Menu
Sign In Search Podcasts Charts Entities Add Podcast API Pricing

Andrej Karpathy

👤 Person
1503 appearances

Podcast Appearances

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, first of all, thank you for having me here.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm excited to be here.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So the quote that you've just mentioned, it's the decade of agents, that's actually a reaction to an existing, pre-existing quote, I should say, where I think some of the labs, I'm not actually sure who said this, but they were alluding to this being the year of agents with respect to LLMs and how they were going to evolve.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I was triggered by that because I feel like there's some over-predictions going on in the industry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And in my mind, this is really a lot more accurately described as the decade of agents.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And we have some very early agents that are actually extremely impressive and that I use daily, you know, Cloud and Codex and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I still feel like there's so much work to be done.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I think my reaction is like, we'll be working with these things for a decade.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're going to get better and it's going to be wonderful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think I was just reacting to the timelines, I suppose, of the

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

of the implication.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And what do you think will take a decade to accomplish?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What are the bottlenecks?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, actually make it work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So in my mind, I mean, when you're talking about an agent, I guess, or what the labs have in mind and what maybe I have in mind as well, is it's, you should think of it almost like an employee or like an intern that you would hire to work with you.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So for example, you work with some employees here.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

When would you prefer to have an agent like Cloud or Codex do that work?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like currently, of course they can't.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What would it take for them to be able to do that?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Why don't you do it today?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And the reason you don't do it today is because they just don't work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, like, they don't have enough intelligence.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're not multimodal enough.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They can't do computer use and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they don't do a lot of the things that you've alluded to earlier.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You know, they don't have continual learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can't just tell them something and they'll remember it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they're just cognitively lacking and it's just not working.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I just think that it will take about a decade to work through all of those issues.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I guess this is where you get into like a bit of, I guess, my own intuition a little bit.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And also just kind of doing a bit of an extrapolation with respect to my own experience in the field, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess I've been in AI for...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

almost two decades.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, it's going to be maybe 15 years or so.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Not that long.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You had Richard Sutton here who was around, of course, for much longer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do have about 15 years of experience of people making predictions of seeing how they actually turned out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And also, I was in the industry for a while and I was in research and I worked in the industry for a while.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, I guess I kind of have just a general intuition that I have left from that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I feel like the problems are tractable.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're surmountable.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But they're still difficult.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And if I just average it out, it just kind of feels like a ticket, I guess, to me.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I mean, that's a giant question because, of course, you're talking about 15 years of stuff that happened.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, AI is actually like so wonderful because there have been a number of, I would say, seismic shifts that were like the entire field has sort of like suddenly looked a different way, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I guess I've maybe lived through two or three of those.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I still think there will continue to be some because they come with some kind of like almost surprising irregularity.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, when my career began, of course, like when I started to work on deep learning, when I became interested in deep learning, this was just kind of like by chance of being right next to Jeff Hinton at University of Toronto.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And Jeff Hinton, of course, is kind of like the godfather figure of AI.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And he was training all these neural networks, and I thought it was incredible and interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But this was not like the main thing that everyone in AI was doing by far.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This was a niche little subject on the side.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's kind of maybe like the first like dramatic sort of seismic shift that came with the AlexNet and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would say like AlexNet sort of reoriented everyone and everyone started to train neural networks, but it was still like very like per task, per specific task.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe I have an image classifier or I have a neural machine translator or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And people became very slowly actually interested in basically kind of agents, I would say.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And people started to think, okay, well, maybe we have a check mark next to the visual cortex or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But what about the other parts of the brain?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How can we get an actual like full agent or in full entity that can actually interact in the world?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I would say the Atari sort of deep reinforcement learning shift in 2013 or so was part of that early effort of agents in my mind, because it was an attempt to try to get agents that not just perceive the world, but also take actions and interact and get rewards from environments.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And at the time this was Atari games, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of feel like that was a misstep, actually.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it was a misstep that actually even the early OpenAI that I was a part of, of course, kind of adopted.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because at that time, the zeitgeist was reinforcement learning environments, games, game playing, beat games, get lots of different types of games.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And OpenAI was doing a lot of that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that was maybe like another prominent part of, I would say, AI where maybe for two or three or four years, everyone was doing reinforcement learning on games.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And basically, that was a little bit of a misstep.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And what I was trying to do at OpenAI actually is like, I was always a little bit suspicious of games as being like this thing that would actually lead to AGI because in my mind, you want something like an accountant or like something that's actually interacting with the real world.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I just didn't see how games kind of like add up to it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so my project at OpenAI, for example, was within the scope of the Universe project on an agent that was using keyboard and mouse to operate web pages.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I really wanted to have something that like interacts with, you know, the actual digital world that can do knowledge work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it just so turns out that this was extremely early, way too early.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So early that we shouldn't have been working on that, you know, because if you're just stumbling your way around and keyboard mashing and mouse clicking and trying to get rewards in these environments, your reward is too sparse and you just won't learn and you're going to burn a forest computing and you're never actually going to get something off the ground.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so what you're missing is this power of representation in the neural network.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so, for example, today, people are training those computer-using agents, but they're doing it on top of a large language model.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so you actually have to get the language model first.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You have to get the representations first.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you have to do that by all the pre-training and all the LLM stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I kind of feel like maybe, loosely speaking, it was like people keep maybe trying to get the full thing too early a few times, where people really try to go after agents too early, I would say.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that was Atari and Universe, and even my own experience.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you actually have to do some things first before you sort of get to those agents.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And maybe now the agents are a lot more competent, but maybe we're still missing sort of some parts of that stack.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I would say maybe those are like the three major buckets of what people were doing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Training neural nets per tasks, trying to the first round of agents, and then maybe the LLMs and actually seeking the representation power of the neural networks before you tack on everything else on top.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, so Sutton was on your podcast, and I saw the podcast, and I had a write-up about that podcast almost that gets into a little bit of how I see things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of feel like I'm very careful to make analogies to animals because they came about by a very different optimization process.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Animals are evolved, and they actually come with a huge amount of hardware that's built in.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And when, for example, my example in the post was the zebra.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A zebra gets born, and a few minutes later, it's running around and following its mother.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's an extremely complicated thing to do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's not reinforcement learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's something that's baked in.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And evolution obviously has some way of encoding the weights of our neural nets in ATCGs.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I have no idea how that works, but it apparently works.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I kind of feel like brains just came from a very different process.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I'm very hesitant to take inspiration from it because we're not actually running that process.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So in my post, I kind of said, we're not actually building animals.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're building ghosts or spirits or whatever people want to call it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're not doing training by evolution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're doing training by basically imitation of humans and the data that they've put on the internet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so you end up with these like sort of ethereal spirit entities because they're fully digital and they're kind of like mimicking humans.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's a different kind of intelligence.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like if you imagine a space of intelligences, we're starting off at a different point almost.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're not really building animals.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think it's also possible to make them a bit more animal-like over time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think we should be doing that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I kind of feel like, sorry, just I guess one more point is, I do feel like Sutton basically has a very, like his framework is like we want to build animals.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I actually think that would be wonderful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If we can get that to work, that would be amazing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If there was a single like algorithm that you can just, you know, run on the internet and it learns everything, that would be incredible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I almost suspect that I'm not actually sure that it exists.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's certainly actually not what animals do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

