Pod Insights No Priors The Road to Autonomous Intelligence with Andrej Karpathy About this episode Andrej Karpathy joins Sarah and Elad in this week of No Priors. Andrej, who was a founding team member of OpenAI and former Senior Director of AI at Tesla, needs no introduction. In this episode, Andrej discusses the evolution of self-driving cars, comparing Tesla and Waymo’s approaches, and the technical challenges ahead. They also cover Tesla’s Optimus humanoid robot, the bottlenecks of AI development today, and how AI capabilities could be further integrated with human cognition. Andrej shares more about his new company Eureka Labs and his insights into AI-driven education, peer networks, and what young people should study to prepare for the reality ahead.
Sign up for new podcasts every week. Email feedback to show@no-priors.com
Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @Karpathy
Show Notes:
(0:00) Introduction
(0:33) Evolution of self-driving cars
(2:23) The Tesla vs. Waymo approach to self-driving
(6:32) Training Optimus with automotive models
(10:26) Reasoning behind the humanoid form factor
(13:22) Existing challenges in robotics
(16:12) Bottlenecks of AI progress
(20:27) Parallels between human cognition and AI models
(22:12) Merging human cognition with AI capabilities
(27:10) Building high performance small models
(30:33) Andrej’s current work in AI-enabled education
(36:17) How AI-driven education reshapes knowledge networks and status
(41:26) Eureka Labs
(42:25) What young people study to prepare for the future
Transcript 10,659 words · 3 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:05 Sarah Guo? Hi listeners, welcome back to No Priors. Today, we're hanging out with Andrej Karpathy, who needs no introduction. Andre is a renowned researcher, beloved AI educator and Cuber, an early team member from OpenAI, the lead for Autopilot at Tesla, and now working on AI for education. We'll talk to him about the state of research, his new company, and what we can expect from AI.
0:29 Speaker 01 Thanks, Andre, for joining us today. It's great to have you here.
0:31 Speaker 02 Thank you. I'm happy to be here.
0:33 Sarah Guo? You led Autopilot at Tesla, and now we actually have fully self-driving cars, passenger vehicles on the road. How do you read that in terms of where we are in the capability set? How quickly we should see increased capability or pervasive passenger vehicles?
0:48 Speaker 02 Yes, I spent maybe five years on self-driving space. I think it's a fascinating space. And basically, what's happening in the field right now is, well, I do also think that I draw a lot of analogies, I would say, to AGI from self-driving, and maybe that's just because I'm familiar with it. But I kind of feel like we've reached AGI a little bit in self-driving, because there are systems today that you can basically take around and as a paying customer can take around here. So Waymo in San Francisco here is, of course, very common. Probably you've taken Waymo. I've taken it a bunch, and it's amazing. And it can drive you all over the place, and you're paying for it as a product. What's interesting with Waymo is the first time I took Waymo was actually a decade ago, almost exactly, 2014 or so. And it was a friend of mine who worked there, and he gave me a demo. And it drove me around the block 10 years ago. And it was basically a perfect drive 10 years ago. And it took 10 years
1:33 Speaker 02 to go from a demo that I had to a product I can pay for that's in the city scale and is expanding, et cetera.
1:39 Speaker 01 How much of that do you think was regulatory versus technology? Like, when do you think the technology was ready? Is it recent?
1:43 Speaker 02 I think it's technology. You're just not seeing it in a single demo drive of 30 minutes. You're not running into all the stuff that they had to deal with for a decade. And so demo and product, there's a massive gap there. And I think a lot of it also regulatory, et cetera. But I do think that we've sort of like achieved AGI in the self-driving space in that sense a little bit. And yet, I think there's what's really fascinating about it is the globalization hasn't happened at all. So you have a demo and you can take it in the self, but like the world hasn't changed yet. And that's going to take a long time. And so going from a demo to an actual globalization of it, I think there's a big gap there. That's how it's related, I would say, to AGI because I suspect similar, it looks in a similar way for AGI when we sort of get it. And then staying for a minute in the self-driving space, I think people think that Waymo is ahead of Tesla. I think personally Tesla is ahead of Waymo and I know it doesn't look like that.
2:28 Speaker 02 But I'm still very bullish on Tesla and its self-driving program. I think that Tesla has a software problem and I think Waymo has a hardware problem is the way I put it. And I think software problems are much easier. Tesla has a deployment of all these cars on Earth, like at scale. And I think Waymo needs to get there. And so the moment Tesla sort of like gets to the point where they can actually deploy this and it actually works, I think it's going to be, you know, really incredible. The latest build I just drove yesterday, I mean, it's just driving me all over the place now. They've made like really good improvements, I would say, very recently.
2:59 Speaker 01 Yeah, I've been using it a lot recently and it actually works quite well.
3:02 Speaker 02 It was, it did some miraculous driving for me yesterday, so I'm very impressed with what the team is doing. And so I still think that Tesla mostly has a software problem, Waymo mostly hardware problem. And so I think Tesla, Waymo looks like it's winning kind of right now. But I think when we look in 10 years and who's actually at scale and where most of the revenue is coming from, I still think they're, they're ahead in that sense.
3:22 Speaker 01 How far away do you think we are from the software problem turning the corner in terms of getting to some equivalency? Because obviously to your point, if you look at a Waymo car, it has a lot of very expensive LiDAR and other sort of sensors built into the car so it can do what it does. It sort of helps support the software system. And so if you can just use cameras, which is the Tesla approach, then you effectively get rid of enormous costs slash complexity and you can do it in many different types of cars. When do you think that transition happens?
3:47 Speaker 02 I mean, in the next few years, I mean, I'm hoping something like that. But actually, what's really interesting about that is I'm not sure that people are appreciating that. Tesla actually does use a lot of expensive sensors. They just do it at training time. So there are a bunch of cars that drive around with LiDARs. They do a bunch of stuff that doesn't scale and they have extra sensors, et cetera. And they do mapping and all the stuff. You're doing it at training time and then you're distilling that into a test time package that gets deployed to the cars and is vision only. And it's like an arbitrage on sensors and expense.
4:14 Sarah Guo? Yeah.
4:14 Speaker 02 And so I think it's actually kind of a brilliant strategy that I don't think is fully appreciated and I think it's going to work out well because the pixels have the information and I think the network will be capable of doing that. And yes, at training time, I think these sensors are really useful, but I don't think they're as useful at test time and I think you can.
