Sergey Levine on Dwarkesh Patel: How Robots Learn on the Job

Dwarkesh Patel 0:00 Today, I'm chatting with Sergey LaVine, who is a co-founder of Physical Intelligence, which is a robotics foundations model company, and also a professor at UC Berkeley. And just generally, one of the world's leading researchers in robotics, RL, and AI. Sergey, thank you for coming on the podcast.

Sergey Levine 0:16 Mm-hmm. Thank you. And thank you for the kind introduction.

Dwarkesh Patel 0:18 (laughs) Let's talk about robotics. So before I pepper you with questions, I'm wondering if you can give, uh, the audience a b- a summary of where Physical Intelligence is at right now. You guys started a year ago.

Sergey Levine 0:28 Yeah.

Dwarkesh Patel 0:28 And what does the progress look like? What are you guys working on?

Sergey Levine 0:31 Yeah. So Physical Intelligence aims to build robotic foundation models, and that basically means general-purpose models that could, in principle, control any robot to perform any task. Uh, we care about this because we, we see this as a very fundamental, uh, aspect of the AI problem. Like, the robot is essentially, uh, encompassing all A- AI technology, so if you can get a robot that's truly general, then you can, uh, do, uh, you know, hopefully a- a large chunk of what people can do. And where we're at right now is I think we've kind of gotten to the point where we've, uh, built out a lot of the basics. (laughs)

Dwarkesh Patel 1:06 (laughs)

Sergey Levine 1:06 And, you know, I think those basics actually are pretty cool. Like, they work pretty well. We can get a robot that will, like, fold laundry and that will go into a new home and, like, try to clean up the kitchen. But in my mind, what we're doing at Physical Intelligence right now is really the very, very early beginnings, just, like, putting in place the basic building blocks on top of which we can then tackle all these, like, really tough problems.

Dwarkesh Patel 1:25 And what's the year-by-year vision? So, um, one year in, now I got a chance to watch some of the robots, and they can do pretty dexterous tasks like folding a box using grippers, and it's, like, I don't know. It's, like, uh, pretty hard to fold a box even with, like, my hands. Um, if you gotta go year by year until we get to the full, like, robotics explosion, wh- what is happening every single year? What is a thing that needs to be unlocked, et cetera?

Sergey Levine 1:47 So there are a few things that we need to get right. Uh, I mean, dexterity, obviously, is one of them, and in the beginning, we really wanted to make sure that we, um, understand whether the methods that we're developing have the ability to tackle, like, the kind of intricate tasks that people can do.

Dwarkesh Patel 2:01 Yeah.

Sergey Levine 2:01 So as you mentioned, like, folding a box, uh, folding different articles of laundry, cleaning up a table-

Dwarkesh Patel 2:07 Yeah.

Sergey Levine 2:07 ... uh, making a coffee, that sort of thing, and that's, like, that's good. Like, that works. Uh, you know, I think that the results we've been able to show are pretty cool, but again, like, the end goal of this is not to fold a nice T-shirt. The end goal is to just, like, confirm our initial hypothesis that, like, the basics are kinda solid.

Dwarkesh Patel 2:21 Yeah.

Sergey Levine 2:21 But from there, there are a number of really major challenges, and I think that, you know, sometimes when, um, results get abstracted to the level of, like, a three-minute video, someone can look at those videos, like, it's like, "Oh, that's cool. Like, that's what they're doing." But it's not. Like, it's a very simple and, uh, basic version of what I think is to come. Like, what you really want from a robot is not to tell it, like, "Hey, please fold my T-shirt." What you want from a robot is to tell it, like, "Hey, robot, like, you're now doing all sorts of, uh, home tasks for me. Uh, I like to have dinner made at 6:00 PM. Uh, I wake up and go to work at 7:00 AM. Uh, I'd like... You know, I like to do my laundry on, on Saturday, so make sure that's ready, this and this and this. Uh, and by the way, check in with me, like, every Monday to see, like, w- you know, wha- what I want you to do to pick up when you do the shopping."

Dwarkesh Patel 3:06 Right.

Sergey Levine 3:06 Right? Like, that's the prompt, and then the robot should go and do this for, like, you know, six months, a year. Like, that's the duration of the task.

Dwarkesh Patel 3:14 Mm-hmm.

Sergey Levine 3:14 So it's a, it's... Ultimately, if, if this stuff is successful, it should be a lot bigger, and it should have that ability to learn continuously. It should have the, uh, understanding of the physical world, the common sense, the ability to go in and pull in more information if it needs it. Like-

Dwarkesh Patel 3:29 Yeah.

Sergey Levine 3:29 ... if I ask it, like, "Hey, um, tonight, like, uh, you know, can you, uh, can you make me this type of salad?" It says, "Okay, you should, like, figure out what that entails, like, look it up, go and buy the ingredients." So there's a lot that goes into this. It requires common sense. It requires understanding that there's certain edge cases that you need to handle intelligently, cases where you need to think harder. Uh, it requires the ability to improve continuously. It requires understanding safety, being reliable at the right time, being able to fix your mistakes when you do make those mistakes.

Dwarkesh Patel 3:56 Yeah.

Sergey Levine 3:56 So there's a lot more that goes into this. Um, but the principles there are you need to leverage prior knowledge, and you need to have the right representations.

Dwarkesh Patel 4:05 So, so this grand vision, what year, if you had to give an esti- uh, median estimate?

Sergey Levine 4:09 Yeah.

Dwarkesh Patel 4:09 Or 25th percentile, uh, 50-

Sergey Levine 4:11 (laughs)

Dwarkesh Patel 4:11 ... 75?

Sergey Levine 4:13 I think it's something where it's not going to be a case where we develop everything in the laboratory and then it's done, and then, you know, come 2030-something, you get a, a robot in a box. I think it'll be the same as what we've seen with AI assistants, that, uh, once we reach some basic level of competence where the robot is delivering something useful, it'll go out there in the world. The cool thing is that once it's out there in the world, it can collect experience and leverage that experience to get better.

Dwarkesh Patel 4:38 Mm-hmm.

Sergey Levine 4:39 So to me, like, what, what I tem- tend to think about a lot in terms of timelines is not the date when it will be done but the date when it will... when, like, the flywheel starts, basically.

Dwarkesh Patel 4:48 Right. Okay, so when does the flywheel start?

Sergey Levine 4:50 I think that could be very soon, uh, and I, and I think there's some decisions to be made. Like, the trade-off there is the more narrow you, you, you scope the thing, the earlier you can get it out into the real world. Um, so, uh, but soon as in, like, this is something we're already exploring. We're already trying to figure out, like, what are, like, the real things this thing can do that could allow us to start spinning the flywheel. But I think in terms of, like, stuff that you would actually care about that you would want to see, um... So I don't know, but I think that single-digit years is very realistic. I'm really hoping it'll be more like one or two before some- something is, like, actually out there, but it's hard to say.

Dwarkesh Patel 5:22 And something being out there means what? Like, what- what- what is out there?

Sergey Levine 5:25 It means that there is a robot that does a thing that you actually care about-

Dwarkesh Patel 5:29 Yeah.

Sergey Levine 5:29 ... that, that you want done, and it does so competently enough to, like, actually do it for real for real people that want it done.

Dwarkesh Patel 5:36 We already have LLMs which are, like, broadly deployed and that hasn't resulted in some sort of, like, flywheel. Um-

Sergey Levine 5:42 That's right.

Dwarkesh Patel 5:43 ... like, at least not some obvious flywheel for the model companies where the... now Claude is, like, learning how to do every single job in the economy, or GPT is learning how to do every single job in the economy. So why did- why doesn't that flywheel work for LLMs?

Sergey Levine 5:55 Well, I think it's actually, uh... I think it's actually very close to working.

Dwarkesh Patel 5:59 Yeah.

Sergey Levine 5:59 And I- I- I am, like, 100% certain that (laughs) many, uh, organizations are working on exactly this.

Dwarkesh Patel 6:06 Right.

Sergey Levine 6:06 In fact, arguably, there is already a flywheel in the sense that-... may, not an automated flywheel, but a human in the loop flywheel-

Dwarkesh Patel 6:12 Right.

Sergey Levine 6:12 ... where everybody who's deploying an LLM is, of course, gonna look at what it's doing and it's gonna use that, uh, to then modify its behavior.

Dwarkesh Patel 6:19 Yeah.

Sergey Levine 6:19 Uh, it, it, it, it's complex because, uh, it comes back to this question of representations and figuring out the right way to derive, uh, supervision signals and ground those supervision signals in the behavior of the system so that it actually improves on what you want.

Dwarkesh Patel 6:35 Mm-hmm.

Sergey Levine 6:35 And I don't think that's, like, a profoundly impossible problem, it's just something where the details get, like, pretty gnarly and challenges with algorithms and stability become pretty complex, so it's just, i- i- it's something that's taken a- a while for, uh, the community collectively to get their hands around it.

Dwarkesh Patel 6:50 Do, do, do you think it'll be easier for robotics or just that, like, the s- the state of this kind of, uh, techniques to label data that you collect out in the world and use it as a reward will just... The, the sort of like, um, the whole wave will rise and robotics will rise as a result or is there some reason why you think robotics will be-

Sergey Levine 7:07 Um...

Dwarkesh Patel 7:07 ... will benefit more from this?

Sergey Levine 7:09 Yeah, I don't think there's, like, a profound reason why robotics is that different, but there are a few small differences that I think, uh, make things a little bit more manageable. So, uh, especially if you have a robot that's doing something i- in cooperation with people, wh- whether it's a person that's supervising it or-

Dwarkesh Patel 7:22 Yeah.

Sergey Levine 7:22 ... directing it, like, there are very natural sources of supervision and there's, like, there's a big incentive for the person to provide the assistance that will make things succeed.

Dwarkesh Patel 7:30 Mm-hmm.

Sergey Levine 7:30 Uh, there are a lot of dynamics where you can make mistakes and recover from those mistakes and then reflect back on what happened and avoid that mistake in the future.

Dwarkesh Patel 7:37 Mm-hmm.

Sergey Levine 7:38 And I think that when you're doing physical things in the real world, that kind of stuff just happens more often, uh, than it does if you're, like, an AI assistant answering a question. Like, if you answer a question and you just, like, answered it wrong, it's like-

Dwarkesh Patel 7:48 Mm...

Sergey Levine 7:49 Well, it's not like you can just, like, go back and, like, tweak a few things. Like, y- the person you told the answer to might not even know that it's wrong.

Dwarkesh Patel 7:54 Right.

Sergey Levine 7:55 Whereas if you're, like, folding the T-shirt and you messed up a little bit, like, yeah, it's pretty obvious. Like, you can reflect on that-

Dwarkesh Patel 7:59 Right.

