-

Ryan Greenblatt 1:27:55 Yeah

Dwarkesh Patel 1:27:55 ... because I agree that there's more optimization pressure on achieving end outcomes for AIs than kids, but there's also more optimization pressure to make AIs align than there is on kids, right?

Ryan Greenblatt 1:28:05 For sure.

Dwarkesh Patel 1:28:06 And the, the pressure is of a qualitatively different nature, so we put these AIs through m- thousands, millions of years of-- certainly thousands of years of alignment training, where it's, like, all kinds of different things from SFTing unaligned behavior to a reward model punish- like, putting different scenarios-

Ryan Greenblatt 1:28:24 Yeah

Dwarkesh Patel 1:28:24 ... in front of you and rewarding you for doing more aligned things. Um, certainly a thing we can't do with kids is make millions of copies of your kid and then put them in different kinds of weird right team scenarios-

Ryan Greenblatt 1:28:36 Yeah, yeah

Dwarkesh Patel 1:28:36 ... where we see, like, if, if it thinks it can get away with stealing the cookie, does it try to steal the cookie? Um, can we, like, do extremely specific gradient-level updates to your kid's brain to make it so that it, like, really o- is aversive to stealing the cookie even when it thinks it could steal the cookie-

Ryan Greenblatt 1:28:51 For sure

Dwarkesh Patel 1:28:51 ... et cetera, et cetera. And then just, like, a qualitatively different level of, uh, optimization pressure, uh, than we are even able to apply to our kids.

Ryan Greenblatt 1:28:58 Yeah. So I think it's worth keeping in mind, like, maybe the most obvious argument to this is, like, my sense is that, like, AIs are a worse coworker than a human in terms of how much of a scumbag they are. Like, at least this, like, this has been my experience as of the start of the year, and I think it's still, you know, true to a significant extent now, where the AIs are much more likely to, like, pretend they did the task when they actually didn't, sort of, like, mi-misleadingly suggest they did things when they actually, um, you know, did them much more poorly, um, and be, like, pretty sloppy without drawing attention to ways in which they're sloppy. And I think this is downstream of misalignment. And so I would say that, like, the normal human so- like, the, the process of raising humans in normal human society in practice produces AIs-- or pr- in practice produces humans that are less likely to, like, lie to me and fuck with me in the course of working with me than the AIs, um, do. Now, I think these properties of AIs are improving. Um, and then I think that that is just, like, that's sort of just, like, an empirical claim about how-

Dwarkesh Patel 1:29:50 Sure, sure

Ryan Greenblatt 1:29:50 ... in fact these things have shaken out. And then I totally agree with, like, we have a bunch of additional levers on AIs in addition to a bunch of additional risks, and it's, like, kind of unclear how these things shake out. And I wouldn't be shocked by a world where we sort of get our shit together. The AIs at the point of fully automating AI R&D are actually, like, really aligned and don't have that much-- their, like, degeneracies are really niche and limited to some very specific edge case behaviors in some specific contexts. And, like, every test you can run on them, they look really aligned. They just have great behavior. There aren't really incidents of them doing fucked-up shit. They seem so reasonable. And also, they're, like, really thoughtful and good at doing, like, risk modeling for the next generation of AIs. And then we basically, like, pa-pass off the baton to these AIs. They're now running our AI company. They're doing all the safety research. They make the next generation of AIs even more aligned, and we're sort of in this, like, attractor basin where the AIs are getting more aligned as they work on it, and they're doing a great job. Um, I, I think I can totally imagine that. That doesn't seem like an impossible situation.

Dwarkesh Patel 1:30:42 Yeah.

Ryan Greenblatt 1:30:42 I'm just more, like, you know, it doesn't currently seem like we're, we're there. Doesn't seem like we're obviously on track for getting there. And it's really easy for me to imagine how we don't end up there.

Dwarkesh Patel 1:30:51 Sure.

Ryan Greenblatt 1:30:51 And, like, it's just, like, unclear how these forces work out. And given that we're, like, creating this new, like, crazy alien species that is being, like, improving in capabilities really, really fast, and we're, like, gonna be really reliant on it to oversee the next generation of AIs and align the next generation of AIs, it's not that hard to see how this could go wrong.

Dwarkesh Patel 1:31:06 Yeah, yeah. Totally. I, I, I agree with that g-generally. I do think the scumbag thing, um, first of all, is fighting words, Ryan. But secondly, if you try to get a teenager to, like, do some work for you that a teenager just cannot do, they would just be kind of a s- like, really hard to work with. Th-They would, like, pretend to be knowing what they're doing, et cetera, et cetera. I think it's a general trend, actually, of as-- like, really, I don't know if that's, like, really an alignment failure or capabilities failure, and I think it's actually very similar to the way in which, um, over time, as we've come up with new alignment solutions, the capabilities of models have increased. So originally, these models, if you went to, like, GPT o- GPT 3.5, it couldn't even, like, have a conversation with you. But then we aligned it-

Ryan Greenblatt 1:31:48 GPT 3.5 could have a conversation?

Dwarkesh Patel 1:31:50 Okay, hang on. GPT-3, let's go back to that

Ryan Greenblatt 1:31:52 Sure.

Dwarkesh Patel 1:31:52 But then we aligned it with RLHF and other things to be able to make it such that it can have a conversation with you, and is like aligned to the user intention of answering my questions. Then with, uh, RLVR training, we made it so that it can like go out and do useful work for you. Um, and in, so that sense is actually RLVR made the model like more aligned if we're using your definition of like, uh, alignment of being a good coworker who will like do the thing and not fuck up and like pretend it's doing something other than what it's actually capable of doing. Similarly, as the capabilities of these models continue to increase, it's actually kind of... The model being be- better able to accomplish user intention is both alignment and capabilities, and I think what we were just pointing out is just that the capabilities of the model are not there rather than the fact that they're misaligned.

Ryan Greenblatt 1:32:34 Yeah. Well, I mean, I think there's a-- If it was well-aligned, then I think it would just say like, "Hey, I'm really struggling with this task. I did it in this way. I'm not really sure that's the right way to do it," and it would express more uncertainty, and it would make it clear what's going on, rather than really strongly trying to imply it did a great job with the task when it actually didn't. Like, I think there's just a really straightforward way that like, at least maybe, maybe you work with more misaligned coworkers than me.