because animals have this outer loop of evolution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And a lot of what looks like learning is actually a lot more maturation of the brain.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think there's actually very little reinforcement learning for animals.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think a lot of the reinforcement learning is actually more like motor tasks.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's not intelligence tasks.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I actually kind of think humans don't actually really use RL, roughly speaking is what I would say.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A lot of the reinforcement learning in my perspective would be things that are a lot more like motor-like, like simple kind of like tasks, throwing a hoop, something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I don't think that humans use reinforcement learning for a lot of intelligence tasks like problem solving and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That doesn't mean we shouldn't do that for research, but I just feel like that's what animals do or don't.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think so.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would agree with you that there's some miraculous compression going on, because obviously the weights of the neural net are not stored in ATCGs.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's some kind of a dramatic compression, and there's some kind of learning algorithms encoded that take over and do some of the learning online.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I definitely agree with you on that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Basically, I would say I'm a lot more kind of like practically minded.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't come at it from the perspective of like, let's build animals.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I come from the perspective of like, let's build useful things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I have a hard hat on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I'm just observing that, look, we're not going to do evolution, because I don't know how to do that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But it does turn out we can build these ghost spirit-like entities by imitating internet documents.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This works.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's actually kind of like, it's a way to bring you up to something that has a lot of sort of built-in knowledge and intelligence in some way, similar to maybe what evolution has done.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that's why I kind of call pre-training this kind of like crappy evolution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's like the practically possible version with our technology and what we have available to us to get to a starting point where we can actually do things like reinforcement learning and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So it's subtle, and I think you're right to push back on it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But basically, the thing that pre-training is doing, so you're basically getting the next token predictor over the internet, and you're training that into a neural net.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's doing two things actually that are kind of like unrelated.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Number one, it's picking up all this knowledge, as I call it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Number two, it's actually becoming intelligent.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

By observing the algorithmic patterns in the internet, it actually kind of like boots up all these like little circuits and algorithms inside the neural net to do things like in-context learning and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And actually, you don't actually need or want the knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I actually think that's probably actually holding back the neural networks overall, because it's actually like getting them to rely on the knowledge a little too much sometimes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

For example, I kind of feel like agents, one thing they're not very good at is going off the data manifold of what exists on the internet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If they had less knowledge or less memory, actually maybe they would be better.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so what I think we have to do kind of going forward, and this would be part of the research paradigms, is actually think we need to start, we need to figure out ways to remove some of the knowledge and to keep what I call this cognitive core.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's this like intelligent entity that is kind of stripped from knowledge but contains the algorithms and contains the magic, you know, of intelligence and problem solving and the strategies of it and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think I'm hesitant to say that in-context learning is not doing gradient descent because, I mean, it's not doing explicit gradient descent, but I still think that, so in-context learning, basically, it's pattern completion within a token window, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it just turns out that there's a huge amount of patterns on the internet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so you're right, the model kind of like learns to complete the pattern, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's inside the weights.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The weights of the neural network are trying to discover patterns and complete the pattern.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's some kind of an adaptation that happens inside the neural network, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Which is kind of magical and just falls out from internet just because there's a lot of patterns.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I will say that...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There have been some papers that I thought were interesting that actually look at the mechanisms behind in-context learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I do think it's possible that in-context learning actually runs a small gradient descent loop internally in the layers of the neural network.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I recall one paper in particular where they were doing linear regression, actually, using in-context learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So basically, your inputs into the neural network are XY pairs.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

x, y, x, y, x, y that happened to be on the line.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then you do x and you expect the y. And the neural network, when you train it in this way, actually does do linear regression.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And normally when you would run linear regression, you have a small gradient descent optimizer that basically looks at x, y, looks at an error, calculates the gradient of the weights, and does the update a few times.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It just turns out that when they looked at the weights of that in-context learning algorithm,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

they actually found some analogies to gradient descent mechanics.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