Sergey Levine 7:59 ... figure out what happened and do it better next time.

Dwarkesh Patel 8:00 Yeah. So I, uh, okay, o- in one year we have robots which are, like, doing some useful things, maybe if you have some, like, relatively simple, like, loopy process, they can, they can do it for you, just, like, you gotta keep folding, like, thousands of boxes or something. But then there's some flywheel, dot, dot, dot, i- i- there's some machine which will, like, just run my house for me, uh, as well as a human housekeeper would. Um, what is the gap between this thing which will be deployed in a year that starts the flywheel and this thing which is like a fully autonomous housekeeper?

Sergey Levine 8:33 Well, I- I think it's actually not that different than what we've seen with, uh, LLMs in some ways, that it's a matter of scope. Like, if you think about coding assistants, right, like, initially, the best tools for coding, they could do, like, a little bit of completion.

Dwarkesh Patel 8:46 Yeah.

Sergey Levine 8:46 Like, you give them a function signature and they'll, like, try their best to type out, like, the whole function and they'll maybe, like, get half of it right.

Dwarkesh Patel 8:52 Yeah.

Sergey Levine 8:53 Uh, and as that stuff progresses, then you're willing to give these things a lot more agency so that, like, the very best coding assistants now, like, if you have, if you're doing something relatively formulaic, maybe it can, like, put together most of a, most of a PR for you for something, you know, fairly accessible.

Dwarkesh Patel 9:08 Right.

Sergey Levine 9:08 So I think it'll be the same thing, that it'll, that we'll see an increase in the scope that we're giving, that we're willing to give to the robots as they get better and better, where initially the scope might be, like, there is a particular thing you do, like, you're making the coffee or something, uh, whereas, a- a- as they get more capable, as their ability to have common sense and a broader repertoire of tasks increases, then we'll give them greater scope.

Dwarkesh Patel 9:28 Is there something-

Sergey Levine 9:28 Like, now you're running the whole coffee shop.

Dwarkesh Patel 9:30 Um, I get that there's a spectrum and I get that there won't be a specific moment that feels like-

Sergey Levine 9:34 Yeah.

Dwarkesh Patel 9:34 ... we've achieved it, but if you had to give a year in which, like, that, your median estimate of when that happens.

Sergey Levine 9:39 I mean, my, my sense there too is that this is a, probably a single digit thing rather than a double digit thing, but the reason it's so hard to really pin down is because as with, as with all research, it does depend on figuring out a few question marks. And I think my, my answer in terms of the nature of those question marks is I don't think these are things that require profoundly deeply different ideas, but it does require the right synthesis of the kinds of things, um, that we already know. And, you know, sometimes synthesis, to be clear, is just as difficult as coming up with, like, profoundly-

Dwarkesh Patel 10:10 Mm-hmm.

Sergey Levine 10:10 ... new stuff, right? Uh, so I, I think it's intellectually a very deep a- and profound problem and, and figuring that out is gonna be, like, very exciting. Uh, but it, uh, I think we kind of, like, know, like, roughly the puzzle pieces and it's something that we need to work on. And I think if we work on it and we're a bit lucky and everything kind of goes as planned, I think single digit is rea- reasonable.

Dwarkesh Patel 10:33 Okay. I'm, I'm just gonna do binary search until I get a year.

Sergey Levine 10:36 (laughs)

Dwarkesh Patel 10:37 Um, okay, so it's less than 10 years, so more than five years? Your median estimate. I know it's like a range.

Sergey Levine 10:41 I, I think, I think five is a good median, yeah.

Dwarkesh Patel 10:43 Okay. Five years. So if you can fully autonomously run a house, then I think you've, like, you can fully autonomously do most blue collar work, so your estimate is in five years it should be able to do most, like, blue collar work in the economy.

Sergey Levine 10:59 So I, I think there's a, there's a nuance here, uh, and, and the nuance is it becomes more obvious if we consider the analogy to the coding assistants, right?

Dwarkesh Patel 11:07 Yeah.

Sergey Levine 11:07 It's not like the, um, the nature of coding assistants today is that there's a switch that flips and suddenly instead of writing software, like, and suddenly, like, all software engineers get fired and everyone's using LLMs-

Dwarkesh Patel 11:19 Right.

Sergey Levine 11:19 ... for everything. Uh, and it, that actually makes a lot of sense that the biggest gain in productivity comes from experts, which is software engineers, whose effe- whose productivity is now augmented by these really powerful tools.

Dwarkesh Patel 11:34 Yeah. I mean, separate from the question of, uh, whether people will get fired or not, there, a, a different question is just, like, what will the economic impact be-

Sergey Levine 11:40 Mm-hmm.

Dwarkesh Patel 11:40 ... in five years?

Sergey Levine 11:41 Yeah.

Dwarkesh Patel 11:42 Um, the reason I'm curious about this is with LLMs, the relationship between the revenues for these models to their inherent... their seeming capability has been sort of mysterious in the sense that, like, you have something which feels like AGI, you can have a conversation with a really, like, is like, you know, like, f- passes a soaring test, it really feels like it can do all this knowledge work, it's obviously doing a bunch of coding, et cetera, but then the revenues for these AI companies are on, like, cumulatively on the order of, like, 20, $30 billion, um, per year, and that's so much less than...... all knowledge work, which is $30, $40 trillion. Um, so in five years, are we in a similar situation that LLMs are in now, or is it more like we have robots deployed everywhere and they're actually, like, doing a whole bunch of real work, et cetera?

Sergey Levine 12:28 It's a- it's a very subtle question. I think what it probably will come down to is this- is this question of scope, right? Like, uh, the reason that LLMs aren't doing all software engineering is because they're good within a certain scope, but there's limits to that.

Dwarkesh Patel 12:41 Right.

Sergey Levine 12:41 And those limits are increasing, to be clear, every year.

Dwarkesh Patel 12:44 Yeah.

Sergey Levine 12:44 Uh, and I think that there's no reason that- that we wouldn't see the same kind of thing with robots, that, um, the scope will have to start out small because there will be certain things that these things- that these systems can do very well and certain other things where more human oversight is really important. And the scope will grow and what that will translate into is increased productivity. And some of that productivity will come from, like, the robots themselves being valuable and some of it will come from the people using the robots-

Dwarkesh Patel 13:15 Right.

Sergey Levine 13:15 ... are now more productive in their- in their work.

Dwarkesh Patel 13:16 But there's so many things which increase productivity, just like wearing gloves increases productivity.

Sergey Levine 13:20 That's correct.

Dwarkesh Patel 13:20 Or, like, I don't know, uh, but then it's like you wanna understand something which, like, increases productivity 100 fold versus, like, uh, you know, wearing glasses inc- uh, or something which has, like, a small increase. So robots already ha- increase productivity for, uh, workers, right? Um, where LLMs are right now in terms of the share of knowledge work they can do, which is it s- I guess probably like 1/1000 of the knowledge work that happens in the economy LLMs are doing, at least in terms of revenue. Um, are you saying, like, that fraction will be possible for robots but for physical work in five years?

Sergey Levine 13:55 That's a very hard question to answer.

Dwarkesh Patel 13:57 Mm-hmm.

Sergey Levine 13:57 Uh, I think I'm probably not prepared to tell you what percentage of all labor work can be done by robots because I don't think right now off the cuff I have a sufficient-

Dwarkesh Patel 14:06 Sure.

Sergey Levine 14:06 ... understanding of, uh, what's involved in, uh-

Dwarkesh Patel 14:10 Yeah.

Sergey Levine 14:10 ... you know, that big of a cross-section of all physical labor. Uh, I think what I can tell you is this, that I think it's much easier to get effective systems rolled out gradually in a human in the loop setup. And again, I think this is, like, exactly what we've seen with- with coding systems, and I think we'll see the same thing with automation where, uh, basically robot plus human is much better than just human or just robot.

Dwarkesh Patel 14:35 Yeah.

Sergey Levine 14:35 Uh, and- and it's- that just, like, makes total sense. It also makes it much easier to get all the technology bootstrapped because when it's robot plus human, now there's a lot more potential for the robot to, like, actually learn on the job, acquire new skills. It's just like, you know, it-

Dwarkesh Patel 14:49 Because a human can label what's happening?

Sergey Levine 14:51 Also because the human can help, it can-

Dwarkesh Patel 14:53 Right.

Sergey Levine 14:53 ... the human c- can give hints. You know, let me tell you this story. Like, um, we, um, when we were working on the, uh, Pi o5 project.

Dwarkesh Patel 15:00 Yeah.

Sergey Levine 15:00 This was the paper that we released, uh, last April. W- we initially controlled our robots with tele-operation in a variety of different settings and then at- at some point, we actually realized that we can actually make significant headway once the model was good enough by supervising it not just with low level actions, but actually literally instructing it through language.

Dwarkesh Patel 15:21 Mm-hmm.

Sergey Levine 15:21 Now, you need a certain level of competence before you can do that, but once you have that level of competence, just standing there and telling the robot, "Okay, now- now pick up the cup."

Dwarkesh Patel 15:27 I see.

Sergey Levine 15:27 "Or put the cup in the sink. Uh, put the- the dish in the sink," just with words-

Dwarkesh Patel 15:32 Yeah.

Sergey Levine 15:32 ... already actually gives the robot information that it can use to get better.

Dwarkesh Patel 15:37 Right.

Sergey Levine 15:37 Now, imagine what this implies for the- the human plus robot dynamic. Like now, like basically learning is not, uh, f- for these systems, it's not just learning from real actions, it's also learning from words, eventually be learning from observing what people do, from the kind of natural feedback that you receive when you're doing a job together with somebody else. And, uh, this kind of- this is also the kind of stuff where the- the prior knowledge that comes from these big models is tremendously valuable.

Dwarkesh Patel 16:03 Yeah.

Sergey Levine 16:03 Because that lets you understand the- that interaction dynamic. So, um, I think that there's a lot of potential for these kind of human plus robot, uh, deployments to make the model better.

Dwarkesh Patel 16:13 Interesting. So I got to go to Label Box and see the robotics setup and try operating some of the robots myself.

Sergey Levine 16:19 So the thing is, like, these triggers, be very mindful of pressing them and don't do some, like, very fast movements.

Dwarkesh Patel 16:25 Yeah.

Sergey Levine 16:25 Keep it, like-

Dwarkesh Patel 16:26 Got it.

Sergey Levine 16:26 ... kind of slow.

Dwarkesh Patel 16:27 Do I need to keep holding it? Oh yeah.

Sergey Levine 16:29 Oh. Oh.

Dwarkesh Patel 16:29 Oh, sorry. Okay.