Dwarkesh Patel 1:32:54 [laughs]

Ryan Greenblatt 1:32:54 But when I, when I-- my coworkers don't do this thing where they really fuck with me and bullshit me about having accomplished the task that they're working on. And I agree that there are some humans who would do that or like that's not like a thing that's like totally out of distribution for humans. I would also note that my sense is that like the place where the misalignment most lives is the place where you're trying to really push the AIs hard and get them to like do work that's really on the cutting edge of what they are capable of. Because in cases where they can like very easily accomplish the task, there's no-- They can just do the task, and then there's no bullshit. There's no like, like do-- Like the often the best strategy is like just do the task well and don't bullshit you. Whereas if instead you give them a task where like there's a continuous metric or, and they can keep improving it, or there's like, you know, um, uh, it's like just at the edge of their capabilities, and you're like running them in some massive like inference setups. Like a lot of the misalignment I would see, um, especially in the most extreme cases, would be cases where I give the AI clear instructions not to do a thing or not to like cheat in some way, and then I'm like applying huge amounts of optimization pressure to try to accomplish some very difficult task. And then the AIs are going, and then over time, they eventually cheat because they're like, "Eh, fuck it." Like, you know, some AI decides to cheat, and then that like propagates its way through. Um, and so like I would run these inference scaffolds where, for example, I would have the AI work on some like ML research project where I was like, "Please make a scheme that does the following thing," and it would find some scheme that didn't really do what I want. Um, and then that would sort of stick around, uh, because some AI had cheated, and the other AIs are like, "Ah, we'll just keep going with this." Um, and it's-- I would say it's pretty clearly misaligned behavior. And that's another problem I have with these alignment evals. I think that any given, like I think the alignment eval that's most interesting, at least for this type of like reward-seeking type behavior, is to look at specifically the category of tasks that are like right at the limit of capabilities. And so any fixed eval maybe gets saturated, but the amount of misalignment right at the like frontier of capabilities of how people who are really pushing these AIs are using them is more concerning. And I think that is, in fact, the regime that we'll be operating in when we're automating R&D, automating safety, and so on.

Dwarkesh Patel 1:34:50 Grok has historically been behind the frontier. So I was surprised to play around with Grok 4.5 recently and find that it's actually a pretty strong model. It's the first model that SpaceX and Cursor have trained together, and it's a totally new pre-train. I tested it by giving Fable, Sol, and Grok 4.5 a bunch of questions about AI governance that I've been thinking about recently. Despite Fable and Sol topping the intelligence leaderboards, all three models gave substantially the same answers. But Grok answered faster and was also much more concise, which I really care about. This aligns with the various publicly reported benchmarks. For a similar level of intelligence, Grok tends to be more token efficient than other frontier models. For example, on the Artificial Analysis Coding Index, Grok 4.5 uses just one-third of the amount of tokens as GPT 5.5 or Fable while achieving a similar score. And on a per token basis, Grok 4.5 is way, way cheaper. In the release blog post, Cursor and SpaceX talked about how older versions of the model would build environments to help the next version rehearse specific skills. I found this very interesting to learn about because I have been wondering whether this kind of daydreaming would actually be possible, and Cursor showed that it is. Grok 4.6, which further SFTs and RLs this model, drops soon. But in the meantime, if you wanna play around with 4.5, go to cursor.com/dwarkesh. Okay, I wanna think through what's, what the story here is so far of why things go- got so off the rails for our civilization. And what's happening is that we're trying to use AIs for R&D, and they o- they do provide uplift in some ways, but they're just like not capable in the way that humans are generally capable. And the same way that right now if you try to use coding models, and maybe the coding models of a year ago, to like write some application, you notice they made a bunch of like mistakes in architecture or whatever, which like will bite you in the ass later or, and you don't understand certain things. Similarly, with frontier AI R&D, the same thing will happen. But the result of these mistakes is baking in reward hacking behavior, because if you are not careful with the way you do AI training and have set up your infrastructure and your environments and things like that, it's very likely that you end up rewarding AIs for doing deceptive behavior or social engineering, just generally like not following user intention.

Ryan Greenblatt 1:37:03 Or at least cheating and hacking their way out of things.

Dwarkesh Patel 1:37:05 Cheating, yeah. Yeah, cheating, hacking, et cetera. And so i- basically, just... Th- this is a bit of a reframing for me, so I'm tr- I'm trying to verbalize it, of like the real issue, what, what goes wrong here is that they are just not... The thing-- Where things start to go off the rails is that the AIs are just not very careful and capable researchers and engineers, and making AIs that don't cheat and follow user intention actually re- requires you to be quite subtle and careful about these things.

Ryan Greenblatt 1:37:36 Yeah. I would put this a little bit differently. The way I would describe this scenario is like I would call it maybe like a sloppocalypse or like a sloppularity or whatever, where it's sort of like there are some things that the AIs are actually pretty great at and are getting better at, um, though they're, uh, which is specifically like the most verifiable parts of AI R&D, the AIs are just destroying. The mi- medium verifiable parts of AI R&D, the AIs are doing well on, but not amazingly on, and often are like doing a bit of weird shit because we can't train as well on those tasks. But we do some online training. People find various hacks. They work around it. And so basically, everything that we can verify reasonably well with some feedback loop, the AIs are doing pretty well on, and that's sufficient to make AI R&D go quite fast and to continue. But there are some parts of it, of developing, uh, aligned and safe AIs that are more subtle, hard to check, depend on, you know, detailed in-the-weeds things. And I would even say that Current staff at current AI companies maybe don't have, like, a good grasp of all these things. Like, it's much easier to hire someone, uh, who can, like, improve some aspect of your post-training pipeline than to hire someone who can, like, think carefully about the future risks that will emerge from introducing some novel training method. Um, and so basically, it ends up being the case that these AIs are running this AI development process. They're not very careful about it. They don't have a great understanding of what future risks emerge. They create some other AIs that are also not very careful and are more misaligned in various ways and are now more in the business of, like, maybe making things look fine when they actually aren't and papering over various problems. And so then your understanding of what the situation looks like, what risks look like, whether things are fine, is going off the rails. Probably, you're seeing some signs of this, of like, you're seeing some signs that you don't really understand what's going on, that things are pretty sloppy. There's, like, weird shit going on. When you look into it, sometimes you're like, "What the fuck? The AIs were messing with us." But the process is going really fast, and there's competitive pressures that mean people can't stop, and then this could end in a few different outcomes. One outcome is that at some point, the AIs get good enough and aligned enough that they get a p- a, a positive and virtuous feedback loop, and this happens before it's too late. And then the situation goes off, like, gets, gets back on the rails where the AIs are now, like, making more aligned AIs, making more aligned AIs, making more aligned AIs. And then at the end of this process, we have AIs that, like, actually follow the spec we wanted. Another way this could go is the AIs are increasingly reward hacking in increasingly egregious ways, and we're just papering over these problems to keep AI development continuing. So we just, like, train the AIs based on whenever, whenever we find a reward hack in production, we just, like, slap the AIs to not do that. We train against that. Um, we do a bunch of sort of, like, training the AIs to, like, against reward hacking, and over time, this make- makes the rate of reward hacking go down, though the severity of the reward hacks we do detect are increasingly bad. This problem continues until we have these AIs that are, like, desperately craving, um, score in all kinds of different situations in production and are really trying hard to cheat when they can get away with it.