In fact, I think even the paper was stronger because they actually hard-coded the weights of a neural network to do gradient descent through attention and all the internals of the neural network.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess that's just my only pushback is that who knows how in-context learning works, but I actually think that it's probably doing a little bit of some kind of funky gradient descent internally, and that I think that that's possible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess I was only pushing back on you're saying it's not doing in-context learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Who knows what it's doing, but it's probably maybe doing something similar to it, but we don't know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think I kind of agree.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, the way I usually put this is that anything that happens during the training of the neural network, the knowledge is only kind of like a hazy recollection of what happened in the training time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's because the compression is dramatic.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're taking 15 trillion tokens and you're compressing it to just your final network of a few billion parameters.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So obviously it's a massive amount of compression going on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I kind of refer to it as like a hazy recollection of the internet documents.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Whereas anything that happens in the context window of the neural network, you're plugging all the tokens and it's building up all this KV cache representation, is very directly accessible to the neural net.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I compare the KV cache and the stuff that happens at test time to more like a working memory.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like all the stuff that's in the context window is very directly accessible to the neural net.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So there's always like these...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

almost surprising analogies between LLMs and humans.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I find them kind of surprising because we're not trying to build a human brain, of course, just directly.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're just finding that this works and we're doing it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do think that...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Anything that's in the weights, it's kind of like a hazy recollection of what you read a year ago.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Anything that you give it as a context at test time is directly in the working memory.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think that's a very powerful analogy to think through things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So when you, for example, go to an LLM and you ask it about some book and what happened in it, like Nick Lane's book or something like that, the LLM will often give you some stuff, which is roughly correct.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But if you give it the full chapter and ask it questions, you're going to get much better results because it's now loaded in the working memory of the model.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I basically agree with your very long way of saying that I kind of agree, and that's why.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I almost feel like just a lot of it still.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe one way to think about it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know if this is the best way, but I almost kind of feel like, again, making these analogies, imperfect as they are.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We've stumbled by with the transformer neural network, which is extremely powerful, very general.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can train transformers on audio or video or text or whatever you want, and it just learns patterns, and they're very powerful, and it works really well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That, to me, almost indicates that this is kind of like some piece of cortical tissue.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because the cortex is famously very plastic as well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can rewire, you know, parts of brains.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there was the slightly gruesome experiments with rewiring, like, visual cortex to the auditory cortex, and this animal, like, learned fine, etc.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think that this is kind of like cortical tissue.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think when we're doing reasoning and planning inside the neural networks, so basically doing reasoning traces for thinking models, that's kind of like the prefrontal cortex.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I think maybe those are like little check marks, but I still think there's many brain parts and nuclei that are not explored.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe, for example, there's a basal ganglia doing a bit of reinforcement learning when we fine tune the models on reinforcement learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But, you know, whereas like the hippocampus, not obvious what that would be.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

some parts are probably not important.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe the cerebellum is, like, not important to cognition, it's thought, so maybe we can skip some of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I still think there's, for example, the amygdala, all the emotions and instincts.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's probably, like, a bunch of other nuclei in the brain that are very ancient that I don't think we've, like, really replicated.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually know that we should be pursuing, you know, the building of an analog of human brain.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm, again, an engineer, mostly at heart.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still feel like maybe another way to answer the question is you're not going to hire this thing as an intern.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's missing a lot of it's because it comes with a lot of these cognitive deficits that we all intuitively feel when we talk to the models.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so it's just like not fully there yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can look at it as like not all the brain parts are checked off yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know that I fully resonate with that because I feel like these models, when you boot them up and they have zero tokens in the window, they're always like restarting from scratch where they were.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I don't actually know in that worldview what it looks like because, again, maybe making some analogies to humans just because I think it's roughly concrete and kind of interesting to think through.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I feel like when I'm awake, I'm building up a context window of stuff that's happening during the day.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I feel like when I go to sleep, something magical happens where I don't actually think that that context window stays around.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think there's some process of distillation into weights of my brain.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this happens during sleep and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We don't have an equivalent of that in large language models.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's, to me, more adjacent to when you talk about continual learning and so on as absent.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

These models don't really have this distillation phase of taking what happened, analyzing it, obsessively thinking through it, basically doing some kind of a synthetic data generation process and distilling it back into the weights, and maybe having a specific neural net per person.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe it's a LoRa, it's not a full...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, it's not a full-weight neural network.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's just some of the small sparse subset of the weights are changed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But basically, we do want to create ways of creating these individuals that have very long contacts.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's not only remaining in the contacts window because the contacts windows grow very, very long.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe we have some very elaborate sparse attention over it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I still think that humans obviously have some process for distilling some of that knowledge into the weights.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're missing it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I do also think that humans have some kind of a very elaborate sparse attention scheme, which I think we're starting to see some early hints of.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So DeepSeek v3.2 just came out, and I saw that they have like a sparse attention as an example.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this is one way to have very, very long context windows.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I almost feel like we are redoing a lot of the...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

cognitive tricks that evolution came up with through a very different process.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But we're, I think, going to converge on a similar architecture cognitively.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, the way I like to think about it is, okay, let's translation invariance in time, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So 10 years ago, where were we?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