Sergey Levine 16:30 That's okay. And don't move it very fast because he can get hurt actually.

Dwarkesh Patel 16:33 Yeah, yeah, okay. Okay, so operating ended up being a bit harder than I anticipated, but I did get to see the Label Box team rip through a bunch of tasks. I also got to see the output data that labs actually have to use to train their robots and asked Manu, Label Box's CEO, about how all this is packaged together.

Narrator 16:54 So what you're looking at is actually the final output that is then delivered to the labs, which then they use to train the models. And so you can see on the left the visualization of the movements of the robot, including its 3D model and so forth, and on the right, you see all the camera streams synchronized with the configuration.

Dwarkesh Patel 17:13 Label Box can get you millions of episodes of robotics data for every single robotics platform and subtask that you want to train on. And if you reach out through labelbox.com/fourcash, Manu will be very happy with me today.

Dwarkesh Patel 17:25 (laughs)

Sergey Levine 17:26 (laughs)

Dwarkesh Patel 17:26 In terms of robotics progress, why won't it be like self-driving cars where we, you know, m- it's been more than 10 years since Google launched its, um... Wasn't it 2009 that they launched the self-driving car initiative? And then I remember when I was a teenager, like, watching demos where we would go buy a Taco Bell, uh, um, and drive back, and only now do we have them actually deployed. And even then, you know, they make- make mistakes, et cetera, and so maybe it'll be many more years before most of the cars are self-driving. So why won't robotics... You know, you're saying five years to this, like, quite robust thing, but actually it'll just feel like 20 years of just, like, once we get the cool demo in five years, then it'll be another 10 years before, like, we have the Waymo or when the Tesla FSD working.

Sergey Levine 18:14 Yeah, that's a really good question. So one of the big things, uh, that is different now than it was in 2009, uh, actually has to do with-... the technology for machine learning systems that understand the world around them. Uh, principally, for autonomous driving, this is perception. Uh, for robots, it can mean a few other things as well. Uh, and perception certainly was not in a good place in 2009. The trouble with perception is that it's one of those things where you can nail a really good demo with a somewhat engineered system, but hit a brick wall when you try to generalize it. Now, at this point in 2025, we have much better technology for generalizable and robust perception systems, and more generally generalizable and robust systems for understanding the world around us. Like, when you say that the system is scalable, in machine learning, scalable really means generalizable. Um, so that gives us a much better starting point, uh, today. So that's not an argument about robotics being easier than autonomous driving. It's just a- an argument for 2025 being a better year than 2009.

Dwarkesh Patel 19:13 Mm-hmm.

Sergey Levine 19:14 But there's also other things about robotics that are a bit different than driving. Like, in some ways, robotic manipulation is a much, much harder problem, but in other ways, it's a- it's a problem space where it's easier to get rolling, to start that flywheel with a more limited scope. Um, so to give you an example, if you're learning how to drive, you'd probably be pretty crazy to learn how to drive on your own without somebody helping you. Uh, like, you- you would not trust your- your- your teenage, uh, child to learn to drive just on their own, just drop them in the car and say, like, "Go for it."

Dwarkesh Patel 19:46 Mm-hmm.

Sergey Levine 19:46 Uh, and that's like a, you know, a 16-year-old who's had, uh, a significant amount of time to learn about the world. You would never-

Dwarkesh Patel 19:53 Right.

Sergey Levine 19:53 ... even dream of putting a five-year-old in a car-

Dwarkesh Patel 19:55 Yeah.

Sergey Levine 19:55 ... and tell him to get started. But if you want somebody to, like, clean the dishes-

Dwarkesh Patel 19:59 Yeah.

Sergey Levine 19:59 ... like, dishes can break too, but you would probably be okay with a child trying to do the dishes, uh, without somebody constantly, like, you know- (laughs)

Dwarkesh Patel 20:07 Mm-hmm.

Sergey Levine 20:07 ... sitting next to them with a- (laughs) with a- with a brake, so to speak. So for a lot of tasks that we want to do with robotic manipulation, there's potential to make mistakes and correct those mistakes, and when you make a mistake and correct it, well, first, you've- you've achieved the task because you've corrected, but you've also gained knowledge that allows you to avoid that mistake in the future. With driving, because of the dynamics of how it's set up, it's very hard to make a mistake, correct it, and then learn from it because the mistakes themselves have significant ramifications. Um, now, not all manipulation tasks are like that. There are truly some, like, very, uh, safety critical stuff, and this is where the next thing comes in, which is common sense. Uh, common sense meaning the ability to make inferences about what might happen, uh, that are reasonable guesses, but that do not require you to experience that mistake and le- and learn from it in advance.

Dwarkesh Patel 20:56 Right.

Sergey Levine 20:57 That's tremendously important, and that's something that we basically had no idea how to do, uh, about five years ago.

Dwarkesh Patel 21:02 Mm-hmm.

Sergey Levine 21:03 But now, uh, you- w- we can actually use LMS and VLMs, ask them questions, and they will make reasonable guesses. Like, they will not give you expert behavior, but you can say, like, "Hey, there's a sign that says slippery floor. Like, what's going to happen when I walk over that?" It's kind of pretty obvious.

Dwarkesh Patel 21:18 Right.

Sergey Levine 21:18 Right? Uh, and no autonomous car in 2009 would have been able to answer that question. (laughs) So common sense plus the ability to make mistakes and correct those mistakes, like, that's sounding like o- o- o- o- an awful lot like what a- what a person does when they're trying to learn something.

Dwarkesh Patel 21:30 Mm-hmm.

Sergey Levine 21:30 All of that doesn't make robotic manipulation easy necessarily, but it allows us to get started with a smaller scope and then grow from there.

Dwarkesh Patel 21:38 So for years, m- m- using, I mean, not since 2009, but we've had lots of video data, language data, and transformers for five, seven, eight years.

Sergey Levine 21:50 Mm-hmm.

Dwarkesh Patel 21:50 And lots of companies have tried to build transformer, uh, based robots with lots of training data, including Google, Meta, et cetera, and what is the reason that they've been hitting roadblocks? What has changed now?

Sergey Levine 22:05 Yeah, that's a really good question. So I'll start out with, uh, maybe, uh, a slight modification to your comment is I think they- they've made a lot of progress.

Dwarkesh Patel 22:14 Mm-hmm.

Sergey Levine 22:14 And in some ways, a lot of the work that, uh, we're doing now at Physical Intelligence is built on the backs of lots of other great work that was done, uh, for example, at, uh, at Google. Like, many of us were actually at Google before.

Dwarkesh Patel 22:25 Right.

Sergey Levine 22:25 We were involved in some of that work. Some of it is work that we're drawing on that others did. So there's definitely, like, been a lot of progress there. But, um, to make robotic foundation models really work, it's not just a laboratory science, uh, kind of experiment. It's also, uh, it also requires kind of industrial scale, uh, building effort. Like, it's- it's like- it's more like the Apollo program than it is like a- a science experiment. And, uh, the excellent research that was done in the past in industrial research labs, and I know I was involved in much of that, uh, was very much framed as a fundamental research effort.

Dwarkesh Patel 23:04 Mm-hmm.

Sergey Levine 23:05 And that's good. Like, the fundamental research is really important, but it's not enough by itself. Uh, you need the fundamental research, and you also need the impetus to make it real, and make it real means, like, actually put the robots out there. Get data that is representative of the kind of tasks that they want to-

Dwarkesh Patel 23:20 Mm-hmm.

Sergey Levine 23:20 ... that they need to do in the real world. Get that data at scale, build out the systems, get all that stuff right. Uh, and that requires a degree of focus, uh, a- a singular focus on really nailing the robotic foundation model for its own sake, not, uh, just as a way to do more science, not just as a way to, like, publish a paper, uh, and not just as- as a way to kind of, like, uh, you know, have a- a- a- a research lab.

Dwarkesh Patel 23:43 What is preventing you now from continuing, uh, scaling that data even more, um, if data is the big bottleneck, why can't you just increase the size of your, uh, office 100X, have 100X more operators who are operating these robots and collecting more data? Yeah, why not ramp it up immediately 100X more?

Sergey Levine 24:03 Yeah, that's a really good question. So the challenge here is in understanding which axes of scale contributes to which axis of capability.

Dwarkesh Patel 24:13 Mm-hmm.

Sergey Levine 24:13 Um, so if we want to expand capability horizontally, meaning, like, the robot knows how to do 10 things now, and I'd like it to do 100 things later, um, you know, that- that- that can be addressed by just directly horizontally scaling what we already have. But-We want to get robots to a level of capability where they can do practical, practically useful things in the real world, and that requires, uh, expanding a- along other axes too. It requires, for example, getting to very high robustness. It requires getting them to perform tasks very efficiently, quickly. Uh, it requires them to recognize edge cases and respond intelligently. And those things, I think, could also be addressed with scaling, but we have to identify the right axes for that, which means figuring out what kind of data to collect, what settings to collect it in, what kind of methods consume that data, how those methods work.

Dwarkesh Patel 24:58 Right.

Sergey Levine 24:58 Uh, so answering those questions more thoroughly will give us a greater clarity on the axes, uh, on- on the- on those, uh, dependent variables, on the things that we need to scale, and we don't fully know right- right now what that will look like. I think we'll figure it out pretty soon, it's something we're working on actively, but we want to really get that right so that when we do scale it up, it'll directly translate into capabilities that are very relevant to practical use.

Dwarkesh Patel 25:25 Just to give an order of magnitude, um, how does the amount of data we- you have collected compare to internet scale pre-training data? And I know it's hard to do, like, a token-

Sergey Levine 25:33 Yeah.

Dwarkesh Patel 25:33 ... by token count because, uh, yeah. What is- how does video information compare to internet information, etc.? But, like-

Sergey Levine 25:39 Yeah. It-

Dwarkesh Patel 25:39 Using your reasonable estimates, what fraction of-

Sergey Levine 25:42 That's right. It- it's very hard to do because, um, robotic experience, uh, consists of time steps that are very correlated with each other.

Dwarkesh Patel 25:48 Yeah.

Sergey Levine 25:49 Uh, so, like, the raw, like, byte representation is enormous-

Dwarkesh Patel 25:53 That's right.

Sergey Levine 25:53 ... but probably the information density is comparatively low.

Dwarkesh Patel 25:55 Yeah.

Sergey Levine 25:56 Um, maybe a better comparison is to, um, the datasets that are used for, uh, for multimodal training.

Dwarkesh Patel 26:01 Yeah.

Sergey Levine 26:02 And there, it's, uh, I believe last time we did that count, it was like between one and two orders of magnitude.

Dwarkesh Patel 26:08 The vision you have of- of robotics will not be possible un- until you collect, like what? 100X, 1,000X more data?