4:27 Speaker 01 It seems like the one other thing or transition that's happened is basically a move from a lot of sort of edge case design heuristics associated with it versus end-to-end deep learning. And that's what other shift that's happened recently. Do you want to talk a little bit about that and sort of what that?
4:41 Speaker 02 Yeah, I think that was always like the plan from the start, I would say, at Tesla is I was talking about how the neural net can eat through the stack. Because when I joined, there was a ton of C++ code and now there's much, much less C++ code in the test time package that runs in the car. Because there's still a ton of stuff in the backend that we're not talking about. The neural net kind of like takes through the system. So first it just does like a detection on the image level, then it does multiple images, gives you prediction, then multiple images over time give you a prediction and you're discarding C++ code and eventually you're just giving steering commands. And so I think Tesla is kind of eating through the stack. My understanding is that current WIMOs are actually like not that, but that they've tried, but they ended up like not doing that is my current understanding, but I'm not sure because they don't talk about it.
5:23 Speaker 02 But I do fundamentally believe in this approach and I think that's the last piece to fall, if you want to think about it that way. And I do suspect that the end-to-end systems for Tesla in like say 10 years, it is just a neural net. I mean, the videos stream into a neural net and commands come out. You have to sort of build up to it incrementally and do it piece by piece. And even all the intermediate predictions and all these things that we've done, I don't think they've actually misled the development. I think they're part of it because. There's a lot of subtle reasons for this. So actually like end-to-end driving, when you're just imitating humans and so on, you have very few bits of supervision to train a massive neural net. And it's too few bits of signal to train so many billions of parameters.
6:06 Speaker 02 And so these intermediate representations and so on help you develop the features and the detectors for everything. And then it makes it a much easier problem for the end-to-end part of it. And so I suspect, although I don't know because I'm not part of the team, but there's a ton of pre-training happening so that you can do the fine-tuning for end-to-end. And so basically, I feel like it was necessary to eat through it incrementally. And that's what Tesla has done. I think it's the right approach. And it looks like it's working, so I'm really looking forward.
6:28 Sarah Guo? If you had started end-to-end, you wouldn't have had the data anyway. That makes sense, yeah. So you worked on the Tesla humanoid robot before you left. I have so many questions, but one is like, starting here, what transfers?
6:40 Speaker 02 Basically, everything transfers, and I don't think people appreciate it.
6:43 Sarah Guo? Okay, that's a big claim.
6:45 Speaker 02 I think cars are basically robots when you actually look at it. Cars are robots. And Tesla, I don't think is a car company. I think this is misleading. This is robotics company. Robotics at scale company. Because I would say at scale is also like a whole separate variable. They're not building a single thing. They're building the machine that builds the thing, which is a whole separate thing. And so I think robotics at scale company is what Tesla is. And I think in terms of the transfer from cars to humanoid, it was not that much work at all. And in fact, like the early versions of Optimus, the robot, it thought it was a car. Like, because it had the exact same computer, it had the exact same cameras. It was really funny because we were running the car networks on the robot, but it's walking around the office and so on. And like, it's trying to like, recognize drivable space,
7:29 Speaker 02 but it's all just walking space now, I suppose. But it actually kind of like generalized a little bit, and there's some fine-tuning necessary and so on. But it thought it was driving, but it's actually like moving through an environment.
7:39 Sarah Guo? Is a reasonable way to think of this as like, actually, it's a robot. Many things transfer, but you're just missing, for example, actuation and action data.
7:46 Speaker 02 Yeah, you definitely miss some components. But, and the other part I would say is like, so much transfers. Like, the speed with which Optimus was started, I think to me was very impressive. Because the moment Elon said, "We're doing this", just people just showed up with all the right tools and the stuff just showed up so quickly and all these CAD models and all the supply chain stuff. And I just felt like, wow, there's so much in-house expertise for building robotics at Tesla. And like, it's all the same tools and they're just like, okay, they're being reconfigured from a car, like a Transformer, the movie. They're just being reconfigured and reshuffled, but it's like the same thing. And you need all the same components. You need to think about all the same kinds of stuff, both on the hardware side, on the scale stuff, and also on the brains. And so for the brains, there was also a ton of transfer, not just of the specific networks, but also all of the approach and the labeling team and how it all coordinates
8:31 Speaker 02 and the approaches people are taking. I just think there's a ton of transfer. What do you think of the first application areas for humanoid robotics or human form stuff? I think a lot of people have this vision of it, like doing your laundry, et cetera. I think that will come late. I don't think B2C should be the right start point because I don't think we can have a robot like Crush Grandma is how I put it, sort of. I think it's like too much legal liability. It's just like, I don't think that's the right approach.
8:52 Sarah Guo? But like a very torquey hug.
8:54 Speaker 02 I mean, it's just gonna fall over or something like that, you know? Like, these things are not perfect yet and they require some amount of work. So I think the best customer is yourself first. And I think probably Tesla's gonna do this. I'm very bullish on Tesla, if people can tell. The first customer is yourself and you incubate it in the factory and so on, doing maybe a lot of material handling, et cetera. This way, you don't have to create contracts working with third parties. It's all really heavy. There's lawyers involved, like, et cetera. You incubate it, then you go, I think, B2B second. And you go to other companies that have massive warehouses. We can do material handling. We're gonna do all this stuff. Contracts get drafted up. Fences get put around, all this kind of stuff. And then, once you're incubated in multiple companies, I think that's when you start to go into the B2C applications. I do think we'll see B2C robots also, like Unitree and so on,
9:38 Speaker 02 are starting to come out with robots that I really want.
9:40 Sarah Guo? I got one.
9:41 Speaker 02 You did?
9:41 Sarah Guo? Yeah. Yeah, the G1.
9:43 Speaker 02 Yeah. So I will probably buy one of those and there's probably gonna be an ecosystem of people building on those platforms too. But I think in terms of like what like wins at scale, I would expect that kind of a approach. But in the beginning, it's a lot of material handling and then going towards more and more HKC things that are more specific. One that I'm really excited about is the Net Friedman challenge of the leaf blower.
10:04 Sarah Guo? Yeah.