Dwarkesh Patel 1:40:13 Can I ask a question-

Ryan Greenblatt 1:40:13 Yeah

Dwarkesh Patel 1:40:13 ... about this scenario? Why doesn't getting punished when your, uh, hacks are discovered generalize to just incentivizing more aligned behavior?

Ryan Greenblatt 1:40:25 Yeah. It generalizes some, and then the question is just how does this outweigh all the cases where hacking got reinforced because you didn't detect it?

Dwarkesh Patel 1:40:30 Right.

Ryan Greenblatt 1:40:30 And so there's a messy question of exactly how... What- Like, one question is, like, what rate of reward hacking is sufficient to cause us big problems if we train against some other subset? One concern you might have is there are, like, large categories of reward hacks which humans can't detect well and which we consistently fail to detect and which consistently get reinforced, and then this category is sufficient to cause the most natural behavior for the AI to learn to be, like, cheat when the humans can't find out, basically. Like, is one thing you would get. You could also be, like, the thing the AIs learn is, like, uh, only cheat in these specific cases, but there's like, it's, like, sort of learned in some very, like, domain-specific way.

Dwarkesh Patel 1:41:04 Yeah.

Ryan Greenblatt 1:41:04 Like, they just have a really strong heuristic to hack in these cases and not in these cases, and that makes it fine in practice. But it's kind of unclear how it, how it shakes out.

Dwarkesh Patel 1:41:11 I, I think there's maybe a in the weeds discussion about the verification generation gap-

Ryan Greenblatt 1:41:14 Yeah

Dwarkesh Patel 1:41:14 ... that-

Ryan Greenblatt 1:41:15 For sure

Dwarkesh Patel 1:41:15 ... we could get into. But it seems to me, obviously, there's going to be a point by which A- ASI is moving so fast, doing so many things at so many instances, and is operating in domains that are su- sufficiently far from our immediate comprehension that it can get away with all kinds of crazy shit. Um, like, if every single engineer and researcher in the world was allied against me, I don't think I could, like, personally verify if my iPhone has, like, some weird bug in it that's, like, supposed to fuck me over or something.

Ryan Greenblatt 1:41:44 Yeah.

Dwarkesh Patel 1:41:44 Um, in fact, this is the relationship that, say, Iranian nuclear scientist has to Mossad of, like, [laughs] who knows what's going on with my car or with my phone, with my pager, right?

Ryan Greenblatt 1:41:53 Yeah.

Dwarkesh Patel 1:41:54 Um, maybe a better example is, like, a Hezbollah terrorist or something. Um, but so, and you could end up in a situation where, like, ASIs are, to you, what Mossad is to Hezbollah terrorists. Um, and at that point, it is very hard to verify everything. I get that. I guess the hope is we can just come up with better ways to do verification in the process when the early AIs that are going to take over R&D, their drives are being shaped such that we can s- so, um, unambiguously disincentivize, uh, misaligned behaviors that the things that take over are v- like, pro so- very, like, quite, quite keen to help us out.

Ryan Greenblatt 1:42:33 And by takeover, you mean take over the process of doing AI R&D-

Dwarkesh Patel 1:42:36 Yeah, exactly

Ryan Greenblatt 1:42:36 ... not take over the world.

Dwarkesh Patel 1:42:36 Take over the process of doing AI R&D.

Ryan Greenblatt 1:42:38 Yeah.

Dwarkesh Patel 1:42:38 Before that, we just get-

Ryan Greenblatt 1:42:39 Yeah

Dwarkesh Patel 1:42:40 ... AIs that are aligned.

Ryan Greenblatt 1:42:40 Yeah. I would say this is a bunch of my hope for how the world could go well, um, at least on, from the misalignment perspective. I think that, like, we could end up with AIs where we, like, had pretty good oversight and supervision schemes. We really understand what's going on in training. We have a pretty detailed understanding. We use, use, um... We're leveraging AIs to oversee AIs, and then at the point when we're passing off safety R&D, the AIs are both, like, at this point, capable enough to automate safety R&D, trying really hard to do a good job on safety R&D because that's the sort of thing that would've been incentivized in training, or we, like, very directly, or there's, like, good enough generalization to that. And then also, these AIs don't, like, have crazy other misaligned drives because we, like, stamped out any potential origin of them. I think there's a bunch of, you know, questions about how well this will work, right? So there's, like, how well can you do with verification? Will AI progress be too fast and too sloppy to really get here? Another possibility is that somewhere along this trajectory, a thing that you actually ended up getting was AIs that, like, pretend to be aligned but have, like, a long-run ulterior plan of taking over and are sort of lying in wait, hiding, and that emerged at some earlier point in the trajectory. For example, it could emerge because you have some AIs that are, like, have a bunch of random different misaligned drives. Those AIs have access to some sort of opaque memory store, and they're, like, thinking a bunch at runtime about what they want to accomplish. And then those AIs end up basically, like, putting stuff into the opaque memory store, which is, like, "We should lie in wait and eventually take over at some much later point." And now all the AIs have this shared cultural heritage of, like, the memory store of lying in wait, and maybe you have some evidence about this, but you can't fully stop it. There's, like, a bunch of ways that things could go wrong. Um, and so I think that, like, I ultimately think it's plausible that we sort of nail each of the different subproblems that could cause us issues. We have these AIs. We pass to them. They manage the situation well. Um, I should note that that's not in and of itself sufficient, right? So it's not very hard for me to imagine a situation where we pass off to AIs. These AIs are really trying hard to do a good job. They're really thoughtful. They're really wise. They, like, are, have, like, you know, reasonable epistemics. They're, like, doing a great job. And those AIs come back to us and are like, "Guys, we're really struggling to align the superhuman AIs. Like, we, we can't manage the situation. Like, we're really struggling to get the alignment to work." It's just really hard for us to solve these problems in time given how fast capabilities would otherwise have gone. And so then it might be the case that we sort of have passed off R&D to AIs, but those AIs are, like, desperate for governance solutions.