2015, we had convolutional neural networks primarily.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Residual networks just came out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So remarkably similar, I guess, but quite a bit different still.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, Transformer was not around.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You know, all these sort of like more modern tweaks on the Transformer were not around.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe some of the things that we can bet on, I think, in 10 years by translational sort of equivariance is we're still training giant neural networks with forward, backward, pass, and update through gradient descent.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But maybe it looks a little bit different.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's just everything is much bigger.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Actually, recently, I also went back all the way to 1989, which was kind of a fun exercise for me a few years ago, because I was reproducing Jan LeCun's 1989 convolutional network, which was the first neural network I'm aware of trained via gradient descent, like modern neural network trained gradient descent on digit recognition.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I was just interested in, okay, how can I modernize this?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How much of this is algorithms?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How much of this is data?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How much of this progress is compute and systems?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I was able to very quickly like half the learning rate, just knowing by time travel by 33 years.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So if I time travel by algorithms to 33 years, I could adjust what Yann LeCun did in 1989, and I could basically half the learning, half the error.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But to get further gains, I had to add a lot more data.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I had to 10x the training set.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I had to actually add more computational optimizations.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I had to basically train for much longer with dropout and other regularization techniques.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so it's almost like all these things have to improve simultaneously.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So we're probably going to have a lot more data.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're probably going to have a lot better hardware.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're probably going to have a lot better kernels and software.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're probably going to have better algorithms.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And all of those, it's almost like no one of them is winning too much.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

All of them are surprisingly equal.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this has kind of been the trend for a while.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess to answer maybe your question, I expect differences algorithmically to what's happening today.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do also expect that some of the things that have stuck around for a very long time will probably still be there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's probably still a giant neural network trained with gradient descent.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That would be my guess.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I guess what was shocking to me is everything needs to improve across the board.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Architecture, optimizer, loss function, and also has improved across the board forever.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I kind of expect all those changes to be alive and well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Building NanoChat?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So NanoChat is a kind of a repository I released.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Was it yesterday or the day before?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I can't remember.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We can see this lead generation that went into the... Well, it's just trying to be a...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's trying to be the simplest, complete repository that covers the whole pipeline end-to-end of building a ChatGPT clone.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so, you know, you have all of the steps, not just any individual step, which is a bunch of... I worked on all the individual steps sort of in the past and released small pieces of code that kind of show you how that's done in algorithmic sense in like simple code.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But this kind of handles all the entire pipeline.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think in terms of learning, it's not so much, I don't know that I actually found something that I learned from it necessarily.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I kind of already had in my mind as like how you build it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this is just a process of mechanically building it and making it clean enough and so that people can actually learn from it and that they find it useful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would probably say, so basically it's about 1,000 lines of code that takes you through the entire pipeline.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would probably put it on the right monitor, like if you have two monitors, you put it on the right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you want to build it from scratch.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You build it from start.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're not allowed to copy-paste.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're allowed to reference, you're not allowed to copy-paste.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe that's how I would do it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I also think the repository by itself, it is like a pretty large beast.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, when you write this code, you don't go from top to bottom.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You go from chunks and you grow the chunks.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that information is absent, like you wouldn't know where to start.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I think it's not just a final repository that's needed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's like the building of the repository, which is a complicated chunk growing process.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that part is not there yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would love to actually add that probably later this week or something in some way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Either it's probably a video or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But maybe, roughly speaking, that's what I would try to do is build the stuff yourself, but don't allow yourself copy-paste.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that there's two types of knowledge almost.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like there's the high-level surface knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But the thing is that when you actually build something from scratch, you're forced to come to terms with what you don't actually understand and you don't know that you don't understand it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it always leads to a deeper understanding.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's like just the only way to build.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's like if I can't build it, I don't understand it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Is that a Feynman quote, I believe, or something along those lines?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I 100%, I've always believed this very strongly because there's all these like micro things that are just not properly arranged and you don't really have the knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You just think you have the knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So don't write blog posts.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Don't do slides.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Don't do any of that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like build a code, arrange it, get it to work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's the only way to go.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Otherwise you're missing knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So the repository, I guess I built it over a period of a bit more than a month.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I would say there's like three major classes of how people interact with code right now.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Some people completely reject all of LLMs, and they are just writing by scratch.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think this is probably not the right thing to do anymore.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The intermediate part, which is where I am, is you still write a lot of things from scratch, but you use the autocomplete that's basically available now from these models.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So when you start writing out a little piece of it, it will autocomplete for you, and you can just tap through, and most of the time it's correct.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Sometimes it's not, and you edit it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But you're still very much the architect of what you're writing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then there's the vibe coding.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You know, hi, please implement this or that, you know, enter, and then let the model do it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's the agents.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do feel like the agents work in very specific settings, and I would use them in specific settings.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But again, these are all tools available to you, and you have to learn what they're good at and what they're not good at and when to use them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So the agents are actually pretty good, for example, if you're doing boilerplate stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Boilerplate code that's just copy-based stuff, they're very good at that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're very good at stuff that occurs very often on the intranet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

because there's lots of examples of it in the training sets of these models.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So there's features of things where the models will do very well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would say NanoChat is not an example of this, because it's a fairly unique repository.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's not that much code, I think, in the way that I've structured it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's not boilerplate code.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's like actually like intellectually intense code almost.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And everything has to be very precisely arranged.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And the models are always trying to, they kept trying to, I mean, they have so many cognitive deficits, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So one example, they keep trying to, they keep misunderstanding the code because they have too much memory from all the typical ways of doing things on the internet that I just wasn't adopting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So the models, for example, I mean, I don't know if I want to get into the full details, but they keep thinking I'm writing normal code and I'm not.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe one example.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe one example is, so the way to synchronize, so we have eight GPUs that are all doing forward backwards.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The way to synchronize gradients between them is to use a distributed data parallel container of PyTorch, which automatically does all the, as you're doing the backward, it will start communicating and synchronizing gradients.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I didn't use DDP because I didn't want to use it because it's not necessary.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I threw it out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I basically wrote my own synchronization routine that's inside the step of the optimizer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so the models were trying to get me to use the DDP container, and they were very concerned about, okay, this gets way too technical, but I wasn't using that container because I don't need it, and I have a custom implementation of something like it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, they couldn't get past that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

and then um they kept trying to like mess up the style like they're way too over defensive they make all these try catch statements they keep trying to make a production code base and i have a bunch of assumptions in my code and it's okay and uh and it's just like i don't need all this extra stuff in there and so i just kind of feel like they're bloating the code base they're bloating the complexity they keep misunderstanding they're using deprecated apis a bunch of times so it's total mess um