Sergey Levine 26:16 Well, that's the thing, that- that we don't know that. Um, uh, it- it's certainly very reasonable to infer that, like, you know, robotics is a tough problem, uh, and probably it requires, you know, as much experience as the language stuff. But because we don't know the answer to that, to me, a much more useful way to think about it is not how much data do we need to get before we're fully done, but how much data do we need to get before we can get started?

Dwarkesh Patel 26:40 Mm-hmm.

Sergey Levine 26:40 Meaning, before we can get, uh, a data flywheel that represents a self-sustaining, uh, and ever-growing data collection or something.

Dwarkesh Patel 26:48 When you say self-sustaining, uh, this is just like learning on the job or do you have something else in mind?

Sergey Levine 26:52 Learning on the job or, uh, acquiring data in a way that the process of acquisition of that data itself is useful and valuable.

Dwarkesh Patel 27:01 I see. Like, just some kind of RL.

Sergey Levine 27:03 Like doing something, uh, like actually real. Yeah. I- I mean, ideally, I would like it to be RL because you can get away with the robot acting autonomously.

Dwarkesh Patel 27:10 Right.

Sergey Levine 27:10 Which is, uh, easier.

Dwarkesh Patel 27:12 So-

Sergey Levine 27:12 But the- that's- that's not out of the question that you can have mixed autonomy, uh, you can- you know, as- as I mentioned before, robots can learn from all sorts of other signals. Uh, I described how we can have a robot that learns from a person talking to it. So there- there's a lot of middle ground in between fully teleoperated robots

Sergey Levine 27:28 and fully autonomous robots.

Dwarkesh Patel 27:29 Yeah. Okay, and how does the PI model work?

Sergey Levine 27:31 Yeah. So, the current model that we have, uh, basically is a vision-language model that has been adapted for motor control. So, uh, to give you a li- a little bit of like a fanciful brain analogy, a VLM, a vision-language model, is basically an LLM that has had a- a little, like, pseudo visual cortex grafted to it, a vision encoder. All right? So our models, they have a vision encoder, but they also have an action expert, an action, uh, decoder essentially. So, it has, like, a little visual cortex and notionally a little motor cortex. And the way that the model actually makes decisions is it reads in the sensory information from the robot, it does some internal processing, and that could involve actually, uh, outputting intermediate steps, like you might tell it to clean up the kitchen and it might think to itself, like, "Hey, to clean up the kitchen, I need to pick up the dish and I need to pick up the sponge and I need to put th- this and this." And then eventually, it works its way through that chain of thought generation down to the action expert which actually produces continuous actions. And that- that has to be a different module because the actions are continuous, they're high frequency, so they ha- have a different data format than, uh, text tokens. But structurally, it's still, uh, an end-to-end transformer and roughly speaking, uh, technically, it- it corresponds to a kinda mixture of experts architecture.

Dwarkesh Patel 28:42 Hmm. And, like, what is actually happening is that it's, like, deco- like, it's, like, predicting I should do X thing, then it's like there's an image token, then some action tokens, like, what it actually ends up doing, and then more image, more, uh, text description, more a- more action tokens. Basically, I'm, like, looking at what- what stream is going on.

Sergey Levine 28:59 Yeah. That- that's right. Uh, with the- with the exception of the actions are actually not represented as discrete tokens, it's- it actually uses a flow matching kind of diffusion-

Dwarkesh Patel 29:06 Mm-hmm.

Sergey Levine 29:06 ... because they're continuous and you need to be very precise with your actions for

Narrator 29:09 ... textures to show.

Dwarkesh Patel 29:09 Right. I find it super interesting that, so you are u- I think you're using the open source GEMMA model which is like Google's, uh, LLM, uh, that they released open source and then adding this action expert on top. And I find it super interesting that the progress in different areas of AI is just based on s- this- not only the same techniques, but literally the same model, that you can just use an open source LLM and then add this action expert on top. It is notable that, like, you naively might think that, oh, there's this, like, separate era of research is robotics and there's a separate era of research called LLMs and, uh, n- natural language processing, and no. It's like, it's literally the same. It's like the considerations are the same, the, um, the architectures are the same, even the weights are the same. I know you do more training on top of these model- open source models, but, y- that I find super interesting.

Sergey Levine 29:59 Yeah. So one theme here that, like, uh, I think is important to keep in mind is that the reason that th- those building blocks are so valuable is because the AI community has gotten a lot better at leveraging prior knowledge.

Dwarkesh Patel 30:11 Mm-hmm.

Sergey Levine 30:12 And a lot of what we're getting from the pretrained LLMs and VLMs is prior knowledge about the world, and it's kind of like, it's a little bit abstracted knowledge, like, you know, you can identify objects, you can figure out, uh, like, roughly where things are in an image, that sort of thing. But I think if- if I had to, like, summarize in one sentence-... the big benefit that recent innovations in AI give to robotics is really that prior, the ability to leverage prior knowledge. And, uh, I, I think the fact that the model is the same model, that's like, that's kind of always been the case in deep learning.

Dwarkesh Patel 30:42 Mm-hmm.

Sergey Levine 30:42 But it's, it's that ability to pull in that prior knowledge, that abstract knowledge that has, that can come from many different sources. That's, uh, really powerful.

Dwarkesh Patel 30:48 Yeah. Today, I'm here with Mark, who is a senior researcher at Hudson River Trading. He has prepared for us a big data set of market prices and historical market data, and we're gonna try to figure out what's going on and whether we can f- predict future prices from historical market data. Mark, let's dig in.

Narrator 31:05 Happy to do it.

Dwarkesh Patel 31:06 So, it sounds like the first fun thing to do is probably to start looking at what an order book actually looks like.

Narrator 31:12 Yeah. I think so.

Dwarkesh Patel 31:13 Mm-hmm.

Narrator 31:13 So I've given you, like, real order book data that is snapshots of the top five levels of the order book, both in the bid and ask side for a couple of different tech stocks: NVIDIA, Tesla, AMD, et cetera.

Dwarkesh Patel 31:26 What is the shape of the prediction? Are, are, are we predicting-

Narrator 31:29 Why don't you, uh, take the dataframe, look at its Y values and just kind of, like-

Dwarkesh Patel 31:32 Oh, that's great.

Narrator 31:33 ... histogram it?

Dwarkesh Patel 31:33 They are centered at zero.

Narrator 31:34 They are roughly centered at zero.

Dwarkesh Patel 31:35 Yeah. But target of what exactly?

Narrator 31:37 So these things are changes in the mid-price from, like, now to some short period-

Dwarkesh Patel 31:41 Mm-hmm.

Narrator 31:42 ... of time in the future.

Dwarkesh Patel 31:42 Thi- this is actually quite interesting, just like a mo- mystery to solve.

Narrator 31:45 And each one of these can be, like, a sizeable chunk of time for a re- researcher.

Dwarkesh Patel 31:48 If this sounds interesting to you, you should consider working at Hudson River Trading. Mark, where can people learn more?

Narrator 31:54 Uh, they can learn more at hudson-trading.com/dwarkish.

Dwarkesh Patel 31:57 Amazing. I was talking to th- um, this researcher, uh, uh, Sander at GDM, and he, he works on video and audio models, and he made the interesting point that the reason, in his view, we aren't seeing that much transfer learning between different modalities, that is to say, like, training a language model on video and images doesn't seem to necessarily make it that much better at textual questions and, um, tasks, is that images are represented at a different semantic level than text. And so his argument is that text has this high-level semantic representation within the model, whereas images and videos are just, like, compressed pixels. There's not really a semantic-

Narrator 32:40 Mm-hmm.

Dwarkesh Patel 32:40 Uh, w- when they're embedded, there's not, they're n- they don't represent some, like, high-level semantic information. They're just, like, compressed pixels. Um, and therefore, there's- th- there's no transfer learning at the level at which they are going through the model.

Narrator 32:53 Yeah.

Dwarkesh Patel 32:53 And obviously this is super relevant to the work you're doing because y- your hope is that by training the model both on the visual data that the robot sees, visual data generally, maybe even in, like, from YouTube or whatever eventually, plus, like, language information, plus action information from the robot itself, some- now all of this together will, like, make it r- generally robust. Um, and then you had a really interesting blog post about, like, why video models aren't as robust as language models. Um, sorry, this is not a super well-formed question. I just wanted to hear you react to some of that.

Sergey Levine 33:22 (laughs) People are like, "What's up with that?"

Dwarkesh Patel 33:23 Yeah, yeah.

Sergey Levine 33:24 Um, yeah. So I, I have, uh, maybe two things I can say there. I have some, like, bad news and some good news.

Dwarkesh Patel 33:29 Yeah.

Sergey Levine 33:30 So the bad news is what you're saying is really getting at the core of a long-running challenge with video and, and image generation models.

Dwarkesh Patel 33:44 Yeah.

Sergey Levine 33:44 Like, in some ways, the idea of getting intelligent systems by predicting video is even older than the idea of getting intelligent systems by predicting text.

Dwarkesh Patel 33:55 Mm-hmm.

Sergey Levine 33:55 But the text stuff turned into use- practically useful things earlier than the video stuff did.

Dwarkesh Patel 34:02 Mm-hmm.

Sergey Levine 34:02 I mean, the video stuff is great. Like, you can generate cool videos, and I think that the work there that's been done recently is, is like amazing. But it's n- But it's not like just generating videos and images has already resulted in systems that have this kind of, like, deep understanding of the world where you can, like, ask them to, like, do stuff beyond just generating more images and-

Dwarkesh Patel 34:19 Yeah.

Sergey Levine 34:19 ... and videos. Whereas with language, c- clearly it has. And I think that this, this point about representations is really key to it. One way we can think about it is this, that, um, if you ha- um, imagine pointing a camera outside this building. There's the sky. There's the clouds are moving around, uh, the water, cars driving around, people. If you wanna predict everything that'll happen in the future, you can do so in many different ways. You can say, "Okay, there's people around, so let me get really good at understanding, like, the psychology of how people behave in crowds-

Dwarkesh Patel 34:46 Mm-hmm.

Sergey Levine 34:46 ... and predict the pedestrians." But you could also say, like, "Well, there's clouds moving around. Let me, like, understand everything about water molecules and, uh, ice particles in the air." And you could go super deep on that.

Dwarkesh Patel 34:55 Right.

Sergey Levine 34:55 Like, if you wanna, like, fully understand, like, all, you know, down to the subatomic level everything that's going on, like, as a person, you could spend, like, decades just thinking about that, and you'll never even get to the pedestrians or the water, right?

Dwarkesh Patel 35:05 Mm-hmm.