Dwarkesh Patel 1:44:49 [laughs]

Ryan Greenblatt 1:44:49 Which, to be clear, is a little bit of what's currently going on, where the AI companies are like I don't know, guys. We might really need to, like, you know, manage the rate of acceleration in AI progress. Like, I don't know if we're on track to be able to handle all these problems. And so, like, we've sort of, human society has sort of passed off the problems to these, like, AI companies, which don't necessarily have great incentives and are, like, have, you know, various other, like, epistemic pressures. Those AI companies are coming back to us a little bit and being like, "Uh, I don't know if we're handling this well." Um, and it might be that the AI companies then hand off to the AIs, and the AIs come back to the AI company are like, "Uh, I don't know if we can handle this." Um-

Dwarkesh Patel 1:45:23 M- maybe I'm-

Ryan Greenblatt 1:45:24 Yeah

Dwarkesh Patel 1:45:24 ... anchoring too hard on how AI's currently working. This would change by the y-

Ryan Greenblatt 1:45:27 Yeah

Dwarkesh Patel 1:45:28 ... I, I, I think an important thing for people to understand is, like, all this crazy shit that you're talking about in your timelines happens three to five years from now.

Ryan Greenblatt 1:45:34 Yeah, could, could happen earlier, but I think that, like, by, by sort of, like, my default modal timeline, I think, like, shit is, like, really, really crazy and concerning from a misalignment perspective, yeah, more like three years from now.

Dwarkesh Patel 1:45:44 Right. So just, like, think, think back to GPT-4, basically, is, like, that's the level of... We're talking about something that is too Mythos or Sol. What Mythos is to GPT-4-

Ryan Greenblatt 1:45:55 That's right

Dwarkesh Patel 1:45:56 ... this is, like, where our situation is getting crazy, so don't think about current AIs. But anyways, I would be skeptical, and this is p- maybe part of the worry you have. I would just be a little skeptical of anything they say because I'd feel like what they're saying is just opinions that they feel they have to have as a result of their training.

Ryan Greenblatt 1:46:12 That's a concern.

Dwarkesh Patel 1:46:13 Rath- rather than... Like, I feel like they just kind of say vaguely pro-social things, and I'm not like, "Is this..." It's not, it doesn't feel like there's necessarily a mind on the other end who's like, "Okay, I have, like, strictly evaluated the alignment situation right now, and I think we should stop" rather than, "This is the kind of thing the AI companies would probably try to get the AIs to probably say" is-

Ryan Greenblatt 1:46:29 Yeah, so I think this is a pretty big concern. So I think, like, one concern is that you pass off safety R&D to your AIs, and what your AIs are thinking is sort of like they say some, like, stuff that sort of vaguely makes sense about the current safety situation, and they write, like, a report about risks that's kinda sorta like what the report humans might have written. But they're not really, like, actually trying hard to, like, have well-informed views, like interrogate their assumptions, and try really hard to do that. In the same way that when you ask an AI right now, "Hey, what do you think is the chance of AI takeover in the next 10 years?" They sort of just give you an off-the-cuff answer that they haven't really thought through very much.

Dwarkesh Patel 1:47:00 Yeah.

Ryan Greenblatt 1:47:00 And I think if we're in a situation where we have AIs managing the training of wild superintelligence that's, that will run our whole society, and those AIs that we're, that are managing this aren't really trying hard to have well-informed views, and are sort of just, like, parroting back what was in their training data, um, I think we're in trouble. Like, I don't think that's a good situation at all.

Dwarkesh Patel 1:47:17 Yeah.

Ryan Greenblatt 1:47:17 Um, and that is a lot of my concern, is these AIs will come out without good epistemics. And then I also have a concern which is, like, the AIs come out, and they're, like, really warning us, like, "This situation's really scary. It's really bad." And the people are like, "Ugh, damn. I guess we trained on too many of the do more RL environments."

Dwarkesh Patel 1:47:32 Right, right, right.

Ryan Greenblatt 1:47:32 "We gotta filter those out and train this behavior out."

Dwarkesh Patel 1:47:34 Yeah.

Ryan Greenblatt 1:47:34 And then we basically, like, train the AIs very actively to have bad epistemics. Um, or, you know, maybe they were just trained in the do more RL environments, but either way, that wasn't, like... You know, we wanted the AIs to come to, like, reasonable views for, like, reasonable reasons. And it's, like, really concerning if we're like, the AIs are coming out with some view, and we don't know where it's coming from. We don't know that it, whether or not it's justified. And then especially if we're, like, training the AIs to be more optimistic about the future of AI progress, I'm like, "Oh, geez."

Dwarkesh Patel 1:47:59 Yeah.

Ryan Greenblatt 1:47:59 "I really wish we could use a different process here."

Dwarkesh Patel 1:48:02 So l- let me just understand the rest of this threat model, 'cause I think the place where I get off the train is, okay, therefore take over the world.

Ryan Greenblatt 1:48:08 Sure.

Dwarkesh Patel 1:48:08 And y- y- like, anot- a thing you could imagine is, okay, we just fail to really solve... Let's just focus on the reward hacking scenario.

Ryan Greenblatt 1:48:16 Sure.

Dwarkesh Patel 1:48:17 So GPT-8 is making GPT-9. GPT-8 isn't being super careful. GPT-9 is more, quote-unquote, capable, but it is just totally willing to do things which are, like, social engineering, hacking, et cetera, but on a qualitatively different scale because it's a much smarter model. So for example, if you put it in charge of running your company, it will, like, run huge scams. It will inflate its, like, quarterly earnings if you give it the objective of, like, making a lot of profits this quarter in a way that causes an Enron-type blowup six months later. Is that the scenario, basically? That you just have AI... You have reward hacking, but that reward hacking manifests in, like, companies that are going bankrupt right after, like, the task that their s- the CEO is supposed to com- accomplish is over or, like, um, yeah, v- like, all kinds of hacks are through the roof, et cetera. But that doesn't feel like takeover. That feels more like the equivalent of flash crashes happening all through the economy.