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

and it's just not that useful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I can go in and I can clean it up, but it's not that useful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I also feel like it's kind of annoying to have to, like, type out what I want in English because it's just too much typing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, if I just navigate to the part of the code that I want and I go where I know the code has to appear and I start typing out the first three letters, autocomplete gets it and just gives you the code.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I think it's... This is a very high-information bandwidth to specify what you want.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If you point to the code where you want it and you type out the first few pieces, and the model will complete it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess what I mean is...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think these models are good in certain parts of the stack.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I actually use the models a little bit in... There are two examples where I actually use the models that I think are illustrative.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

One was when I generated the report.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's actually more boilerplate-y.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I actually bytecoded partially some of that stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That was fine.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because it's not like mission-critical stuff and it works fine.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then the other part is when I was rewriting the tokenizer in Rust...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm actually not as good at Rust because I'm fairly new to Rust.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I was doing, there's a bit of vibe coding going on when I was writing some of the Rust code.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I had Python implementation that I fully understand and I'm just making sure I'm making a more efficient version of it and I have tests.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I feel safer doing that stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so basically they lower or like they increase accessibility to languages or paradigms that you might not be as familiar with.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think they're very helpful there as well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because there's a ton of Rust code out there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The models are actually pretty good at it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I happen to not know that much about it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So the models are very useful there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think you're getting at some of my, like why my timelines are a bit longer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think, yeah, they're not very good at code that has never been written before.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe it's like one way to put it, which is like what we're trying to achieve when we're building these models.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's tough.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think they kind of know, but they don't fully know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they don't know how to fully integrate it into the repo and your style and your code and your place and some of the custom things that you're doing and how it fits with all the assumptions of the repository and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think they do have some knowledge, but...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

they haven't gotten to the place where they can actually integrate it, make sense of it, and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that a lot of the stuff, by the way, continues to improve.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think currently probably state-of-the-art model that I go to is the GPT-5 Pro.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's a very, very powerful model.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So if I actually have 20 minutes, I will copy-paste my entire repo and I go to GPT-5 Pro, the Oracle, for like some questions.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And often it's not too bad and surprisingly good compared to what existed a year ago.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do think that overall the models are – they're not there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of feel like the industry, it's over – it's making too big of a jump.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's trying to pretend like this is amazing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's not.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's slop.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think they're not coming to terms with it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And maybe they're trying to fundraise or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm not sure what's going on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But we're at this intermediate stage.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The models are amazing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They still need a lot of work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

For now, autocomplete is my sweet spot.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But sometimes, for some types of code, I will go to a nullim agent.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe you guys discussed one other kind of thought that is like, I do feel like I have a hard time differentiating where AI begins and stops.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because I do see AI as fundamentally an extension of computing in some pretty fundamental way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I feel like I see a continuum of this kind of like recursive self-improvement or like of speeding up programmers all the way from the beginning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like even like I would say like code editors.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

syntax highlighting, syntax or like checking even of the types, like data type checking.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

All these kinds of tools that we've built for each other, even search engines, like why aren't search engines part of AI?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, I don't know, like ranking is kind of AI, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

At some point, Google was like, even early on, they were thinking of themselves as an AI company doing Google search engine, which I think is totally fair.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I kind of see it as a lot more of a continuum than I think other people do, and I don't, it's hard for me to draw the line.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of feel like, okay, we're now getting a much better autocomplete.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And now we're also getting some agents which are kind of like these loopy things, but they kind of go off rails sometimes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And what's going on is that the human is progressively doing a bit less and less of the low-level stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

For example, we're not writing the assembly code because we have compilers, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like compilers will take my highlight language in C and write the assembly code.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So we're abstracting ourselves very, very slowly.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's this what I call autonomy slider of like more and more stuff is automated of the stuff that can be automated at any point in time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And we're doing a bit less and less and raising ourselves in the layer of abstraction over the automation.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, maybe the way I would put it is humans don't use reinforcement learning is maybe what I, as I've said it all.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think they do something different, which is, yeah, you experience.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So reinforcement learning is a lot worse than I think the average person thinks.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Reinforcement learning is terrible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It just so happens that everything that we had before is much worse.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because previously, we were just imitating people, so it has all these issues.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So in reinforcement learning, say you're working with, you're solving a math problem.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This is very simple.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're given a math problem, and you're trying to find a solution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Now, in reinforcement learning, you will try lots of things in parallel first.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So you're given a problem, you try hundreds of things,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

different attempts.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And these attempts can be complex, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They can be like, oh, let me try this, let me try that, this didn't work, that didn't work, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then maybe you get an answer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And now you check the back of the book and you see, okay, the correct answer is this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then you can see that, okay, this one, this one, and that one got the correct answer, but these other 97 of them didn't.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So literally what reinforcement learning does is it goes to the ones that worked really well, and every single thing you did along the way, every single token gets up-weighted of, like, do more of this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The problem with that is, I mean, people will say that your estimator has high variance, but, I mean, it's just noisy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's noisy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So basically, it kind of almost assumes that every single little piece of the solution that you made that right at the right answer was the correct thing to do, which is not true.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, you may have gone down the wrong alleys