Sergey Levine 35:06 So, if you want to really predict everything that's going on in that scene, there's just so much stuff that even if you're doing a really great job and capturing, like, 100% of something, by the time you get to everything else, like, you know, ages will have passed.

Dwarkesh Patel 35:19 Right.

Sergey Levine 35:19 Whereas with text, it's already been abstracted into those bits that we as humans care about. So the representations are already there, and they're not just good representations. They actually, like, focus in on what really matters. Okay, so that's, that's the bad news. Here's the good news. The good news is that we don't have to just get everything out of, like, pointing a camera outside this building, because when you have a robot, that robot is actually trying to do a job. So it has-

Dwarkesh Patel 35:44 (coughs)

Sergey Levine 35:44 ... uh, a purpose.

Dwarkesh Patel 35:45 Yeah.

Sergey Levine 35:45 Uh, and its perception is in service to fulfilling that purpose. And that is, like, a really great, uh, focusing factor. We know that for people, this really matters. Like, literally what you see is affected by what you're trying to do. Uh, like, th- there's been no shortage of psychology experiments showing that people have, like, almost a shocking degree of tunnel vision, uh, where they will, like, literally not see things right in front of their eyes-

Dwarkesh Patel 36:06 Yeah.

Sergey Levine 36:06 ... if it's not relevant to what they're trying to achieve.

Dwarkesh Patel 36:08 Yeah.

Sergey Levine 36:08 And that is tremendously powerful. Like, there must be a reason why people do that, because you know, certainly if you're in- out in the jungle, seeing more is better than seeing less.

Dwarkesh Patel 36:15 Right.

Sergey Levine 36:15 So if you have that powerful focusing mechanism, it must be darn important-

Dwarkesh Patel 36:18 Yeah.

Sergey Levine 36:18 ... for getting it to achieve your goal. And I think robots will have that focusing mechanism because they're trying to achieve a goal.

Dwarkesh Patel 36:23 By the way, the fact that video models aren't as robust, is that bearish for robotics because it will... s- so much of the data you will have to use will not... I, I guess some of... You're saying a lot of it will be labeled, but like-Ideally, you just wanna be able to, like, throw all of, e- everything on YouTube, every video we have ever recorded, and have it learn how the physical world works and how to, like, move about, et cetera, just see humans performing tasks and learn from that. But if, yeah, I guess you're saying, like, it's hard to learn just from that, and it actually, like, needs to practice the task itself.

Sergey Levine 36:56 Well, let me put it this way. Like, let's say that I, um, gave you lots of, uh, videotapes or lots of recordings of different sporting events-

Dwarkesh Patel 37:05 Yeah.

Sergey Levine 37:05 ... and gave you a year to just watch sports. Uh, and then after that year, I told you, "Okay, now your job, you're gonna be playing tennis."

Dwarkesh Patel 37:11 Yeah.

Sergey Levine 37:11 (laughs) Okay, that's like, that's pretty, pretty dumb, right? Whereas if I told you first, like, "You're gonna be playing tennis," and then I, e- and then I let you-

Dwarkesh Patel 37:18 Right.

Sergey Levine 37:18 ... study up, right? Like, now you, you really know what, what you're looking for.

Dwarkesh Patel 37:21 Right.

Sergey Levine 37:21 So, I think that actually, like there's, there's a very real challenge here. I don't wanna understate the challenge, but I do think that there's also a lot of potential for foundation models that are embodied that, uh, learn from interaction, from controlling robotic systems, to actually be better at absorbing the other data sources because they know what they're trying to do. Uh, I don't think that that by itself is like a silver bullet. I don't think it solves everything, but I think that, uh, it does help a lot, and I think that, uh, we've already seen the beginnings of that where we can see that, uh, including web data in training for robots-

Dwarkesh Patel 37:54 Yeah.

Sergey Levine 37:55 ... really does help, uh, with generalization, and I, I actually have the suspicion that in the long run, it'll make it easier to use those sources of data that have been tricky to use up until now.

Dwarkesh Patel 38:03 Uh, uh, famously, LLMs have all these emergent capabilities that were never engineered in because somewhere in internet text is the data to train and to give it the knowledge to do a certain kind of thing. With robots, it seems like you are collecting all the data manually, so there won't be this mysterious new capability that, like, is somewhere in the dataset that you haven't purposefully collected, which seems like it should make it even harder to then have w- uh, robust, out of distribution kind of, uh, capabilities. And so I wonder if the trek over the next five, ten years will just be, like, each sub-task, you have to give it thousands of episodes, and then it's very hard to actually automate much work just by doing sub-tasks. So, if you look, think about what a barista does, what a waiter does, um, what a chef does, um, very little bit involves just, like, sitting at one station and, like, doing stuff, right? Just, like, you gotta move around, you gotta restock, you gotta, um, fix the machine or, um, et cetera, um, go between, like, the counter and, uh, the cashier and the machine, et cetera. So yeah, will th- will it just be, like, will there just be this long tail of things that you gotta keep, skills you gotta keep, like adding episodes for manually, then labeling and seeing how well they did, et cetera? Or is there some reason to think that it will progress more generally than that?

Sergey Levine 39:23 Yeah. So there's a subtlety here. Emergent capabilities don't just come from the fact that internet data has a lot of stuff in it. They also come from the fact that generalization, once it reaches a certain level, becomes compositional.

Dwarkesh Patel 39:36 Mm.

Sergey Levine 39:37 Uh, there is a (laughs) a cute example that, uh, one of my students, uh, really liked to use, uh, in some of his presentations which is, um, uh, you know what I- international phonetic alphabet is?

Dwarkesh Patel 39:49 No.

Sergey Levine 39:49 IPA? So if you look in a dictionary, uh, they'll have the pronunciation of a word and written in, like, kind of funny letters. That's basically international-

Dwarkesh Patel 39:56 Mm-hmm.

Sergey Levine 39:56 ... phonetic alphabet. So it's, it, it's an alphabet that is pretty much exclusively used for writing down pronunciations of individual words in dictionaries.

Dwarkesh Patel 40:03 Mm-hmm.

Sergey Levine 40:03 And you can ask an LLM to write you a recipe, uh, for like making some meal in international phonetic alphabet-

Dwarkesh Patel 40:10 Mm-hmm.

Sergey Levine 40:10 ... and it will do it.

Dwarkesh Patel 40:11 Right.

Sergey Levine 40:11 And that's like, like, holy crap. Like, that is definitely not something that has, uh, that it has ever seen because IPA is only ever used for writing down-

Dwarkesh Patel 40:18 Right.

Sergey Levine 40:19 ... pronunciations of individual words. So that's, that's, that's compositional generalization.

Dwarkesh Patel 40:22 Yeah.

Sergey Levine 40:22 It's putting together things you've seen like that in new ways. And it's like, you know, arguably, there's nothing, like, profoundly new here because, like, yes, you've seen different words written that way, but you've figured out that now you can compose the words in this other language the same way-

Dwarkesh Patel 40:34 Yeah.

Sergey Levine 40:34 ... uh, that you've composed words, uh, in English. So, um, that's actually where the emergent capabilities come from, and because of this, in principle, if we have a sufficient diversity of behaviors, the model should figure out that those behaviors can be composed in new ways, uh, as the situation calls for it. And we've actually seen things even with our current models which, you know, I should say that I think they're, in the grand scheme of things, like looking back five years from now, we'll probably think that these are tiny in scale. But we've already seen what I would call emergent capabilities. When we were playing around with some of our laundry folding policies, actually we discovered this by accident. Uh, the robot accidentally picked up two T-shirts out of the bin instead of one. Starts folding the first one, the other one gets in the way, picks up the other one, throws it back in the bin.

Dwarkesh Patel 41:17 Mm.

Sergey Levine 41:17 And we're like, "We, we didn't know, we didn't know it would do that." Like-

Dwarkesh Patel 41:20 Yeah.

Sergey Levine 41:20 ... holy crap. And then we tried to play around with it, and it's like, yep, it does that every time. Like, you can drop in, you know, it's, it's doing its work, drop something else on the table, just pick it up, put it back.

Dwarkesh Patel 41:27 Ah.

Sergey Levine 41:28 Right? Okay, that's cool. Uh, shopping bag, it starts putting things in the shopping bag. The shopping bag tips over, it picks it back, back up, and stands it upright.

Dwarkesh Patel 41:35 Mm.

Sergey Levine 41:35 We didn't, we didn't tell anybody-

Dwarkesh Patel 41:36 Yeah.

Sergey Levine 41:36 ... to collect data for that. I'm sure somebody accidentally at some point or may- maybe intentionally picked up the shopping bag, but it's just, you have this kind of compositionality that emerges when you do learning at scale, and that's really, uh, where all these remarkable capabilities come from.

Dwarkesh Patel 41:50 Yeah.

Sergey Levine 41:50 And now you put that together with language, you put that together with, uh, all sorts of chain of thought reasoning, and there's a lot of potential for the model to compose things in new ways.

Dwarkesh Patel 41:58 Right. I had an example like this when I got a tour of the, um, robots, by the way, at your, um, office. So, it was folding shorts, and I don't know if there was an episode like this in the, um, in the training set, but just for fun, I, like, took one of the shorts and, like, uh, turned it inside out.

Sergey Levine 42:14 Mm-hmm.

Dwarkesh Patel 42:15 And then it was able to understand that it first needed to get... So first of all, the grippers are just like this, like two-

Sergey Levine 42:24 Yes. (laughs)

Dwarkesh Patel 42:24 ... two limbs, uh, or just like opposable finger and thumb-like thing. And, um, it's actually shocking how much you can do with just that. Yeah, it understood that I first needed to fold it inside out before folding it correctly. I mean, what's surp- especially surprising about that is...It seems like this model only has, like, one second of context. So as compared to these language models which can often, like, see the entire code base and they're, like, observing hundreds of thousands of tokens and thinking about them before outputting, and they're observing their own train of thought for thousands of tokens before making a plan about how to code something up, your model is like seeing one image of, like, what happened in the last second, and it vaguely knows, like, it's supposed to fold this short-

Sergey Levine 43:04 Mm-hmm.

Dwarkesh Patel 43:05 Um, and it's seeing, like, the image of what's happened in the last second, and I guess it works. It's, like, crazy that, like, no... It will just see the last thing that happened and then keep executing on the plan, so fold it inside out, then fold it correctly. But it's shocking that a second of context is enough to execute on a minute-long task. W- yeah, I'm curious why you made that choice in the first place.

Sergey Levine 43:28 Mm.

Dwarkesh Patel 43:28 And why it's possible to actually do tasks... Like, if a human could only, like, think, had like a second of memory-

Sergey Levine 43:33 Yeah.

Dwarkesh Patel 43:33 ... and had to, like, do physical work, I feel like that would just be impossible.