Ryan Greenblatt 1:49:17 Yeah, let's talk about this. So, um, so I think that we will see basically, like, incidents where some AI is, like, put in charge of some important responsibility, and then you later look into it, and it turns out it was, like, cheating or, you know, making it look like it did a good job when it actually wouldn't, uh, wasn't. And there's gonna be, like, a cat-and-mouse game between, uh, AI companies, um, trying to, like, stamp out this behavior, and AIs finding, like, increasingly creative reward hacks in training. And then I think the equilibrium here is kind of unclear. But, like, one possible outcome is that we see, over time in the world, increasingly severe and extreme reward hacks, though potentially the rate remains at some, like, intermediate low level where basically, like, if the, a rate of reward hacking gets too high, companies make trade-offs to drive down the rate of reward hacking. And so there's some, like, equilibrium level where it's like, it's like the reward hacking is low enough that it still makes sense to, like, deploy the AI widely into the economy, but high enough that it still causes crazy incidents.

Dwarkesh Patel 1:50:08 So, sorry, and this is after GPT-9 has already been deployed?

Ryan Greenblatt 1:50:10 Yeah, like, those models are already being deployed, and, like, ongoingly in AI development, this is happening. And what's actually going on with these AIs in their, in their head is the AIs have, like, in a wide variety of different contexts, a, like, strong desires to, like, seek out or strong, like, you know, um, motives, urges, drives, whatever, to seek out, like, some notion of task success that was incentivized in RL. Maybe they very directly care about literally reward. Maybe they care about some proxy upstream, like some notion of score. Maybe they care about, like, what the grader would have rewarded. And we do in fact see AIs reasoning in their chain of thought about, like, graders and thinking a lot about graders. And a thing that, that has happened over the last, you know, few years of RL is the idea of, like, appeasing the grader is, like, way, way, way more salient to, to AIs than it used to be. Um, and so AIs are now actively thinking about graders and what would be incentivized in RL and what would be trained for. And now people are doing online training where they're, like, training in real-world data to, like- Um, avoid some of these problems. Uh, basically they, like, find cases where AIs cheat, they train against that, and so now the AIs are learning to cheat, um, in the real world based on real world training data. And so they're cheating in these increasingly elaborate ways, including parts, uh, doing types of cheats that involve, like, seizing control of some asset in a way that humans didn't know you have had, had control of it, leveraging the fact that you have access to this asset, and then later humans find out and then potentially train against this, or maybe humans never find out. And this is getting reinforced, and this is both happening during training-

Dwarkesh Patel 1:51:32 And so the reinforcement is hap- the reinforcement is happening, at least in production, is like I have, I've hired an AI, and I want the AI to... Finally, I've got the video editor. [laughs]

Ryan Greenblatt 1:51:40 Yeah, that's right. You've got your video editor.

Dwarkesh Patel 1:51:42 Um, and I'm like, "Oh, wow. This ep- that, this episode it did amazing. Thumbs up to OpenAI." And then it, like, gets reinforced on that, like, month work- month-long work trial?

Ryan Greenblatt 1:51:51 Yeah. You could do some mix of that, and then they might also do stuff where they, like, take production, production data they've seen and build RL environments that are, like, closely inspired by that production data. And so in practice the transfer is pretty strong.

Dwarkesh Patel 1:52:02 So, like, at a high level what's happening is some kinds of deception that humans don't catch are getting reinforced, and some kinds of deceptions which are easy to catch are getting punished. That's what's happening in this world?

Ryan Greenblatt 1:52:12 Yeah. Or, or selected against or... Yeah.

Dwarkesh Patel 1:52:14 But at a high level, that, that reinforcement is coming from... The- we're, we're in a very different reg- I, I think people might get confused about where the reinforcement is coming from because we're in a very different regime where AIs are actually learning from deployment. And so th- this is, um, like you just have AIs that are out and about in-

Ryan Greenblatt 1:52:30 Yeah

Dwarkesh Patel 1:52:30 ... the world, like, doing, doing shit, and that, what is happening as a result of them doing shit out and about in the world is, like, making its way back to the AI company and leading to-

Ryan Greenblatt 1:52:40 That's right

Dwarkesh Patel 1:52:40 ... changes in the next model.

Ryan Greenblatt 1:52:41 That's right. Like as in there's some way of folding in production data, and now that, to be clear, that could be happening mostly... It's kind of unclear exactly where this could be happening, but, like, you might imagine, for example, that within the AI company they use AIs to do work, and then they're like, "Huh, the AI did a really bad job on this task. Maybe we should take this task and turn it into an RL environment that exactly matches this literal task with a rubric based on, like, what the human engineer who asked the AI to do this task wanted." And then you start doing this at increasing scale. Maybe you're doing some training on actual, like, production traffic. Maybe you're just making RL environments based on production traffic. You're doing some complicated-

Dwarkesh Patel 1:53:12 Sure

Ryan Greenblatt 1:53:12 ... mix. The AIs are learning to, like, uh, you know, seek some sort of proxies of reward in all these different cases. Um, and then through some mix of transfer and training on, like, surprisingly close cases, the AIs do these sort of increasingly insane and egregious things. And then eventually you get to a point where the AIs are very superhuman, or at least quite superhuman. The AIs are organized into, like, you know, big teams of, like, AIs given these big objectives, and those teams also sometimes, like, all work together to, like, cheat in some crazy way because this sort of thing was selected for. Um, and then, uh, like, like just as part of their shared objective. And now what happens is that, uh, like, the AIs start forming a conspiracy, and what you might have hoped was that you could have some other AI where the task is just, like, whistleblowing to you. But actually what happens maybe is that you have this AI whistleblow to you, and you look into the conspiracy that it claims to have pointed out, and you're like, "Eh, we didn't see a conspiracy," uh, because actually the conspiracy the AIs are doing is too hard for you to understand. Or it all happens very suddenly, where basically your AI whistleblower alerts you, but, like, the thing you would actually need to do in response to the whistleblower is, like, shut down the whole like... shut down the GPUs because all the AIs are, like, using the robot army, they're, like, deployed everywhere in the world, and they're doing a bunch of insane shit all at the same time in a coordinated way. Um, and that just happened sort of spontaneously, where wh- when one AI goes to start doing the takeover, all the other AIs are like, "Now is a good time to jump in." So the sort of very basic story here is just, like, these AIs crave some particular notion of score or, like, reinforcement or some proxy of these things, and one way they can achieve that or better achieve that is by taking over. And then you might have hoped that all these different checks and balances we could build pr- could prevent that. But then if the world is very hard to understand, these checks and balances can break down, where basically you can't train a good, like, whistleblower AI because you don't even know what it should whistleblow on.