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

until you write the right solution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Every single one of those incorrect things you did, as long as you got to the correct solution, will be up-weighted as do more of this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's terrible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's noise.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You've done all this work only to find a single, at the end, you get a single number of like, oh, you did correct.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And based on that, you weigh that entire trajectory as like up-weight or down-weight.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so the way I like to put it is you're sucking supervision through a straw because you've done all this work that could be a minute to roll out and you're like sucking the bits of supervision of the final reward signal through a straw and you're like putting it, you're like, you're basically like, yeah, you're broadcasting that across the entire trajectory and using that to upweigh or downweigh that trajectory.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's crazy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A human would never do this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Number one, a human would never do hundreds of rollouts.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Number two, when a person sort of finds a solution, they will have a pretty complicated process of review of like, okay, I think these parts that I did well, these parts I did not do that well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I should probably do this or that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they think through things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's nothing in current LLMs that does this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's no equivalent of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do see papers popping out that are trying to do this because it's obvious to everyone in the field.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I kind of see it as like, the first imitation learning actually, by the way, was extremely surprising and miraculous and amazing that we can fine-tune by imitation in humans.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that was incredible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because in the beginning, all we had was base models.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Base models are autocomplete.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it wasn't obvious to me at the time, and I had to learn this, and the paper that blew my mind was InstructGPT.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because it pointed out that, hey, you can take the pre-trained model, which is autocomplete,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And if you just fine-tune it on text that looks like conversations, the model will very rapidly adapt to become very conversational.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it keeps all the knowledge from pre-training.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this blew my mind because I didn't understand that this just like stylistically can adjust so quickly and become an assistant to a user through just a few loops of fine-tuning on that kind of data.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It was very miraculous to me that that worked.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So incredible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that was like two years, three years of work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And now came RL.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And RL allows you to do a bit better than just imitation learning, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because you can't have these reward functions and you can hill climb on the reward functions.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so some problems have just correct answers.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can hill climb on that without getting expert trajectories to imitate.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that's amazing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And the model can also discover solutions that the human might never come up with.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So this is incredible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And yet, it's still stupid.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think we need more.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I saw a paper from Google yesterday that tried to have this reflect and review page idea in mind.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What was the memory bank paper or something?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I've actually seen a few papers along these lines.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I expect there to be some kind of a major update to how we do algorithms for LLMs coming in that realm.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I think we need three or four or five more.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So process-based supervision just refers to the fact that we're not going to have a reward function only at the very end of after you've made 10 minutes of work, I'm not going to tell you you did well or not well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm going to tell you at every single step of the way how well you're doing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this is basically the reason we don't have that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's tricky how you do that properly because you have partial solutions and you don't know how to assign credit.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So when you get the right answer, it's just an equality match to the answer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Very simple to implement.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If you're doing basically process supervision, how do you assign, in an automatable way, partial credit assignment?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's not obvious how you do it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Lots of labs, I think, are trying to do it with these LLM judges.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So basically, you get LLMs to try to do it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So you prompt an LLM, hey, look at a partial solution of a student.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How well do you think they're doing if the answer is this?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they try to tune the prompt.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The reason that I think this is kind of tricky is quite subtle.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's the fact that anytime you use an LLM to assign a reward, those LLMs are giant things with billions of parameters and they're gameable.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And if you're reinforcement learning with respect to them, you will find adversarial examples for your LLM judges almost guaranteed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can't do this for too long.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You do maybe 10 steps or 20 steps, maybe it will work, but you can't do 100 or 1,000 because it's not obvious.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because I understand it's not obvious, but basically the model will find little cracks,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

it will find all these spurious things in the nooks and crannies of the giant model and find a way to cheat it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So one example that's prominently in my mind is, I think this was probably public, but basically, if you're using an element judge for a reward, so you just give it a solution from a student and ask it if the student will or not,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We were training with reinforcement learning against that reward function, and it worked really well, and then suddenly the reward became extremely large.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It was a massive jump, and it did perfect.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you're looking at it like, wow, this means the student is perfect in all these problems.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's fully solved math.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But actually what's happening is that when you look at the completions that you're getting from the model, they are complete nonsense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They start out okay, and then they change to da-da-da-da-da-da-da.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So it's just like, oh, okay, let's take two plus three, and we do this and this, and then da-da-da-da-da-da-da-da.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you're looking at it and it's like, this is crazy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How is it getting a reward of one or 100%?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you look at the LLM judge and it turns out that the, the, the, the, the is an adversarial examples for the model and it assigns 100% probability to it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's just because this is an out-of-sample example to the LLM.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's never seen it during training, and you're in pure generalization land.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's never seen it during training, and in the pure generalization land, you can find these examples that break it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Not even that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Prompt injection is way too fancy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're finding adversarial examples, as they're called.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

These are nonsensical solutions that are obviously wrong, but the model thinks they're amazing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think the labs are probably doing all that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, okay, so the obvious thing is like the should not get 100% reward.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

okay, well, take the, the, the, the, put in the training set of the LLM judge and say, this is not 100%, this is 0%.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can do this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But every time you do this, you get a new LLM and it still has adversarial examples.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's infinity adversarial examples.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think probably if you iterate this a few times, it'll probably be harder and harder to find adversarial examples.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I'm not 100% sure because this thing has a trillion parameters or whatnot.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I bet you the labs are trying.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually, I still think, I still think we need other ideas.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So like this idea of like a review solution and come up with synthetic examples such that when you train on them, you get better and like meta-learn it in some way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think there's some papers that I'm starting to see pop out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I only am at a stage of like reading abstracts because a lot of these papers, you know, they're just ideas.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Someone has to actually like make it work on a frontier LLM lab scale.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