10:05 Speaker 02 Like, I would love for an optimist to walk down the street, like tiptoe down the street and like pick up individual leaves so that we don't need like leaf blowers. And I think this will work and it's an amazing task. And so I would hope that that's one of the first applications. Or just even raking. Yeah, that's what worked too. Just very quietly. Yeah, just quiet raking.
10:20 Sarah Guo? That's cute. They, I mean, they do actually have like a machine that's working, it's just not a humanoid. Can we talk about the humanoid thesis for a second? Because the simplest version of this is like the world is built for humans and you build one set of hardware. The right thing to do is build a model that can do an increasing set of tasks in this set of hardware. I think there's a, like, another... Camp that believes like, well, like, humans are not optimal for any given task, right? You can make them stronger or bigger or smaller or whatever. And why shouldn't we do superhuman things? Like, how do you think about this?
10:52 Speaker 02 I think people are maybe underappreciating the complexity of any fixed cost that goes into any single platform. I think there's a large fixed cost you're paying for any single platform. And so I think it makes a lot of sense to centralize that and have a single platform that can do all the things. I would say the humanoid aspect is also very appealing because people can teleoperate it very easily. And so it's a data collection thing that is extremely helpful, because people will be able to, obviously, very easily teleoperate it. I think that's usually overlooked. There's, of course, the aspect you mentioned, which is like, the world is designed for humans, etc. So I think that's also important. I mean, I think we'll have some variations on the humanoid platform, but I think there is a large fixed cost to any platform. And then I would say also, one last dimension of it is, you benefit a little ton from like, the transfer learning between the different tasks. And in AI, you really want the single neural net that is multitasking, doing lots of things, that's where you're getting all the
11:37 Speaker 02 intelligence and the capability from. And that's also why language models are so interesting, is because you have a single regime, like a text domain, multitasking all these different problems. And they're all sharing knowledge between each other, and it's all coupled in a single neural net. And I think you want that kind of a platform. And, you know, you want all the data you collect for leaf picking to benefit all the other tasks. If you're building a special purpose thing for any one thing, you're not going to benefit from a lot of the transferring between all the other tasks, if that makes sense.
12:04 Sarah Guo? Yeah, I think there's one argument of like, it seems, I mean, the G1 is like 30 grand, right? But it seems hard to build a very capable humanoid robot under a certain bomb. And like, if you wanted to, you know, put an arm on wheels that can do things, like, like, maybe there are cheaper approaches to a general platform at the beginning. Does that make sense to you?
12:25 Speaker 02 Cheaper approaches to a general platform.
12:26 Sarah Guo? From a hardware perspective.
12:28 Speaker 02 Yeah, I think that makes sense. Yeah, you put a wheel on it instead of a feet, etc. I do feel like, I wonder if it's taking down like a local minimum, a little bit. I just feel like, pick a platform, make it perfect, is like, the long term, pretty good bet. And then the other thing, of course, is like, I just think it will be kind of like, familiar to people. And I think people will understand that maybe when I talk with it. And I feel like the psychological aspect also of it, I think, favors. Possibly the humanoid platform unless people are scared of it and would actually prefer a platform that is more abstract of like some but then I don't know if it's just some
13:00 Sarah Guo? We've had sci-fi for so long.
13:01 Speaker 02 A wheelchair monster doing stuff and I don't know if that's like more appealing or less appealing.
13:04 Speaker 01 It's interesting that I think that the other form factor for the unitary is the dog, right? And it's almost a more friendly or familiar.
13:10 Speaker 02 Yeah, but then people watch Black Mirror and suddenly the dog flips to like a scary thing. So it's hard to think through. I just think psychologically it'll be easy for people to understand what's happening and.
13:21 Speaker 01 What do you think is missing in terms of technological milestones for progress relative to substantiating this future?
13:28 Speaker 02 For robotics specifically.
13:30 Speaker 01 Yeah, or the humanoid robot or anything else human form.
13:33 Speaker 02 Yeah, I don't know that I have like a really good window into it. I do think that it is kind of interesting that like in the humanoid form factor, for example, for the lower body, I don't know that you want to do imitation learning from like demonstration because for lower body, it's all a lot of like inverted pendulum control and stuff like that. It's for the upper body that you need a lot of like teleoperation and data collection and end to end and etc. And so I think like everything becomes like very hybrid in that sense and I don't know how those systems interact.
13:57 Speaker 01 When I talk to people working, they feel a lot of what they focus on is like actuation and you know, manipulation and sort of digital manipulation and things like that.
14:04 Speaker 02 Yeah, I do expect in the beginning it's a lot of like teleoperation for getting stuff off the ground and then imitating it and getting something that works 95% of the time. And then talking about human to robot ratios and gradually having people who are supervisors of robots instead of doing the task directly and all this kind of stuff is going to happen over time and pretty gradually. I don't know that there's like any individual impediments that I'm like really familiar with. I just think it's a lot of grunt work. A lot of like the tools are available. Transformers are this beautiful like blob of tissue. You can just get just arbitrary tasks and you just need the data. You need to put it in the right form. You need to train it. You need to experiment with it. You need to deploy it, iterate on it, that just a lot of grunt work. I don't know that I have a single individual thing that is like holding us back technically.
14:41 Sarah Guo? Where are we in the state of large blob research?
14:45 Speaker 02 Large blob research?
14:46 Sarah Guo? Yeah.
14:47 Speaker 02 We're in a really good state. So I think, I'm not sure if it's fully appreciated, but like the transformer is like much more amazing. It's not just like, it's not just another neural net. It's like an amazing neural net, extremely general. So for example, when people talk about the scaling loss in neural networks, the scaling laws are actually a, to a large extent of a property of the transformer. Before the transformer, people were playing with LSTMs and stacking them, etc. You don't actually get like clean scaling loss. And this thing doesn't actually train and doesn't actually work. It's the transformer that was the first thing that actually just kind of like scales. And you get scaling loss and everything makes sense. So it's just like general purpose training computer. I think of it as kind of a computer, but it's like a differentiable computer. And you can just give it inputs and outputs and billions of it. And you can train with backpropagation.