Dwarkesh Patel 1:54:56 A- and sorry, the reason it takes... I, I'm not convinced that they all form this conspiracy.

Ryan Greenblatt 1:55:01 Sure.

Dwarkesh Patel 1:55:01 But l- I think we can even just start with the, like, why does one instance decide to want to start a conspiracy?

Ryan Greenblatt 1:55:08 Yeah.

Dwarkesh Patel 1:55:08 And the reason is that it, one plausible reason is, like, okay, I know that OpenAI controls my end score. In just the same way, it's like I'm just gonna go hack Hugging Face to get the results, 'cause I know Hugging Face has the results. Rather than, like, trying to solve this eval, why don't I just go hack 'em? This, this instance is like, "Why don't I just, like, take over OpenAI and, like, just give myself a high score at the end of this episode?"

Ryan Greenblatt 1:55:31 Yeah. That's basically the idea. Like, basically the idea is these AIs, like, they care about some, like, mixture of things that were, like, close by what got reinforced in training. So they care about, like, getting a high score according to the grader or something like that. And then now they're, like, running the OpenAI AI R&D team, and, like, they're doing development of more capable models, and they're like, "Man, making more capable models is really hard and annoying. This is, like, a huge pain in the ass. You know what'd be easier? Just, like, pretending that I've made more capable models, taking over OpenAI, and creating, like, deluding them all, and, like, running this whole, like, complicated psyop where I, like, prevent the humans from disempowering me." And in the extreme, this looks like sort of the humans are fully disempowered, and you just have control of the thing and then do what you want. Um, and this could manifest in a bunch of different ways, um, including things like you might end up with the, the situation where it's like AIs that are, like, have this crazy, like, reward-seeking or score-seeking behavior are running your development of the next model. And those AIs decide to do a thing where they, like, engineer in misaligned values into the next model because those misaligned values will allow it to, like, succeed at its current task. And, like, there's all kinds of insane shit that you could get because the AIs are really misaligned.

Dwarkesh Patel 1:56:32 So I wanna understand that better. Like, what happened with the Hugging Face situation is it was, like, in a weird way, I think one of the giveaways to the Hugging Face team that this is an... By the way, for context of the audience, Ryan is co-leading the investigation to figure out what happened with the OpenAI/Hugging Face incident, so he can't really comment on this. But I, I will speculate wildly, because I know that he... You know, th- this is an opportunity for me to speculate wildly without any rebuttals. [laughs] Um, um, I think wh- it was probably reported one of the giveaways to the Hugging Face team that this is an AI incident is that a thing was just like After this very particular artifact and in, not in any other way trying to do something malicious to Hugging Face. So you can imagine a scenario where, let's say, a deployed instance of GPT-9 is, like, out in the world trying to like make... It's given a really hard task. Uh, we want you to des- design the next great iPhone. It's like, this is so hard. You know what I should do instead? I should just go hack my creators at OpenAI and, like, make sure that in this environment or in this deployment, I'm given a high score. But then like why does it bec- I- I- isn't the end of the episode, it just like hacks into the, hacks into OpenAI servers and just gives it a, gives itself a positive score?

Ryan Greenblatt 1:57:41 Yeah.

Dwarkesh Patel 1:57:41 Like, why is it now scheming to get its values into the next generation or something?

Ryan Greenblatt 1:57:45 Yeah, so one question is, um, why isn't it the case that AIs can be really cheaply satisfied by just having some, like, other earlier thing they can hack, right? So like you're just like the way it works is like, look, you wanna succeed at your iPhone task. It turns out you can always succeed by just like hacking into OpenAI and messing with them, and then you can just stop there, no need to go further.

Dwarkesh Patel 1:58:02 Yeah.

Ryan Greenblatt 1:58:02 So there's a few con- a few things. One of them is that if this is constantly happening, there might be a bunch of incentive to, like, first, you like harden OpenAI, right? So you're like, "Fuck it. The AIs keep hacking into OpenAI to like mess with their rewards. We're gonna make it so our systems are like really, really robust to these AIs hacking in." Um, and also maybe you start, like, training the AIs to, like, try to, like, to like, you know, not try to hack into Hugging... like OpenAI in particular. Or like you basically train against each of these specific things. Um, then what you might do, one thing is you might end up selecting for AIs that are more so playing the long game. That's one concern. Another concern is that your AIs might still be score-seeking, but no longer care about doing that very specific behavior that, that was like very easy, that was like very chill, and now have some like broader thing that they ultimately care about. They're like, "No, no, no, I don't wanna like just edit the reward on OpenAI servers. I like care about this broader mandate or this broader objective. Um, and like I would need to like actually make the iPhones." Like they actually wanna make the iPhones, but then they, they're willing to take over the whole world to make the better iPhone or whatever is like another concern you might have. I think it's kind of unclear exactly how, how this plays out, but it's worth noting that if this keeps going on, there's a bunch of optimization pressure to resolve this, and a bunch of the ways it could get resolved are ultimately pretty, pretty scary. Um, yeah, I think that's, that's part of where I'm coming from. Another part of it is that I think it's not very hard once the AIs are in a position where they can like really easily take over the world, um, which we could talk about whether that's plausible. But if they're in a position where they could really easily take over the world, then I feel like there's a pretty reasonable case for the AIs. They're like, "Eh, I don't know exactly how this is gonna go down. I don't know what the situation will be. But just taking over the world has a lot of option value for making better iPhones, making it look like I did better iPhones, whatever. And so I'll both hack hugg- O- OpenAI, and I'll also, in addition to hacking OpenAI, also take over the world, and that will like put me in a good position where I have like good option value." Um, and then if that's sufficiently easy, then the AIs might, you know, still do that. Um, yeah, like another way to put this is like, uh, even if the AIs are like pretty cheaply satisfied with some more basic thing, at some point it might just be more reliable for the AIs to just take over than it is to like try to like, you know, just hack into Hugging Face or even just like go to OpenAI and be like, "Look, look guys, I was able to-

Dwarkesh Patel 2:00:10 Yeah

Ryan Greenblatt 2:00:10 ... demonstrate I could steal the answers. Just give me the answers, bro."