Sergey Levine 43:35 Yeah. I mean, it, it's not that there's something good about having less memory, to be clear.

Dwarkesh Patel 43:39 Yeah.

Sergey Levine 43:39 Like, I think that, uh, adding memory, adding longer context, all that stuff, adding higher resolution images, I think those things will make them all better.

Dwarkesh Patel 43:47 Yeah.

Sergey Levine 43:47 But the reason why it's not the most important thing for the kind of skills that you saw when you visited us, it... At some level, I think it comes back to Moravec's paradox. So Moravec's paradox is basically the, the... It- it's like, you know, if you know one thing about... If you wanna know one thing about robotics, it's like that's, that's the thing. Moravec's paradox says that basically, uh, in AI, the easy things are hard and the hard things are easy, meaning, like, the things that we take for granted, like picking up objects, seeing a- uh, you know, perceiving the world, all that stuff, those are all the hard problems in AI. And the things that we find challenging, like playing chess and doing calculus, actually are often the easier problems.

Dwarkesh Patel 44:25 Mm-hmm.

Sergey Levine 44:25 And I think this memory stuff is actually Moravec's paradox in disguise, where we think that the cognitively demanding tasks that we do that we find hard, that kind of cause us to think, like, "Oh, man, I'm sweating. I'm working so hard," those are the ones that require us to keep lots of stuff in memory, lots of stuff in our minds. Like, if you're solving some big math problem, if you're having a, uh-

Dwarkesh Patel 44:45 Yeah.

Sergey Levine 44:45 ... a complicated technical conversation on a podcast-

Dwarkesh Patel 44:47 Yeah.

Sergey Levine 44:47 ... like, those are things we have to keep all those pieces, all those puzzle pieces in your head. If you're doing a well-rehearsed task, if you are an Olympic swimmer and you're swimming with perfect form and you're, like, right there in the zone-

Dwarkesh Patel 45:00 Mm-hmm.

Sergey Levine 45:01 ... like, people even say, like, "It's in the moment." It's in the moment, right?

Dwarkesh Patel 45:04 Yeah.

Sergey Levine 45:04 Like, it's like you've practiced it so much, you, you, you've baked it into your neural network in your brain-

Dwarkesh Patel 45:09 Yeah.

Sergey Levine 45:10 ... that, uh, you don't have to think carefully about keeping all that context.

Dwarkesh Patel 45:13 Yeah.

Sergey Levine 45:14 Right? So it really is just Moravec's mar- M- Moravec's paradox manifesting itself, but that doesn't mean that we don't need the memory. It just means that if we want to match the level of dexterity f- and physical proficiency that people have, there's other things we should get right first, and then gradually go up that stack into the more cognitively demanding areas, into reasoning, into context, into planning, all that kind of stuff.

Dwarkesh Patel 45:36 Mm-hmm.

Sergey Levine 45:36 And that stuff will be important

Sergey Levine 45:37 too.

Dwarkesh Patel 45:37 And how physically will... So you have, you have this, like, trilemma. You have three different things which all take more compute during inference that you wanna opt, uh, you wanna increase at the same time. You have the inference speed, and so humans are processing 24 frames a second or whatever it is. That we're just like... We can react to things extremely fast. Then you have the context length, and for, I think, the kind of robot which is just, like, cleaning up your house, I think it has to kind... It has to be aware of, like, things that happened minutes ago or hours ago and how that influences its plan about the next task it's doing. And then you have the model size, and espe- I guess at least with LLMs we've seen that there's gains from increasing the amount of, uh, parameters. And I think currently you have 100 millisecond, uh, inference speeds. You have a second long context, and then the model is what? A co- couple billion parameters? How many?

Sergey Levine 46:34 Mm-hmm.

Dwarkesh Patel 46:34 Okay. And so each of these, at least two of them are many orders of magnitude smaller than what seems to be the human equivalent, right? Like the model... If a human brain has like trillions of parameters and this has like two billion parameters, and then if humans are processing at least as fast as a model, like, uh, uh, actually a decent bit faster, and we have hours of context

Narrator 46:55 (clears throat)

Dwarkesh Patel 46:56 ... depends on how you define human context, but hours of context, minutes of context.

Sergey Levine 46:59 Sometimes decades of context.

Dwarkesh Patel 47:00 Yeah, exactly. So you have to have many order of magnitude improvements across all this thing, all of these three things which seem to oppose each other, or, like, increasing one reduces the amount of, um, r- reduces the amount of compute you can dedicate towards the other one in inference. So how are we gonna, yeah, how are we gonna solve this? (laughs)

Sergey Levine 47:19 Yeah. Uh, well, that's a very big question.

Dwarkesh Patel 47:22 (laughs)

Sergey Levine 47:22 Um, yeah. Let- let's, let's try to unpack this a little bit. I think there's, there's a lot going on in there. One thing that, um, I would say is a really interesting technical problem, and I think that it's something where we'll see perhaps a lot of really interesting innovation over the next few years, is the question of representation for context.

Dwarkesh Patel 47:43 Mm-hmm.

Sergey Levine 47:44 So, um, if you imagine the... Like s- some of the examples you gave, like if you, if you have a home robot that's doing something and needs to keep track, as a person, there are certainly some things where you keep track of them very symbolically, like almost in language. Like, you know, I have my checklist. Like, I'm going shopping and I, you know, at least for me, I can, like, literally visualize in my mind, like, my checklist. Like, you know, pick up the, the yogurt, pick up the milk, pick up whatever. And that, and I'm not, like, picturing the milk shelf with the milk sitting there. I'm just thinking, like, milk.

Dwarkesh Patel 48:13 Right.

Sergey Levine 48:13 Right? But then there's other things that are much more spatial, almost visual. Uh, you know, when, when I was, uh, trying to get to your, to your studio, I was thinking like, "Okay, uh, here's the, what this street looks like. Here's what that street looks like."

Dwarkesh Patel 48:27 Right.

Sergey Levine 48:27 "Here's the, you know, what I expect the doorway to look like."

Dwarkesh Patel 48:29 Yeah.

Sergey Levine 48:29 So representing your context in the right form that captures what you really need to achieve, uh, your goal, uh, and otherwise kind of discards all the unnecessary stuff, I think that- that's like, that's a really important thing.

Dwarkesh Patel 48:41 Yeah.

Sergey Levine 48:41 And I think we're, we're seeing the beginnings of that with multimodal models, but I think that multimodality has so much more to it than just, like, image plus text, and I think that that's a place where there's a lot of room for really exciting innovation.

Dwarkesh Patel 48:53 Ooh, do you mean in terms of, um...... how we represent?

Sergey Levine 48:57 Mm-hmm.

Dwarkesh Patel 48:57 Okay.

Sergey Levine 48:58 Yeah, how we represent both context, both what happened in the past, and also plans or reasoning-

Dwarkesh Patel 49:03 Yeah.

Sergey Levine 49:03 ... as you can call it in the LLM world, uh, which is what we would like to happen in the future, or intermediate processing stages in solving a task. I think do- doing that in a variety of modalities, including potentially learn modalities that are suitable for the job is something that has, I think, enormous potential, uh, to overcome some of these challenges.

Dwarkesh Patel 49:19 Interesting. Another question I have as we're dis- as- as we're discussing these, like, um, tough trade-offs in terms of, um, uh, inference is comparing it to the human brain and figuring out the human brain is able to have hours, decades of context while being like, being able to act on the order of 10 milliseconds while having 100 trillion parameters or however you wanna count it. And I wonder if the best way to understand what's happening here is that human brain hardware is just way more advanced than the hardware we have in GPUs, or that the algorithms for encoding video information are, like, way more efficient-

Sergey Levine 50:02 Mm-hmm.

Dwarkesh Patel 50:03 ... uh, and maybe it's, like, some crazy mixture of experts where-

Sergey Levine 50:07 Mm-hmm.

Dwarkesh Patel 50:08 ... the active parameters is also on the order of billions-

Sergey Levine 50:11 Mm-hmm.

Dwarkesh Patel 50:11 ... low billions, or some mixture of the two. Basically, if you had to think about, like, why do we have these models that are across many dimensions orders of magnitude less efficient? Is it hardware or algorithms than-

Sergey Levine 50:25 Yeah.

Dwarkesh Patel 50:25 ... c- compared to the brain?

Sergey Levine 50:26 Yeah, that's a really good question. So I definitely don't know the answer to this. Uh, I, I am not by any means well-versed in, in neuroscience, but if I had to guess and also provide an answer that leans more on things I know, it's something like this, that the brain is extremely parallel.

Dwarkesh Patel 50:42 Mm-hmm.

Sergey Levine 50:42 Uh, it kind of has to be just out of- just because of the biophysics. Um, but i- like, it's, it's even more parallel than your GPU.

Dwarkesh Patel 50:51 Yeah.

Sergey Levine 50:51 Uh, if you think about how a modern multimodal language model processes, uh, the input, uh, if you give it some images and some text, like, first it reads in the images, then it reads in the text and, and then proceeds, uh, one token at a time to generate the output. It makes a lot more sense to me for an embodied system to have parallel, uh, processes. Now, mathematically, you can actually make close equivalences between, uh, parallel and sequential stuff. Like, transformers aren't actually fundamentally sequential. Like, you kind of make them sequential by putting in position embeddings.

Dwarkesh Patel 51:22 Mm-hmm.

Sergey Levine 51:23 Transformers are fundamentally actually very parallelizable things.

Dwarkesh Patel 51:25 Right.

Sergey Levine 51:25 And that's what makes them so great. So I don't think that actually mathematically this, this, like, highly parallel thing where you're doing perception and proprioception and planning all at the same time-

Dwarkesh Patel 51:35 Right.

Sergey Levine 51:35 ... is actually, actually necessarily needs to look that different from a transformer.

Dwarkesh Patel 51:38 Yeah.

Sergey Levine 51:38 Although its practical implementation will be different. And you could imagine that the system will in parallel think about, uh, okay, here's, like, my long-term memory. Like, here's what I've seen, you know, a decade ago. Here's my short-term kind of spatial stuff. Here's my semantic stuff. Uh, here's what I'm seeing now. Here's what I'm planning. And all of that can be implemented in a way that there's some, you know, very familiar kind of attentional mechanism but in practice all, all running in parallel maybe at different rates, maybe with the more complex things running slower, the faster reactive stuff running faster.