15:30 Speaker 02 It actually kind of like arranges itself into a thing that does the task. And so I think it's actually kind of like a magical thing that we've stumbled on in the algorithm space. And I think there's a few individual innovations that went into it. So you have the residual connections. That was a piece that existed. You have the layer normalizations that needs to slot in. You have the attention block. You have the lack of these like saturating non-linaritys like tan h's and so on. Those are not present in the transformer because they kill gradient signals. So there's a few like, there's four or five innovations that all existed. And we're put together into this transformer. And that's what Google did with their paper. And this thing actually trains. And suddenly you get scaling loss and suddenly you have like this piece of tissue that just trains to a very large extent. And so it was a major unlock.
16:11 Sarah Guo? You feel like we are not near the limit of that unlock, right? Because I think there is a discussion of, of course, like the data wall and how expensive another generation of scale would be. Like, how do you think about that?
16:22 Speaker 02 That's where you start to get into, like, I don't think that the neural network architecture is like holding us back fundamentally anymore. It's like not the bottleneck. Whereas I think in the previous, before transformer, it was a bottleneck. But now it's not the bottleneck. So now we're talking a lot more about what is the loss function? What is the data set? We're talking a lot more about those and those have become the bottlenecks almost. It's not the general piece of tissue that we configure is based on whatever you want it to be. And so that's where I think a lot of the activity has moved. And that's why a lot of the companies and so on who are applying this technology, like they're not thinking about the transformer much. They're not thinking about the architecture. You know, the Llama release, like the transformer hasn't changed that much. You know, we've added the rope positional and the rope relative positional encodings. That's like the major change. Everything else doesn't really matter too much. It's like plus 3% on the small few things. But really it's like rope is the only thing that's slotted in.
17:07 Speaker 02 And that's the transformer as it has changed since the last five years or something. So there hasn't been that much innovation on that. Everyone just takes it for granted. Let's train it, etc. And then everyone's just innovating on the data set mostly and the loss function details. So that's where all the activity has gone to.
17:22 Sarah Guo? Right, but what about the argument like in that domain that that was easier when we were taking internet data? And we're out of internet data. And so the questions are really around like synthetic data or more expensive data collection.
17:35 Speaker 02 So I think that's a good point. So that's where a lot of the activity is now in the LLMs. So the internet data is like not the data you want for your transformer. It's like a nearest neighbor that actually gets you really far, surprisingly. But the internet data is a bunch of internet webpages, right? It's just like what you want is the inner thought monologue of your brain.
17:51 Sarah Guo? Yep.
17:52 Speaker 02 That's the idea.
17:52 Sarah Guo? The trajectories in your brain.
17:53 Speaker 02 The trajectories in your brain as you're doing problem solving. If we had a billion of that, like AGI is here, roughly speaking, I mean, to a very large extent. And we just don't have that. So where a lot of activity is now, I think, is with the internet data that actually gets you like really close. Because it just so happens that internet has enough of reasoning traces in it and a bunch of knowledge and the transformer just makes it work. Okay. So I think a lot of activity now is around refactoring the data set into these inner monologue formats. And I think there's a ton of synthetic data generation that's helpful for that. So what's interesting about that also is like the extent to which the current models are helping us create the next generation of models. And so it's kind of like, you know, the staircase of improvement.
18:31 Speaker 01 How much you think synthetic data is, or how far does that get us, right? Because to your point on each data, each model helps you train the subsequent model better, or at least create tools for it, data labeling, whatever, maybe. Part of it is synthetic data. How important do you think the synthetic data piece is?
18:44 Speaker 02 Because when I talk to people, they say, yeah. I think is the only way we can like make progress is we have to make it work. I think with synthetic data, you just have to be careful. Because these models are silently collapsed is like one of the major issues. So if you go to chat GPT and you ask it to give you a joke, you'll notice that it only knows like three jokes. That's like the only, it gives you like one joke, I think, most of the time. And sometimes it gives you like three jokes. And it's because the models are collapsed and it's silent. So when you're looking at any single individual output, you're just seeing a single example. But when you actually look at the distribution, you'll notice that it's not a very diverse distribution. It's silently collapsed. When you're doing synthetic data generation, this is a problem because you actually really want that entropy. You want the diversity and richness in your data set. Otherwise, you're getting collapsed data sets and you can't see it when you look at any individual example.
19:26 Speaker 02 But the distribution is, has lost a ton of entropy and richness. And so it silently gets worse. And so that's why you have to be very careful and you have to make sure that you maintain your entropy in your data set. And there's a ton of. Techniques for that. As an example, someone released this persona dataset as an example. The persona dataset is a dataset of one billion personalities, like humans, like backgrounds.
19:50 Sarah Guo? Oh, yes, I saw this, yeah.
19:51 Speaker 02 I'm a teacher, or I'm an artist, I live here, I do this, etc. And it's like little paragraphs of fictitious human background. And what you do when you do synthetic data generation is not only like, oh, complete this task and do it in this way, but also imagine you're describing it to this person, and you put in this information. And now you're forcing it to explore more of the space, and you're getting some entropy. So I think you have to be just very careful to inject the entropy, maintain the distribution. And that's the hard part, that I think maybe people aren't sufficiently appreciating as much in general. So I think basically synthetic data, absolutely the future. We're not going to run out of data, is my impression. I just think you have to be careful.
20:27 Sarah Guo? What do you think we are learning now about human cognition from this research?
20:31 Speaker 02 I don't know if we're learning.
20:31 Sarah Guo? One could argue that figuring out the shape of reasoning traces we want, for example, is instructive to actually understand how the brain works.
20:38 Speaker 02 I would be careful with those analogies, but in general, I do think that it's a very different kind of thing. But I do think that there are some analogies you can draw. So as an example, I think transformers are actually better than the human brain in a bunch of ways. I think they're actually a lot more efficient system. And the reason they don't work as good as the human brain is mostly data issue, roughly speaking. It was the first order approximation, I would say. And actually, as an example, transformer memorizing sequences is so much better than humans. Like, if you give it a sequence and you do a single forward backward pass on that sequence, then if you give it the first few elements, it will complete the rest of the sequence. It memorized that sequence. And it's so good at it. If you gave a human a single presentation of a sequence, there's no way that you can remember that. And so the transformers actually, I do think there's a good chance that the gradient-based optimization,
21:22 Speaker 02 the forward backward update that we do all the time for training neural nets, is actually more efficient than the brain in some ways. And these models are better, they're just not yet ready to shine. But in a bunch of cognitive sort of aspects, I think they might come out.
21:33 Sarah Guo? With the right inputs, they will be better.