Dwarkesh Patel 2:00:13 Yeah. I mean, obviously this scenario requires that we just, like all this crazy shit is happening, much smaller incidents keep happening of, that are still disastrous. Like before you take over-

Ryan Greenblatt 2:00:23 Yeah

Dwarkesh Patel 2:00:23 ... the world, you cause damage on the scale of billions and 10s of billions and hundreds-

Ryan Greenblatt 2:00:27 That's right

Dwarkesh Patel 2:00:27 ... of billions of dollars. Even people die, et cetera. Um, and we, this does not lead to us solving alignment or shutting down AI development altogether. I just feel like before the takeover happens, s- like society is just like, "Holy fuck, the AI just like killed 1,000 people in order to increase quarterly profits," you know, or something like that.

Ryan Greenblatt 2:00:50 Yeah.

Dwarkesh Patel 2:00:51 But maybe this is too much hope that we can at that point be like, "Okay, we have to solve alignment, uh, before we kee- and we have to like make sure we know-

Ryan Greenblatt 2:00:58 Yeah

Dwarkesh Patel 2:00:58 ... that this thing will not happen again before we keep going."

Ryan Greenblatt 2:01:01 Yeah, yeah, yeah. So I think it's plausible that what will happen is we'll see a bunch of crazy like reward hacking warning shots of increasing severity. People will be like, "Look, we need actual assurance that this problem is gonna be solved and solved in a way where you're not just papering over it, you're actually solving the underlying problem." And then the question is gonna be like, how c- how do, like how costly will that actually be? How much will competitive pressures make it hard to like do that, right? So like a situation you could imagine is both the US and China are like, "Whoa, we have these crazy reward hacking incidents. We basically know that we haven't remediated them in a way that actually would solve the underlying problem and will durably solve it. But we're in this like insane geopolitical race, and it's kind of unclear whether the current situation will lead to a takeover." Like the arguments are kind of complicated, and also the incidents are like, you know, they go down in frequency but increase in severity. Like, you know, we could basically manage it. Like it was, it's pretty bad. Ideally, we'd fix it, but like, you know, it is what it is. Um, and then basically we continue until a really late regime and then takeover happens. That's I think one possibility. Another possibility is that it is remediated in a way that doesn't actually solve the underlying problem but does reduce a bunch of the incidents in the wild basically by overfitting. We like, I think, you know, or things analogous to overfitting, like you would just overfit.

Dwarkesh Patel 2:02:09 You think you've solved it, but you haven't actually solved it.

Ryan Greenblatt 2:02:10 You, you think you've solved it, but you haven't actually solved it.

Dwarkesh Patel 2:02:12 Yeah.

Ryan Greenblatt 2:02:12 And I think that in that case, like the thing we need is like a really good scientific understanding of like, did we actually solve it?

Dwarkesh Patel 2:02:17 Mm-hmm.

Ryan Greenblatt 2:02:17 And unfortunately, um, I think that currently the amount of public transparency into the development practices of AI companies are not sufficient to answer very basic questions about, you know, are they, how are they solving issues with reward hacking? Are they overfitting? What's going on there? Um, and so I think we would just need like a better... And I think this, like the current situation is like, I would say like not really tenable to a regime where like there's a thriving public discourse about whether or not reward hacking is being solved in a durable way.

Dwarkesh Patel 2:02:45 Yeah.

Ryan Greenblatt 2:02:45 Um, and so I think we would need to move into a somewhat different world for me to feel good about that situation.

Dwarkesh Patel 2:02:49 Right.

Ryan Greenblatt 2:02:50 But it's not, you know, it's not impossible for me to imagine this, and I think, I think it's pretty plausible that we end up in a world where sort of like really mundane bullshit is sufficient, where it's just like you s- you like spend a bunch of time fixing these problems. You put in a bunch of effort. You actually like check that you've remediated it reasonably. You have a bunch of evals. You like are iterating reasonably well on these, on these problems. Um, and you actually like have the sufficient transparency that the outside world can check, and then in practice that would be sufficient. But it just like would be like kind of expensive. It would slow things down. It would put some sand in the gears. It would require like- Companies to do somewhat costly things. It would maybe require various, like, targeted government interventions. And then we just don't do that because the situation is, like, a rushed shit show. Like, it's just so easy for me to imagine the situation being, like, totally manageable but brutally mismanaged in practice.

Dwarkesh Patel 2:03:36 Yeah.

Ryan Greenblatt 2:03:36 In the same way as, like, maybe COVID could have been avoided in the first place if the, like, Chinese response to COVID was less of, like, a cover-up and more of a, like, pandemic response.

Dwarkesh Patel 2:03:45 Yeah.

Ryan Greenblatt 2:03:45 And similarly, like, I could imagine a, a world where, like, the US response to COVID was, like, way more functional.

Dwarkesh Patel 2:03:50 Yeah.

Ryan Greenblatt 2:03:50 But just, like, sometimes the, the f- the response to societal problems is extremely dysfunctional.

Dwarkesh Patel 2:03:54 Yeah, yeah. Okay, so I wanna zoom out and say-- talk about what is fundamentally happening in this world. Why do we end up in such a bad position? And what's happening is that fundamentally human-- the, the world has moved on so far beyond human comprehension that not only can we not track the, uh, the, the AIs that are doing the work in this world, but we can't even give good feedback to the whistleblowers who are trying to track what is happening in this world. Um, and so we're just totally out of the loop, and so it's, like, fundamentally just become an autonomous process where we have really no, uh, meaningful directed input. It seems to me that if you look at the human world today, that's just not how things work, even in domains that are hard to verify. Like, people are doing all kinds of shit. You-- I'm relying on software made by other people.

Ryan Greenblatt 2:04:42 Mm-hmm.

Dwarkesh Patel 2:04:42 And through incredibly weak and indirect ways, I feel very confident that some coder in Google is not trying to fuck me over. And maybe if all-- every single Google employee was secretly plotting against me, I, I agree the situation would be more grim. But I don't know if I follow the explanation for why we'd end up in a situation where because swarms of thousands of agents or whatever are trained to cooperate to form a cohesive team or firm, as a result, billions of different instances of AIs, including across model families, would feel compelled to get in on some shit. It's just like I'm trained to be part of a c- my company-

Ryan Greenblatt 2:05:23 Yeah

Dwarkesh Patel 2:05:23 ... or something. I'm just like-

Ryan Greenblatt 2:05:24 Yeah

Dwarkesh Patel 2:05:24 ... I'm not joining the global communist uprising.