in full generality.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because when you see these papers, they pop up and it's just like a little bit of noisy, you know?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's cool ideas, but I haven't actually seen anyone convincingly show that this is possible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That said, the LLM labs are fairly closed, so who knows what they're doing now, but...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I do think that we're missing some aspects there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So as an example, when you're reading a book,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I almost feel like currently when LLMs are reading a book, what that means is we stretch out the sequence of text and the model is predicting the next token and it's getting some knowledge from that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's not really what humans do, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So when you're reading a book, I almost don't even feel like the book is like exposition I'm supposed to be attending to and training on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The book is a set of prompts for me to do synthetic data generation.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

or for you to get into a book club and talk about it with your friends.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's by manipulating that information that you actually gain that knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think we have no equivalent of that, again, with LLMs.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They don't really do that, but I'd love to see during pre-training some kind of a stage that thinks through the material and tries to reconcile it with what it already knows and thinks through for some amount of time and gets that to work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so there's no equivalence of any of this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This is all research.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's some subtle, very subtle that I think are very hard to understand reasons why it's not trivial.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So if I can just describe one.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Why can't we just synthetically generate and train on it?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, because every synthetic example, like if I just give synthetic generation of the model thinking about a book, you look at it and you're like, this looks great.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Why can't I train on it?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, you could try, but the model will actually get much worse if you continue trying.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's because all of the samples you get from models are silently collapsed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're silently, this is not obvious if you look at any individual example of it, they occupy a very tiny manifold of the possible space of sort of thoughts about content.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So the LLMs, when they come off, they're what we call collapsed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They have a collapsed data distribution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If you sample, one easy way to say it is go to ChatGPT and ask it, tell me a joke.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It only has like three jokes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's not giving you the whole breadth of possible jokes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's giving you like, it knows like three jokes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're silently collapsed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So basically, you're not getting the richness and diversity and the entropy from these models as you would get from humans.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So humans are a lot more sort of noisier, but at least they're not biased.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're not in a statistical sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're not silently collapsed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They maintain a huge amount of entropy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So how do you get synthetic data generation to work despite the collapse and while maintaining the entropy is a research problem.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Say we have a chapter of a book and I ask an alum to think about it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It will give you something that looks very reasonable.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But if I ask it 10 times, you'll notice that all of them are the same.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, yeah, yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So any individual sample will look okay, but the distribution of it is quite terrible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's quite terrible in such a way that if you continue training on too much of your own stuff, you actually collapse.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I actually think that there's no like fundamental solutions to this possibly.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I also think humans collapse over time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think this is, again, these analogies are surprisingly good, but humans collapse during the course of their lives.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This is why children have completely, you know, they haven't overfit yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they will say stuff that will shock you because it's kind of, you can see where they're coming from, but it's just not the thing people say.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And because they're not yet collapsed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But we're collapsed, we end up revisiting the same thoughts, we end up saying more and more of the same stuff, and the learning rates go down, and the collapse continues to get worse, and then everything deteriorates.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's an interesting idea.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I do think that...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

When you're generating things in your head and then you're attending to it, you're kind of like training on your own samples.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're training on your synthetic data.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And if you do it for too long, you go off rails and you collapse way too much.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So you always have to like seek entropy in your life.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So talking to other people is a great source of entropy and things like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe the brain has also built some internal mechanisms for increasing the amount of entropy in that process.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But yeah, maybe that's an interesting idea.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think there's something very interesting about that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, 100%.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that humans actually, they do kind of like have a lot more of an element compared to LLMs of like seeing the forest for the trees.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And we're not actually that good at memorization, which is actually a feature.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because we're not that good at memorization, we actually are kind of like forced to...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

to find the patterns in a marginal sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think LLNs, in comparison, are extremely good at memorization.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They will recite passages from all these training sources.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can give them completely nonsensical data, like you can hash some amount of text or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You get a completely random sequence.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If you train on it, even just, I think, a single iteration or two, it can suddenly regurgitate the entire thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It will memorize it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's no way a person can read a single sequence of random numbers and recite it to you.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's a feature, not a bug almost, because it forces you to like only learn the generalizable components.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Whereas LLMs are distracted by all the memory that they have of the pre-trained documents.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's probably very distracting to them in a certain sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that's why when I talk about the cognitive core, I actually want to remove the memory, which is what we talked about.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'd love to have less memory so that they have to look things up.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they only maintain the algorithms for like thought and the idea of an experiment and all this cognitive glue of acting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm not sure.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think it's almost like a separate axis.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's almost like the models are way too good at memorization and somehow we should remove that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think people are much worse, but it's a good thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I think that's a great question.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, you can imagine having a regularization for entropy and things like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I guess they just don't work as well empirically because right now, like, the models are collapsed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I will say...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Most of the tasks that we want of them don't actually demand the diversity.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's probably the answer of what's going on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so it's just that the frontier labs are trying to make the models useful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of just feel like the diversity of the outputs is not so much.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Number one, it's much harder to work with and evaluate and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But maybe it's not what's actually capturing most of the value.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Or like maybe if you're doing a lot of writing help from LLMs and stuff like that, I think it's probably bad because the models will give you these like silently all the same stuff, you know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So they're not, they won't explore lots of different ways of answering a question, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I kind of feel like maybe this diversity is just not as big of a, yeah, maybe like, yeah, not as many applications needed so the models don't have it, but then it's actually a problem with synthetic generation time, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So we're actually shooting ourselves in the foot by not allowing this entropy to maintain in the model.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think possibly the labs should try harder.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually know if it's super fundamental.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually know if I intended to say that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I haven't done these experiments, but I do think that you could probably regularize the entropy to be higher.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So you're encouraging the model to give you more and more solutions.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But you don't want it to start deviating too much from the training data.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's going to start making up its own language.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's going to start using words that are extremely rare.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So it's going to drift too much from the distribution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think controlling the distribution is just like a tricky... It's just like someone just has to...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's probably not trivial in that sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So it's really interesting in the history of the field because at one point everything was very scaling-pilled in terms of like, oh, we're going to make much bigger models, trillions of parameter models.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And actually what the models have done in size is they've gone up and now they've actually kind of like