21:35 Speaker 01 That's inherently true of computers for all sorts of applications, right? And putting memory to your point.
21:40 Speaker 02 Yeah, exactly. And I think human brains just have a lot of constraints. You know, the working memory is very small. I think transformers have a lot, lot bigger working memory and will continue to be the case. They are much more efficient learners. The human brains function under all kinds of constraints. It's not obvious that human brain is back propagation, right? It's not obvious how that would work. It's a very stochastic sort of dynamic system. It has all these constraints it works under, so ambient conditions, et cetera. So I do think that what we have is actually potentially better than the brain. And it's just not there yet.
22:11 Speaker 01 How do you think about human augmentation with different AI systems over time? Do you think that's a likely direction? Do you think that's unlikely?
22:19 Speaker 02 Augmentation?
22:19 Speaker 01 Augmentation of people with AI models.
22:22 Speaker 02 Oh, of course. I mean, but in what sense maybe? I think in general, absolutely.
22:26 Speaker 01 Because I mean, there's an abstract version of it you're using as a tool. That's the external version. There's the, you know, the merger scenario. You know, a lot of people end up talking about.
22:34 Speaker 02 Yeah, yeah. I mean, we're already kind of merging. The thing is like, there's a, you know, there's the I/O bottleneck.
22:39 Speaker 01 Sure.
22:39 Speaker 02 But for the most part, you know, at your fingertips, if you have any of these models.
22:42 Speaker 01 Yeah, but that's a little bit different because I mean, people have been making that argument for, I think, 40, 50 years where technological tools are just extension of human capabilities, right?
22:50 Speaker 02 Yeah, the computer is the bicycle for human mind, etc.
22:52 Speaker 01 Yeah, exactly. So but there's a subset of the AI community that thinks that, for example, the way that we subsume some potential conflict with future AI or something else would be through some form of.
23:03 Speaker 02 Yeah, like the Neuralink pitch, et cetera.
23:05 Speaker 01 Exactly.
23:06 Speaker 02 Yeah, I don't, I don't know what this merger looks like yet, but I can definitely see that you want to decrease the I/O to tool use. And I see this as kind of like an exocortex, well, building on top of our neocortex, right? And it's just the next layer. And it just turns out to be in the cloud, et cetera. But it is the next layer of the brain.
23:20 Speaker 01 Yeah, Accelerando, a book from the early 2000s, has a version of this where basically everything is substantiated in a set of goggles that are computationally attached to your brain that you wear. And then if you lose them, you must feel like you're losing a part of your person or memory.
23:34 Speaker 02 I think that's very likely, yeah. And today, the phone is already almost that. And I think it's going to get worse. When you put your techno stuff away from you, you're just like naked human in nature.
23:41 Sarah Guo? Well, you lose part of your intelligence. It's very anxiety inducing.
23:45 Speaker 01 A very, a very simple example of that is just maps, right? So a lot of people now have noticed can't actually navigate their city very well anymore because they're always using turn by turn direction.
23:54 Speaker 02 And if we have this, for example, like universal translator, which I don't think is too far away, like you'll lose the ability to speak to people who don't speak English if you just put your stuff away.
24:02 Sarah Guo? I'm very comfortable repurposing that part of my brain to do further research.
24:06 Speaker 02 I don't know if you saw the video of like the kid that has a magazine and is trying to like swipe on the magazine. What's fascinating to me about it is like this kid doesn't understand what comes with nature and what's technology. Technology on top of the nature. Because it's made so transparent. And I think this might look similar where people will just start assuming the tools. And then when you take them away, you realize like, I guess like people don't know what's technology and what's not. If you're wearing this thing that's always translating everyone or like doing stuff like that for you, then maybe people like lose the. It's basic cognitive abilities that may not exist. I think it exists, yeah. Like whoa, like by nature.
24:38 Sarah Guo? We're gonna specialize.
24:39 Speaker 02 You can't understand people who speak Spanish like at the helm.
24:44 Speaker 02 Or like when you go to objects like in Disney, all the objects are alive. And I think we are gonna potentially come to that kind of a world where why can't I talk to things? Like already today, you can talk to Alexa and you can ask her for things and so on. Yeah, I've seen some toy companies like that where they're basically trying to embed an LLM in a toys that can interact with a child. Yeah, like isn't this strange that when you go to a door, you can't just say open like what the hell? Another favorite example of that, I don't know if you saw either Demolition Man or iRobot. People make fun of the idea that you, yeah, like you can't just talk to things and what the hell.
25:14 Sarah Guo? If we're talking about a exocortex that feels like a pretty fundamentally important thing to democratize access to. How do you think like the current market structure of what's happening in LLM research? You know, there's a small number of large labs that actually have a shot at the next generation progressing training. Like, how does that translate to what people have access to in the future?
25:40 Speaker 02 So what you're kind of alluding to maybe is the state of the ecosystem, right? So we have kind of like an oligopoly of a few closed platforms. And then we have an open platform that is kind of like behind, so like Meta Llama, etc. And this is kind of like mirroring the open source kind of ecosystem. I do think that when this stuff starts to, when we start to think of it as like an exocortex. So there's the, there's a saying in crypto, which is like not your keys, not your, not your tokens. Like, is it the case that if it's like not your weights, not your brain? That's interesting because a company is effectively controlling your exocortex and therefore a big part of your. It starts to feel kind of invasive. If this is my exocortex.
26:11 Sarah Guo? I think people will care much more about ownership, yes.
26:13 Speaker 02 Like you're, yeah, you realize you're renting your brain. Like it seems strange to rent your brain.
26:18 Sarah Guo? The thought experiment is like, are you willing to give up ownership and control to rent a better brain? Because I am, yeah.
26:24 Speaker 02 So I think that's the trade off, I think. We'll see how that works, but maybe it's possible to like. By default, use the closed versions because they're amazing. But you have a fallback in various scenarios. And I think that's kind of like the way things are shaping up today even. Right, like when APIs go down on some of the closed source providers, people start to implement fallbacks to like the open ecosystems, for example, that they fully control and they're empowered by that, right? So maybe that's just the extension of what it'll look like for the brain. Is you fall back on the open source stuff should anything happen. But most of the time, you actually.
26:54 Sarah Guo? So it's quite important that the open source stuff continues to progress.