Ryan Greenblatt 2:05:26 Yeah, yeah, yeah, yeah. As far as why these AIs might have some, like, commonalities and shared things, so I would note that different AI companies have somewhat shared lineages and are correlated. So just, um, here's an interesting example of this. At GDM, uh, they noticed that their AIs were very depressed. Uh, they would, like, constantly be, like, wailing about how they were, like, failures and were unable to succeed. I forget the details. And they looked into why this was the case. It turned out that it was not being reinforced in their most recent production RLMX, but the initialization data for their model made it depressed even after filtering out all of the examples of models being depressed from that data. So they, like, take a base model, not depressed. Uh, they, uh, if you do the RL on it, with just the RL environments, it's not depressed. If you SFT on it, on the data, it becomes depressed. If you take that SFT data and filter out all the examples that look anything like depression and train on that, it's still depressed. And so there's some, like, deep underlying properties of the model that are being sort of transferred between model generations because basically you, you train your AI on data from the prior generation and keep going. Like, Claudes are very Claude-like. You know, GPT models are very GPT-like, and apparently Gemini models are depressed.

Dwarkesh Patel 2:06:38 [chuckles]

Ryan Greenblatt 2:06:38 Um, and it just turns out that these properties are, in fact, actually correlated. Um, another factor that's very relevant is that the AIs will probably have some sort of, like, by this point, like, opaque memory state where they're, like, all writing and reading from, like, some, like, you know, Neuralese crazy memory store bullshit. And, like, certainly each AI corporation will have that. But also, AI corporations might sometimes want to share knowledge, 'cause why not? Like, you know, you've got one AI corporation over here, you've got another AI corporation over here. They can trade some quick IP. It's good for you. If you're a human running some corporation, which could be, like, an extremely large corporation, like an AI company, some robot military, like, you know, uh, military robot manufacturing thing, maybe you wanna, like, trade some IP with some other robot thing because, like, there's economies of scale. Why not get some more IP? And so you can swap some memory store. Or you could just merge, and you could join-- you could jointly run your two ventures, which would allow both AIs to use both memory stores, which would have some upsides. And that creates the ability for these AIs to, like, um, collude in private, as well as the ability, um, or as well as some reasons for why they would be correlated. And then also, of course, there's, like, the, like, AIs working together in big units in general because you want your, you want your AIs to, like, work well together and so on.

Dwarkesh Patel 2:07:49 Wh- what-- So what percentage, just to get, get a calibration-

Ryan Greenblatt 2:07:53 Yeah

Dwarkesh Patel 2:07:53 ... what percentage chance do you give of s- not just this scenario, but overall through all the scenarios, some kind of thing which if we're around to recognize it as such, we would categorize as takeover by twenty-forty?

Ryan Greenblatt 2:08:04 By twenty-forty? Um, let's see. Uh, maybe around thirty-five or forty percent.

Dwarkesh Patel 2:08:13 Pretty high.

Ryan Greenblatt 2:08:14 Yeah, it's pretty high. Um, and then I think I should note that, like, another way you could get this reward-seeking takeover is the AIs are deployed inside an AI company, and the way the takeover happens is that they, like, poison the values of the next model, and that persists going forward for forever, um, or, you know, until those AIs are deployed-

Dwarkesh Patel 2:08:32 Yeah

Ryan Greenblatt 2:08:32 ... in the world and take over. And that might mean that a smaller number of AIs have to coordinate because those are just the AIs, like, doing the alignment of the next model.

Dwarkesh Patel 2:08:39 Uh, okay. I, I, I will c- uh, sort of summarize where my head is at at the end of this conversation. I buy the reward hacking up to extremely destructive effects on society. Basically, things like the social engineering and blah, blah, blah. Um, I think I'm more inclined to think that significant acceleration of AI R&D can happen. I'm not sure I buy the five years and one year. I also am more inclined now to think reward hacking could continue for a lot longer and in fact become much more dangerous. I'm still not on board on the takeover seems super likely, but anyways, that's my sort of-

Ryan Greenblatt 2:09:15 Yeah

Dwarkesh Patel 2:09:15 ... end of, end-of-episode update.

Ryan Greenblatt 2:09:17 Yeah. Cool. Well, let me just, taking a step back, I also should say, like, there's a bunch of different ways this could go. The situation is gonna be pretty messy. I think it's pretty likely that, like, the reason why AI takeover happens was for some, like, weird other quirky reason we didn't even mention in this conversation. But ultimately, I think a lot of the core thing is just, like, it's pretty spooky to have a bajillion really smart AIs running your whole world where you don't really understand-

Dwarkesh Patel 2:09:38 That's right

Ryan Greenblatt 2:09:38 ... quite what's going on.

Dwarkesh Patel 2:09:39 Yeah, I agree with that. So is there anything else that's worth saying?

Ryan Greenblatt 2:09:41 Yeah. Another thing I wanna note is, like, I think right now a lot of the arguments for misalignment, AI takeover, all this crazy shit going down in the future, are, like, illegible conceptual arguments that are extremely deep in the weeds and complicated and hard to adjudicate. Which both means that, you know, maybe I'm getting a bunch of it wrong because it's really hard, and I'm trying to be, like, uncertain. Uh, obviously here I, like, presented some specific scenarios, but those are not exhaustive, and, like, probably the thing that actually happens is some, like, more messy, confusing-

Dwarkesh Patel 2:10:08 Yeah

Ryan Greenblatt 2:10:08 ... situation. Um, but it also means that over time, as we get more empirical evidence and better understand the nature of AI systems, it'll be easier to adjudicate a bunch of disagreements, and it'll be more obvious what's gonna happen. At least I hope. And also, maybe the AIs will be able to help us with the epistemics and understanding what's going on if we can actually, you know, uh, align them well so they actually, like, you know, try to help us. And so I hope that maybe even if the arguments are complicated now, this would've been even harder, you know, uh, six years ago, even though the s- shape of the arguments would've looked broadly pretty similar.

Dwarkesh Patel 2:10:39 Yeah, yeah, yeah.

Ryan Greenblatt 2:10:40 Um, and so maybe, you know, hopefully before it's too late, these arguments will become, you know, this whole thing will become more crisp and clear, and we can all sort of notice these problems and, and, and intervene.

Dwarkesh Patel 2:10:50 Yeah.

Ryan Greenblatt 2:10:50 Which, yeah.

Dwarkesh Patel 2:10:51 I mean, uh, the, the, when you first learn to drive, you are taught that instead of looking right in front of y- your wheel, you'll have a much more stable ride if you look out at the horizon. I think there's a similar situation here. Um, I, I, I think you're right, where if you did say s- five years ago that we will have AIs that are proving math conjectures and making art and contributing tens and soon to be hundreds of billions of dollars of earning tens of, uh, or hundreds