Dwarkesh Patel 52:08 I'm sure you've been seeing a bunch of fun images that people have been generating with Google's new image generation model, Nano Banana. My X feed is full of wild images. But you might not realize that this model can also help you do less flashy tasks like restoring historical pictures or even just cleaning up images. For example, I was reading this old paperback as I was prepping to interview Sarah Payne, and it had this really great graph of World War II Allied shipping that I wanted to overlay in the lecture. Now in the past, this would have taken one of my editors 20 or 30 minutes to digitize and clean up manually. But now, we just took a photo of the page and then dropped into- into Nano Banana and got back a clean version. This was a one shot, but if Nano Banana doesn't nail it on the first attempt, you can try to just go back and forth with it until you get a result that you're super happy with. We keep finding new use cases for this model, and honestly, this is one of those tools that just doesn't feel real. Check out Gemini 2.5 Flash Image Model, AKA Nano Banana, on both Google AI Studio and the Gemini app. All right, back to Sergey. If in five years we have a system which is, like, as robust as a human in terms of interacting with the world, then what- what- what has happened that makes it physically possible to be able to run those kinds of models, um, to have video information that is streaming at real time or hours of prior video information is somehow being encoded and considered while decoding-

Sergey Levine 53:30 You know-

Dwarkesh Patel 53:31 ... in, like, a millisecond scale and with many more parameters? Uh, is it just that, like, uh, NVIDIA has shipped much better GPUs or that you guys have come up with much better, like, uh, encoders and stuff? Or, like-

Sergey Levine 53:41 Yeah.

Dwarkesh Patel 53:41 ... what's happened in the five years?

Sergey Levine 53:43 I- I- I think there's- there are a lot of things, uh, to this question. I think certainly there's, like, a really fascinating systems problem. Um, I'm by no means a systems expert, but I would imagine that the right architecture in practice, especially if you want an affordable low-cost system, would be to externalize at least part of the thinking.

Dwarkesh Patel 53:59 Yeah.

Sergey Levine 54:00 Uh, you know, you could imagine maybe in the future you'll have a robot that has, like, uh... You know, if your internet connection is not very good, the robot is in kind of like a dumber reactive mode.

Dwarkesh Patel 54:08 Mm-hmm.

Sergey Levine 54:08 But if you have a good internet connection, then it can-

Dwarkesh Patel 54:10 (laughs)

Sergey Levine 54:10 ... like, be a little smarter.

Dwarkesh Patel 54:11 Right.

Sergey Levine 54:11 That's pretty cool. Um, but I think there is- there are also research and algorithms things that can help here, um, like, uh, figuring out the right representations, concisely representing both, uh, your past observations, but also changes in observation.

Dwarkesh Patel 54:25 Yeah.

Sergey Levine 54:25 Right? Like, you know, your sensory stream is extremely temporally correlated, which means that the marginal information gained from each additional observation is not the same as the entirety of that observation because the image that I'm seeing now is very correlated to the image I saw before.

Dwarkesh Patel 54:38 Right. Yes.

Sergey Levine 54:38 So in principle, if I want to represent it concisely, I could get away with a much more compressed representation-

Dwarkesh Patel 54:42 Right.

Sergey Levine 54:42 ... than if I represent the images independently.

Dwarkesh Patel 54:44 Yeah.

Sergey Levine 54:44 So there's a lot that can be done on the algorithm side to get this right, and that's really interesting algorithms work. I think there's also, like, a really fascinating systems problem. Uh, to be truthful, like, I haven't gotten to the systems problem because, you know, you want to implement the system once you sort of know-

Dwarkesh Patel 54:58 Right.

Sergey Levine 54:58 ... the shape of the, of the machine learning-

Dwarkesh Patel 54:59 Yeah.

Sergey Levine 55:00 ... uh, solution. But I think there's a lot of cool stuff to do there.

Dwarkesh Patel 55:02 Yeah. E- e- and maybe you guys just need to hire, like, the people who run the YouTube data centers 'cause, like-

Sergey Levine 55:05 (laughs)

Dwarkesh Patel 55:05 ... they know how to, like, encode video information. Um, uh, okay, this actually is an interesting question, which is that with LLMs, of course-... they're being, theoretically, you could run your own model on this laptop or whatever. But realistically, what happens is that the largest, most effective models are being run in batches, um, of thousands and millions of users at the same time, uh, not locally. Well, the same thing happened in robotics because of the inherent deficiencies of batching, plus the fact that we have to do this incredibly computive, uh, compute intensive inference task. Um, and so you don't want to be carrying around, like, uh, you know, like a f- $50,000 GPUs per robot or something. You want, you just want that to happen somewhere else. Um, so yeah, should- we sh- this robotics world, should we just be anticipating something where you need connectivity everywhere, you need robots that are like, have like super fast, um... and you're streaming video information back and forth, right? Or at least video, video information one way. So does that have interesting implications about like how this, um, how this deployment of robots will actually be instantiated?

Sergey Levine 56:13 I, I don't know. Uh, but if I were to guess, I would guess that it'll, like, we'll actually see both.

Dwarkesh Patel 56:18 Mm-hmm.

Sergey Levine 56:18 That we'll see low cost systems with, uh, off-board inference and, uh, more reliable systems, uh, for example, in settings where, like, if you have an outdoor robot or a s- something where you can't rely on connectivity, that are costlier-

Dwarkesh Patel 56:32 Right.

Sergey Levine 56:32 ... and have onboard inference. Um, a few things I'll say from a technical standpoint that might contribute to understanding this. While a real time system obviously needs to be controlled in real time, often at high frequency, the amount of thinking you actually need to do for every time step-

Dwarkesh Patel 56:50 Mm-hmm.

Sergey Levine 56:50 ... might be surprisingly low. Uh, and again, we s- we see this in humans and animals. Uh, when we plan out movements, there is definitely a, a, a real planning process that happens in the brain.

Dwarkesh Patel 57:02 Yeah.

Sergey Levine 57:02 Like if, if you re- re- record like, uh, from a, from a monkey brain, you will actually find neural correlates of planning.

Dwarkesh Patel 57:09 Right.

Sergey Levine 57:09 And there is something that happens in advance of a movement, and when that movement actually takes place, the shape of the movement correlates with what happened before the movement.

Dwarkesh Patel 57:18 Yeah.

Sergey Levine 57:18 Like, that's planning, right? Uh, so that means that you put something in place and, you know, set the initial conditions of some kind of process and then unroll that process, and that's the movement.

Dwarkesh Patel 57:27 Yeah.

Sergey Levine 57:27 And that means that during that movement, you're doing less processing, uh, and you kind of batch it up in advance. But you're s- you're not, like, entirely in open loop. It's not like you're playing, you're playing back a tape recorder. You are actually reacting as you go, you're just reacting at a different level of abstraction.

Dwarkesh Patel 57:40 Mm-hmm.

Sergey Levine 57:40 Uh, uh, a, a more basic level of abstraction. And again, this comes back to representations. Figure out which representations are sufficient for kind of planning in advance and then unrolling. Which representations require a tight feedback loop?

Dwarkesh Patel 57:51 Right.

Sergey Levine 57:51 And for that tight feedback loop, like, what i- wha- what is, what are you doing feedback on? Like, you know, if I'm driving a vehicle, maybe I'm doing feedback on the position of the lane marker so that I stay straight,

Sergey Levine 57:59 and then at, at a lower frequency, I sort of gauge where I am in traffic.

Dwarkesh Patel 58:02 And then so you have a couple lectures from a few years back where you say like, even for robotics, RL is, in many cases, better than imitation learning. But f- so far, the models are exclusively doing imitation learning. So I'm curious how your, ho- how is your thinking on this has changed, or maybe it's not changed, but then you, you need to do this for the RL. Like, why, w- why can't you do RL yet?

Sergey Levine 58:22 Yeah. So the key here is prior knowledge.

Dwarkesh Patel 58:24 Yeah.

Sergey Levine 58:25 Uh, so in order to effectively learn from your own experience, it turns out that it's really, really important to already know something about what you're doing. Otherwise, it takes far too long. Uh, it's just like it, it, it takes, uh, uh, a person, when they're a child, a very long time to learn very basic things.

Dwarkesh Patel 58:41 Yeah.

Sergey Levine 58:41 To learn to write for the first time, for example. Once you already have some knowledge, then you can learn new things very quickly. So the purpose of training the models with supervised learning now is to build out that foundation that provides the prior knowledge so they can figure things out much more quickly later.

Dwarkesh Patel 58:57 Mm-hmm.

Sergey Levine 58:57 And again, this is not a new idea. This is exactly what we've seen with, uh, LMs, right? LMs started off, uh, being trained purely with next token prediction, and that provided an excellent starting point, first for all sorts of synthetic data generation and then, uh, for RL.

Dwarkesh Patel 59:11 Hmm.

Sergey Levine 59:12 So I, I, I think it, it makes total sense that we would expect basically any foundational effort to follow that same trajectory, where we first build out the foundation, essentially in like a somewhat brute force way.

Dwarkesh Patel 59:21 Right.

Sergey Levine 59:22 And the stronger that foundation gets, the easier it is to then make it even better with much more accessible-

Dwarkesh Patel 59:28 Right.

Sergey Levine 59:28 ... training.

Dwarkesh Patel 59:29 In, um, in 10 years, will the best model for knowledge work also be a robotics model or have like a r- action expert attached to it? And the reason I ask is like-

Sergey Levine 59:37 Mm-hmm.

Dwarkesh Patel 59:38 ... so far, we've seen advantages from using more general models for things.

Sergey Levine 59:42 Yeah.

Dwarkesh Patel 59:42 And will robotics fall into this bucket of we will just have the model which does everything, including physical work and knowledge work? Or do you think they'll continue to stay separate?

Sergey Levine 59:53 I really hope that they will actually be the same. And, um, you know, obviously, I'm extremely biased. I, I love robotics. I think it's like, it's very fundamental to AI. But I think that it's optimistically that it's actually the other way around, that the robotics, uh, element of the equation will make all the other stuff better. And there are two, uh, reasons for this that I could, that I, that I could tell you about. One has to do with representations and focus. So what I said before, with, uh, video prediction models, if you just want to predict everything that happens, it's very hard to figure out what's relevant.

Dwarkesh Patel 1:00:27 Yeah.

Sergey Levine 1:00:27 If you have the focus that comes from actually trying to do a task, now that acts to structure how you see the world in a way that, uh, allows you to more fruitfully utilize the other signals. That could be extremely powerful.

Dwarkesh Patel 1:00:39 Yeah.

Sergey Levine 1:00:40 The second one is that understanding the physical world at a, at a very deep, fundamental level, at a level that goes beyond just what we can articulate with language, can actually help you solve other problems.

Dwarkesh Patel 1:00:49 Mmm.

Sergey Levine 1:00:50 Uh, and we, we see, we, we experience this all the time. Like, when we talk about abstract concepts, we say like, "This company has a lot of momentum."

Dwarkesh Patel 1:00:57 Yeah.