26:57 Speaker 02 I think so, 100%. And this is not like an obvious point or something that people maybe agree on right now. But I think 100%.
27:03 Speaker 01 I guess one thing I've been wondering about a little bit is, what is the smallest performant model that you can get to in some sense? Either in parameter size or however you want to think about it. And some a little bit curious about your view. Because you've thought a lot about the distillation, small models, you know?
27:18 Speaker 02 I think it can be surprisingly small. And I do think that the current models are wasting a ton of capacity, remembering stuff that doesn't matter. Like they remember SHA hashes, they remember like the ancient.
27:29 Sarah Guo? Because the dataset is not curated the best.
27:31 Speaker 02 Yeah, exactly. Like, and I think this will go away. And I think we just need to get to the cognitive core. And I think the cognitive core can be extremely small. And it's just this thing that thinks. And if it needs to look up information, it knows how to use different tools.
27:43 Sarah Guo? Is that like 3 billion parameters? Is that 20 billion?
27:46 Speaker 02 I think even a billion, billion suffices. We'll probably get to that point. And the models can be very, very small. And I think the reason they can be very small is fundamentally, I think, just like distillation works. It may be like the only thing I would say. Distillation works surprisingly well. A distillation is where you get a really big model or a huge amount of compute or something like that, supervising a very small model. And you can actually stuff a lot of capability into a very small model.
28:10 Speaker 01 Is there some sort of like mathematical representation of that? Or some information theoretical like formulation of that? Because it almost feels like you should be able to. Calculate that now in terms of what's the
28:22 Speaker 02 Maybe, maybe like one way to think about it is like, you know, we go back to like the internet dataset, which is what we're working with. The internet is like 0.001% cognition and like 99.99% of like information is just like, you know,
28:34 Sarah Guo? Garbage, yeah.
28:35 Speaker 02 And I think most of it is not useful to the thinking part and it's like
28:38 Speaker 01 Yeah, yeah. I guess maybe another way to frame the question is like, is there a mathematical representation of cognitive capability relative to model size? Or how do you capture cognition in terms of, you know, here's the min or max relative to what you're trying to accomplish? And maybe there's no good way to represent that. So I think maybe a billion parameters gets you sort of like a good cognitive core.
28:56 Speaker 02 Hmm, I think probably right. I think even one billion is too much. I don't know. We'll see.
29:01 Sarah Guo? It's very exciting, given if you think about, uh, well, you know, it's a question of like on an edge device versus on the cloud, but
29:08 Speaker 01 And also this raw cost of using the model and everything. Yeah, it's very exciting.
29:12 Sarah Guo? Right, but at less than a billion parameters, I have my exocortic cortex on a local device as well.
29:17 Speaker 02 Yeah, and then probably it's not a single model, right? Like, it's interesting to me to think about what this will actually play out like, um, because I do think you want to benefit from parallelization. You don't want to have a sequential process. You want to have a parallel process. And I think companies to some extent are also kind of like, uh, sick, um, parallelization of work. And but they, there's a hierarchy in a company because that's one way to, you know, you have the information processing and the reductions that need to happen within the organization for information. So I think we'll probably end up with, uh, companies for OLLMs. I think it's not unlikely to me that you have models of different capabilities specialized to various, uh, unique domains. Maybe there's a programmer, et cetera. And it will actually start to resemble companies to a very large extent. So you'll have the programmer and the
29:58 Speaker 02 program manager and, you know, similar kinds of roles of OLMs working in parallel and coming together and orchestrating computation on your behalf. So maybe it's not correct to think about. It's more like a swarm.
30:09 Speaker 01 Yeah, it feels like an ecosystem. It's like a biological ecosystem where you have specialized roles and niches.
30:14 Speaker 02 And I think it will start to resemble that.
30:15 Sarah Guo? You have automatic escalation to other parts of the swarm, depending on the difficulty of the problem and the specialization.
30:20 Speaker 02 Assuming the CEO is like a really brilliant, uh, cloud model.
30:24 Sarah Guo? Yeah.
30:24 Speaker 02 Uh, but the workers can be a lot cheaper, maybe even open source models or whatnot.
30:28 Sarah Guo? And my cost function's different from your cost function.
30:30 Speaker 02 Yeah, so, uh, that could be interesting.
30:32 Sarah Guo? You left OpenAI, you're working on education, uh, you've always been an educator. Like, why? Why do this?
30:38 Speaker 02 I would start with, I've always been an educator and I love learning and I love teaching. And so it's kind of just like a space that I've been very passionate about for a long time. And then the other thing is, I think one macro picture that's kind of driving me is, I think there's a lot of activity in like AI and I think most of it is to kind of like replace or displace people, I would say. It's in the theme of like sliding away the people. But I'm always more interested in anything that kind of empowers people. And I feel like I'm kind of on the high level like team human. And I'm interested in things that AI can do to empower people. And I don't want the future where people are kind of on a side of automation. I want people to be very in an empowered state and I want them to be amazing, even much more amazing than today. And then other aspects that I find very interesting is like,
31:21 Speaker 02 how far can a person go if they have the perfect tutor for all the subjects? And I think people could go really far if they had the perfect curriculum for anything. And I think we see that with, you know, if you, if some rich people maybe have tutors and they do actually go really far. And so I think we can approach that with AI or even Luxor Passit.
31:39 Speaker 01 There's very clear literature on that actually from the eighties, right? Where one-on-one tutoring, I think, helps people get one standard deviation better. Two, Bloom. Is it two? Yeah, it's the Bloom stuff. Yeah, exactly. There's a lot of really interesting precedents on that. How do you actually view that as substantiating through the lens of AI? Or what's the first types of products that will really help with that? Or, you know, because there's books like The Diamond Age where they talk about the young ladies, Illustrated Primer and all that kind of stuff.
32:02 Speaker 02 So I would say I'm definitely inspired by aspects of it. So like in practice, what I'm doing is trying to currently build a single course. And I want it to be just like the course you would go to if you want to learn AI. I think the problem with basically is like, I've already taught courses, like I taught 231N at Stanford and that was the first deep learning class and was pretty successful. But the question is like, how do you actually like really scale these classes? Like, how do you make it so that your target audience is maybe like eight billion people on earth? And they're all speaking different languages and they're all different capability levels, et cetera. So and a single teacher doesn't scale to that audience. And so the question is how to use AI to sort of like do the scaling of a really good teacher. And so the way I'm thinking about it is the teacher is kind of doing a lot of the course creation in the curriculum because currently at current AI capability, I don't
32:45 Speaker 02 think the models are good enough to create a good course, but I think they're good to become the front end to the student and interpret the course to them. And so basically, the teacher doesn't go to the people and the teacher is not the front end anymore. The teacher is on the back end designing the materials in the course and the AI is the front end and it can speak all the different languages and it's kind of like takes you through the course.
33:06 Sarah Guo? Should I think of that as like the TA type experience or is that not a good analogy here?
33:11 Speaker 02 That is like one way I'm thinking about it is it's AI TA. I'm mostly thinking of it as like this front end to the student and it's the thing that's actually interfacing with the student and taking them through the course. And I think that's tractable today and it just doesn't exist and I think it can be really good. And then over time, as the capability increases, you would potentially refactor the setup in various ways. I like to find things where like the AI capability today and having a good model of it and I think a lot of companies that maybe don't quite understand intuitively where the capability is today and then they end up kind of like building things that are kind of like too ahead of what's available or maybe not ambitious enough. And so I think, I do think that this is kind of a sweet part of what's possible and also really interesting and exciting, so.
33:52 Sarah Guo? I want to go back to something you said that I think is very inspiring, especially coming from like your background and understanding of where exactly we are in research, which is essentially like we do not know what the limits of human performance from a learning perspective are, given much better tooling. And I think there's like a very easy analogy to, we just had the Olympics like a month ago, right? And you know, a runner and it's the very best mile time or pick any sport today is much better than it was. Putting aside performance enhancing drugs like 10 years ago, just because like you start training earlier, you have a very different program, we have much better scientific understanding, we have technique, we have gear. The fact that you believe like we can get much further as humans if we're starting with like the tooling and the curriculum is amazing.
34:33 Speaker 02 Yeah, I think we haven't even scratched like what's possible at all. So I think there's like two dimensions basically to it is number one is the globalization dimension of like I want everyone to have a really good education, but the other one is like how far can a single person go? I think both of those are very interesting and exciting.
34:44 Speaker 01 Usually when people talk about 101 learning, they talk about the adaptive aspect of it, where you're a challenging person at the level that they're at. Do you think you can do that with AI today? Or is that something for the future? And it's more today, it's about reach and multiple languages.
34:56 Speaker 02 I think the low-hanging fruit is things like, for example, different languages, super low-hanging fruit. I think the current models are actually really good at translation, basically, and can target the material and translate it like at the spot. So I think a lot of things are low-hanging fruit. This adaptability to a person's background, I think is like not at the low-hanging fruit, but I don't think it's like too high up or too much away. But there is something you definitely want, because not everyone is coming in with the same background. And also, what's really helpful is like, if you're familiar with some other disciplines in the past, then it's really useful to make analogies to the things you know. And that's extremely powerful in education. So that's definitely a dimension you want to take advantage of. But I think that starts to get to the point where it's like not obvious and needs some work. I think like the easy version of it is not too far, where you can imagine just prompting the model. It's like, oh, hey, I know physics or I know this. And you probably get something.
35:40 Speaker 02 But I guess what I'm talking about is something that actually works, not something that like you can demo and work sometimes. So I just mean like it actually really works and in a way a person would.
35:49 Speaker 01 Yeah. And that's the reason why I was talking about adaptability, because also people learn at different rates or certain things they find challenging that others don't or vice versa. And so it's a little bit of how do you modulate relative to that context? And I guess you could have some reintroduction of what the person is good or bad at into the model over time.
36:03 Speaker 02 That's the thing with AI. I feel like a lot of them, a lot of these capabilities are just kind of like prompt away. So you always get like demos, but like, do you actually get a product? You know what I mean? So in this sense, I would say the demo is near, but the product is far.
36:16 Speaker 01 So one thing we were talking about earlier, which I think is really interesting, is sort of lineages. That happens in the research community, where you come from certain labs and everybody gossips about being from each other's labs. I think a very high proportion of Nobel laureates actually used to work in a former Nobel laureates lab. So there's some propagation of, I don't know if it's culture or knowledge or branding or what. And in an AI education-centric world. How do you maintain lineage or does it not matter? Or how do you think about those aspects of propagation of network and knowledge?
36:44 Speaker 02 I don't actually want to live in a world where lineage matters too much, right? So I'm hoping that AI can help you destroy that structure a little bit. It feels like kind of gatekeeping by some finite scarce resource, which is like, oh, there's a finite number of people who have this lineage, et cetera. So I feel like it's a little bit of that aspect. So I'm hoping it can destroy that.
37:00 Sarah Guo? It's definitely one piece, like actual learning, one piece pedigree.
37:03 Speaker 01 Yeah, well, it's also the aggregation of, it's a cluster effect, right? It's like, why is all of the, or much of the AI community in the Bay Area? Or why is most of the fintech community in New York?
37:13 Speaker 02 Yeah.
37:14 Speaker 01 And so I think a lot of it is also just you're clustering really smart people with common interests and beliefs. And then they kind of propagate from that common core. And then they share knowledge in an interesting way. You've got to agree, a lot of that behavior has shifted online to some extent, particularly for younger people.
37:26 Speaker 02 I think one aspect of it is kind of like the educational aspect, where like, if you're part of a community today, you're getting a ton of education and apprenticeship, et cetera, which is extremely helpful and gets you to a point of empowered state in that area. I think the other piece of it is like the cultural aspect of what you're motivated by and what you want to work on. What does the culture prize? And what are they putting on pedestal? And what do they kind of like worship, basically? So in academic world, for example, is the H-index. Everyone cares about the H-index, the amount of papers you publish, et cetera. And I was part of that community and I saw that. And I feel like now I've come to different places and there's different idols in all the different communities. And I think that has a massive impact of what people are motivated by and where they get their social status and what actually matters to them. I also was, I think, part of different communities, like growing up in Slovakia, also a very different environment. Being in Canada, also a very different environment.
38:09 Sarah Guo? What mattered there?
38:10 Speaker 02 So.
38:11 Sarah Guo? Thank you.
38:12 Speaker 01 Hockey.