Sergey Levine 1:00:57 (laughs) Uh, right? I, I, I, like you, w- we'll use like social metaphors to describe inanimate objects, like, "My computer hates me." (laughs) Right?

Dwarkesh Patel 1:01:05 Right.

Sergey Levine 1:01:05 Like, we experience the world in a particular way and our subjective experience shapes how we think about it in very profound ways. And then we u- use that as a hammer to basically hit all sorts of other nails that are far too abstract, uh, to handle any other way.

Dwarkesh Patel 1:01:17 I guess but th- there might be other considerations that are relevant to...... uh, physical robots in terms of, like, inference speed and model size, et cetera. Which might be different than the considerations for knowledge work. But then maybe you can... Maybe that doesn't change... Maybe it's still the same model, but then you can serve it in different ways and the advantages of co-training are high enough that... Yeah. Whe- whenever I'm... Like, I'm wondering in five years if I'm using a model to code for me, does it also know how to do robotic stuff and... yeah, may- maybe the advantages of co-training and robotics are high enough that it's worth.

Sergey Levine 1:01:49 Mm-hmm. Well, a- and I should say that the coding is probably, like, the pinnacle of, uh, of, uh, abstract knowledge work.

Dwarkesh Patel 1:01:55 Right.

Sergey Levine 1:01:55 In the sense that, like, just by, by the, by the mathematical nature of computer programming-

Dwarkesh Patel 1:01:59 Yeah.

Sergey Levine 1:01:59 ... it's an extremely abstract activity.

Dwarkesh Patel 1:02:00 Yeah.

Sergey Levine 1:02:00 Which is why people struggle with it so much.

Dwarkesh Patel 1:02:02 Yeah. I'm a bit confused about why simulation doesn't work better for, uh, robots. Like, if I look at humans, smart humans do a good job of... if they're intentionally trying to learn, noticing what about the simulation is similar to real life and paying attention to that and learning from that. So, if you have, like, pilots who are learning in simulation or F1 drivers who are learning in simulation, should it be expected to be the case that as robots get smarter, they will also be able to learn more things from simulation or, uh-

Sergey Levine 1:02:33 Yeah.

Dwarkesh Patel 1:02:33 ... or is this a curse and we need real world data forever?

Sergey Levine 1:02:36 Th- this is a very subtle question. Um, your example with the airplane pilot using simulation is really interesting, but something to remember is that when a pilot is using a simulator to learn to fly an airplane, they're extremely goal-directed. So, their goal in life is not to learn to use a simulator, their goal in life is to learn to fly the airplane.

Dwarkesh Patel 1:02:55 Yeah.

Sergey Levine 1:02:55 They know there will be a test afterwards and they know that eventually they'll be in charge of, like, a few hundred passengers-

Dwarkesh Patel 1:02:59 Right.

Sergey Levine 1:02:59 ... and they really need to not crash that thing.

Dwarkesh Patel 1:03:02 (laughs)

Sergey Levine 1:03:02 Um, and when we train, um, models on data from multiple different domains, the models don't know that they're supposed to solve a particular task. They just see, like, "Hey, here's one thing I need to master. Here's another thing I need to master."

Dwarkesh Patel 1:03:15 Right.

Sergey Levine 1:03:15 So, maybe, like, a better analogy there is if you, if you're, like, playing a video game where you can fly an airplane and then eventually someone puts you in the cockpit of a real, of a real one. Like, it's not that the video game is useless, but it's, it's not the same thing. And if you're trying to play that video game and your goal is to, like, really, like, master the video game, um, you're not gonna go about it in quite the same way.

Dwarkesh Patel 1:03:33 Oh. Oh, isn't, um... Can you do some kind of meta RL on this, which is, like, almost identical actually to the... uh, there's this really interesting paper you wrote in 2017 where maybe the loss function is not how well it does at a particular video game or particular simulation, but how well being trained at different video games makes it better at some other downstream task? I, I did a terrible job explaining, but I, I, I-

Sergey Levine 1:03:56 I understand what you mean.

Dwarkesh Patel 1:03:57 Yeah, yeah. Okay. Maybe, can you do a better job explaining what I, what I was trying to say?

Sergey Levine 1:03:59 Yeah. So, so, so I think what you're trying to say is basically that, uh, well, maybe if we have, like, a really smart model that's doing meta learning, perhaps it can figure out that its performance on a downstream problem, a real world problem, is increased by doing something in a simulator.

Dwarkesh Patel 1:04:13 And then specifically make that the loss function, right?

Sergey Levine 1:04:16 Yeah. That's right.

Dwarkesh Patel 1:04:16 Yeah.

Sergey Levine 1:04:16 But here's the thing with, with, with this. There's a set of these I- ideas that are all gonna be, like, something like train to make it better on the real thing by leveraging something else.

Dwarkesh Patel 1:04:26 Yeah.

Sergey Levine 1:04:27 And the key linchpin for all of that is the ability to train it to be better on the real thing.

Dwarkesh Patel 1:04:31 Yeah.

Sergey Levine 1:04:32 Um, the thing is, like, I actually suspect in reality we might not even need to do something quite so explicit because meta learning is emergent, as you pointed out before, right? Like, uh, LMS essentially do a kind of meta learning in, via in-context learning. I mean, we can debate as to how much that's learning or not, but the point is that large powerful models trained on the right objective on real data get much better at leveraging all the other-

Dwarkesh Patel 1:04:53 Yeah.

Sergey Levine 1:04:54 ... stuff. And I think that's actually the key. And coming back to your, uh, airplane pilot, like, the airplane pilot is trained on a real world objective. Like, their objective is to be a good airplane pilot, to be successful, to have a good career, and all of that kind of propagates back into the actions they take-

Dwarkesh Patel 1:05:08 Right.

Sergey Levine 1:05:08 ... in leveraging all these other data sources. So, what I think is actually the key here to leveraging auxiliary data sources, including simulation, is to build the right foundation model that is really good that has those emergent abilities.

Dwarkesh Patel 1:05:19 Mm-hmm.

Sergey Levine 1:05:20 And, to your point, uh, to get really good like that, it has to have the right objective. Uh, now, we know how to get the right objective out of real world data. Maybe we can get it out of other things, but that's, that's harder right now.

Dwarkesh Patel 1:05:32 Right.

Sergey Levine 1:05:33 And I think that it... again, we can look to the examples of what happened in other fields. Like, these days, if someone trains an LLM for solving complex problems, they're using lots of synthetic data.

Dwarkesh Patel 1:05:43 Mm-hmm.

Sergey Levine 1:05:44 But the reason they're able to leverage that synthetic data effectively is because they have this starting point that is trained on lots of real data that kinda gets it-

Dwarkesh Patel 1:05:50 Yeah.

Sergey Levine 1:05:50 ... and once it gets it, then it's more able to leverage all this other stuff.

Dwarkesh Patel 1:05:53 Right.

Sergey Levine 1:05:54 So, uh, I think, like, perhaps ironically, the key to leveraging other data sources, including simulation, is to get really good at using real data, understand what's up with the world, and then now y- now you can fruitfully use all this other stuff.

Dwarkesh Patel 1:06:04 So, so once we have this, like, um... in 2035, 2030, what... basically the sci-fi world, uh, are you optimistic about the ability of, like, true AGIs to build simulations in which they are rehearsing skills that no human or AI has ever had a chance to practice before? Um, some e- you know, they need to, like, practice to be astronauts because we're building the Dyson sphere and they can just do that in simulation or, like, will the issue with simulation continue to be one e- regardless of how smart the models get?

Sergey Levine 1:06:34 So here's what I would say, that deep down at a very fundamental level, the synthetic experience that you create yourself doesn't allow you to learn more about the world. It allows you to rehearse things, it allows you to consider counterfactuals, but somehow information about the world needs to get injected into this system. So, um, a- and I think the way you posed this question actually elucidates this very nicely because in robotics classically people have often thought about simulation as a way to inject human knowledge because a person knows how to write down, like, differential equations-

Dwarkesh Patel 1:07:09 Right.

Sergey Levine 1:07:09 ... they can code it up and that, like, gives the robot more knowledge than it had before. But, uh, I think that increasingly what we're learning from, uh, experiences in other fields, from, uh, how, like, the video generation stuff goes, from synthetic data for LMSs, that actually probably the most powerful way to create synthetic experience is from a really good model...... uh, because, you know, the model probably knows more than a person does about those fine-grain details. But then, of course, where does that model get the knowledge? From experiencing the world.

Dwarkesh Patel 1:07:34 Yeah.

Sergey Levine 1:07:35 So in a sense, what you said, uh, I think is actually quite right in that a very powerful AI system can simulate a lot of stuff, but also at that point, it kind of almost doesn't matter because viewed as a black box, what's going on with that system is that information comes in and capability comes out.

Dwarkesh Patel 1:07:51 Yeah.

Sergey Levine 1:07:52 And whether the way it processes that information is by imagining some stuff and simulating or by some model-free method, uh, is kind of irrelevant in our understanding its capabilities.

Dwarkesh Patel 1:07:59 Do you have any sense of what it, what the equivalent is in humans? Like, whatever we're doing when we're daydreaming or sleeping or, um... I don't know if you have some sense of, like, what this auxiliary thing we're doing is, but if you had to make an ML analogy for it, what is it?

Sergey Levine 1:08:14 Well, uh, yeah, I mean, cer- certainly when, uh, uh, when you sleep, your brain does stuff that looks an awful lot like what it does when it's awake, uh, that looks an awful lot like playing back experience or perhaps generating new statistically similar experience.

Dwarkesh Patel 1:08:27 Right.

Sergey Levine 1:08:28 Um, a- and so I think it's like, it's very reasonable to guess that perhaps simulation through a learned model is, like, part of how your brain figures out, like, counterfactuals basically.

Dwarkesh Patel 1:08:40 Yeah.

Sergey Levine 1:08:40 But, um, something that's kind of even more fundamental than that is that optimal decision-making at its core, regardless of how you do it, requires considering counterfactuals. You basically have to ask yourself, "If I did this instead of that, would it be better?" And you have to answer that question somehow, and whether you answer that question by using a learned simulator or whether you, you answer that question by using a value function, uh, or something like that, by using a reward model, in the end it's kind of all the same. Like, as long as you have some mechanism for considering counterfactuals and figuring out-

Dwarkesh Patel 1:09:10 Yeah.

Sergey Levine 1:09:10 ... which counterfactual is better, you've got it.

Dwarkesh Patel 1:09:12 Yeah.

Sergey Levine 1:09:12 Uh, so that, that, that... I, I like thinking about it this way because it k- it kinda simplifies things. It tells us that the key is

Sergey Levine 1:09:18 not necessarily to do really good simulation. The key is to figure