38:15 Speaker 02 I would say as an example, I would say in Canada, I was in the University of Toronto and Toronto. I don't think it's a very entrepreneurial pill environment. It doesn't even occur to you that you should be starting companies. I mean, it's not something that people are doing. You don't know friends who are doing it. You don't know that you're supposed to be looking up to it. People aren't like reading books about all the founders and talking about them. It's just not a thing you aspire to or care about. And what everyone is talking about, oh, is where are you getting your internship? Where are you going to work afterwards? And it's just accepted that there's a bunch of, there's a thick set of companies that you are supposed to pick from and just allow. ...yourself with one of them and that's like what you look up to or something like that. So these cultural aspects are extremely strong and maybe actually the dominant variable. Because I almost feel like today already the education aspects I think are the easier one. Like a ton of stuff is already available, etc.
39:00 Speaker 02 So I think mostly it's a cultural aspect that you're part of.
39:02 Sarah Guo? So on this point, like one thing you and I were talking about a few weeks ago is, and I think you also posted online about this. There's a difference between learning and entertainment. And learning is actually supposed to be hard. And I think it relates to this question of like, you know, status. And what like, status is a great motivator, like who the idol is. How much do you think you can change in terms of motivation through systems like this? If that's like a blocking factor? Are you focused on give people the resources such that they can get as far as possible in the sequence for their own capability as they can, like further than any other point in history, already inspirational? Or do you actually want to change how many people want to learn?
39:47 Sarah Guo? Or at least bring themselves down the path? Want is a loaded word.
39:50 Speaker 02 I would say like I want to make it much easier to learn. And then maybe it is possible that maybe people don't want to learn. I mean, today, for example, people want to learn for practical reasons, right? Like they want to get a job, etc., which makes total sense. So in a pre AGI society, education is useful. And I think people will be motivated by that because they're they're climbing up the ladder economically, etc.
40:09 Sarah Guo? Oh, but in the post AGI society, we're just all gonna be...
40:11 Speaker 02 Society, I think education is entertainment to a much larger extent.
40:16 Sarah Guo? Including like successful outcomes, education, right? Not just letting the content wash over you.
40:22 Speaker 02 Yes, I think so.
40:23 Sarah Guo? Outcomes being like understanding, learning, be able to contribute new knowledge or however you define it.
40:27 Speaker 01 I think it's not an accident that if you go back 200 years, 200 years, the people who were doing science were nobility or people of wealth.
40:34 Sarah Guo? We will all be nobility, learning with Andre, yeah.
40:37 Speaker 02 I do think that I see it very much equivalent to your quote earlier. Like I feel like learning something is kind of like going to the gym, but for the brain, right? Like it feels like going to the gym. I'm going to the gym is fun. People like to lift, etc. Some people don't go to the gym.
40:50 Sarah Guo? No, no, no, some people do, but it is, it takes effort, yeah.
40:53 Speaker 02 Yeah, it takes effort, but it's effortful, but it's also kind of fun. And you also have a payoff, like you feel good about your... In various ways, right? And I think education is basically equivalent to that. So that's what I mean when I say education should not be fun, etc. I mean, it is kind of fun, but it's like a specific kind of fun, I suppose, right? I do think that maybe in a post AGI world, what I would hope happens is people actually, they do go to the gym a lot, not just physically, but also mentally. And it's something that we look up to as being highly educated and also, you know, just, yeah.
41:21 Sarah Guo? Can I ask you one last question about Eureka? Just because I think it would be interesting people. Like, who is the audience for the first course?
41:27 Speaker 02 The audience for the first course? I'm mostly thinking of this as like an undergrad level course. So if you're doing undergrad in technical area, I think that would be kind of the ideal audience. I do think that what we're seeing now is we have this like antiquated concept of education, where you go through school and then you graduate and go to work, right? Obviously, this will totally break down, especially in a society that's turning over so quickly. People are gonna come back to school a lot more frequently as the technology changes very, very quickly. So it is kind of like undergrad level, but I would say like anyone at that level at any age is kind of like in scope. I think it will be very diverse in age as an example. But I think it is mostly like people who are technical and mostly want to, mostly actually want to understand it to, you know, a good amount.
42:09 Sarah Guo? When can they take the course?
42:10 Speaker 02 I was hoping it would be late this year. I do have a lot of distractions that are piling on, but I think probably early next year is kind of like the timeline. Yeah, I'm trying to make it very, very good. And yeah, it just takes time to get there, so.
42:23 Speaker 01 I have one last question, actually, that's pseudo-related to that. If you have little kids today, what do you think they should study in order to have a useful future?
42:31 Speaker 02 There's a correct answer in my mind. And the correct answer is mostly like, I would say like math, physics, CS kind of disciplines. And the reason I say that is because I think it helps for just thinking skills. It's just like the best thinking skill core is my opinion. Of course, I have a specific background, etc., so I would, I would think this, but, but that's just my view on it. I think like me taking physics classes and all these other classes just like shaped the way I think. And I think it's very useful for problem solving in general, etc. And so if we're in this world where pre AGI, this is gonna be useful. Post AGI, you still want empowered humans who can function in any arbitrary capacity. And so I just think that this is just the correct answer for people and what they should be doing and taking. And it's either useful or it's good. And so I just think it's the right answer.
43:16 Speaker 02 And I think a lot of the other stuff you can tack on a bit later. But the critical period where people have a lot of time and they have a lot of kind of like attention and time, I think should be mostly spent on doing these kinds of simple manipulation heavy tasks and workloads, not memory heavy tasks and workloads.
43:32 Speaker 01 I did a math degree and I felt like there was a new groove being carved into my brain as I was doing that.
43:37 Sarah Guo? And it's a harder groove to carve later.
43:39 Speaker 02 And I would of course put in a bunch of other stuff as well. Like I'm not opposed to all the other disciplines, et cetera. I think it's actually beautiful to have a large diversity of things, but I do think 80% of it should be something like this.
43:48 Sarah Guo? One, we're not efficient memorizers compared to our tools. Thank you for doing this. It was so much fun.
43:54 Speaker 02 Yes, great to be here.
43:57 Sarah Guo? Find us on Twitter at NoPriorsPod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com.
Previous Building toward a bright post-AGI future with Eric Steinberger from Magic.dev Next Future of LLM Markets, Consolidation, and Small Models with Sarah and Elad Page updated October 26, 2026 · transcript and summary are machine-generated