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

actually even come down.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

State-of-the-art models are smaller.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And even then, I actually think they memorized way too much.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think I had a prediction a while back that I almost feel like we can get cognitive cores that are very good at even like a billion, billion parameters.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It should be all very like, like if you talk to a billion parameter model, I think in 20 years, you can actually have a very productive conversation, it thinks.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's a lot more like a human.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But if you ask it some factual question, it might have to look it up.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But it knows that it doesn't know and it might have to look it up and it will just do all the reasonable things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

No, because I basically think that the training data is, so here's the issue.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The training data is the internet, which is really terrible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So there's a huge amount of gains to be made because the internet is terrible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like if you actually, and even the internet, when you and I think of the internet, you're thinking of like a Wall Street Journal or that's not what this is.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

When you're actually looking at a pre-training data set in the front of your lab and you look at a random internet document, it's total garbage.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like I don't even know how this works at all.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's some like stock ticker symbols.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's a huge amount of slop and garbage from like all the corners of the internet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's not like your Wall Street Journal article that's extremely rare.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I almost feel like because the internet is so terrible, we actually have to sort of like build really big models to compress all that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Most of that compression is memory work instead of like cognitive work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But what we really want is the cognitive part to actually delete the memory.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then, so I guess what I'm saying is like we need

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

intelligent models to help us refine even the pre-training set to just narrow it down to the cognitive components.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I think you get away with a much smaller model because it's a much better data set and you could train it on it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But probably it's not trained directly on it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's probably distilled for a much better model still.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I just feel like distillation works extremely well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So almost every small model, if you have a small model, it's almost certainly distilled.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, come on, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

At some point, it should take at least a billion knobs to do something interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're thinking it should be even smaller?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I almost feel like I'm already contrarian by talking about a billion-parameter cognitive core, and you're outdoing me.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think, yeah, maybe we could get a little bit smaller.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I still think that there should be enough.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, maybe it can be smaller.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that, practically speaking, you want the model to have some knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You don't want it to be looking up everything.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because then you can't think in your head.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're looking up way too much stuff all the time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I do think it needs to be some basic curriculum needs to be there for knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But it doesn't have esoteric knowledge, you know?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I don't know that I have a super strong prediction.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that the labs are just being practical.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They have a flops budget and a cost budget.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it just turns out that pre-training is not where you want to put most of your flops or your cost.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that's why the models have gotten smaller, because they are a bit smaller, the pre-training stage is smaller, et cetera, but they make it up in reinforcement learning and all this kind of stuff, mid-training and all this kind of stuff that follows.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So they're just being practical in terms of all the stages and how you get the most bang for the buck.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess like forecasting that trend, I think, is quite hard.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do still expect that there's so much longing for it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's my basic expectation.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I have a very wide distribution here.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Probably most part, yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I expect the data sets to get much, much better because when you look at the average data sets, they're extremely terrible.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like so bad that I don't even know how anything works, to be honest.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like look at the average example in the training set.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like factual mistakes, errors, nonsensical things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Somehow when you do it at scale, the noise washes away and you're left with some of the signal.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

um so data sets will improve a ton it's just everything gets better so um our hardware um our all the kernels um all the kernels for running the hardware and maximizing what you get with the hardware you know so nvidia is slowly tuning the actual hardware itself tensor course and so on all that needs to happen and will continue to happen uh all the kernels will get better and utilize the chip to the max extent all the algorithms will probably improve over optimization architecture and just all the modeling components of how everything is done and what the algorithms are that we're even training with

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I do kind of expect like a just very just everything.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Nothing dominates.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Everything plus 20%.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This is like roughly what I've seen.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess I have two answers to that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Number one, I'm almost tempted to, like, reject the question entirely because, again, like, I see this as an extension of computing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Have we talked about, like, how to chart progress in computing or how do you chart progress in computing since 1970s or whatever?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What is the x-axis?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I kind of feel like the whole question is kind of, like, funny from that perspective a little bit.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I will say, I guess, like, when people talk about AI and the original AGI and how we spoke about it when we – when OpenAI started –

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

AGI was a system you can go to that can do any task that is economically valuable, any economically valuable task at human performance or better.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Okay.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that was the definition.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I was pretty happy with that at the time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of feel like I've stuck to that definition forever.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then people have made up all kinds of other definitions.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I feel like I like that definition.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Now, number one, the first concession that people make all the time is they just take out all the physical stuff because we're just talking about digital knowledge work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I feel like that's a pretty major concession compared to the original definition, which was like any task a human can do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I can lift things, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like AI can't do that, obviously.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, okay, but we'll take it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What fraction of the economy are we taking away by saying, oh, only knowledge work?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually know the numbers.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I feel like it's about 10% to 20%, if I had to guess, is only knowledge work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like someone could work from home and perform tasks, something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still think it's a really large market.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, yeah, what is the size of the economy and what is 10%, 20%?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, we're still talking about a few trillion dollars of, even in the U.S., of market share almost, or like work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, still a very massive bucket.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, but I guess like going back to the definition, I guess what I would be looking for is to what extent is that definition true?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, are there jobs or lots of tasks?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If we think of tasks as, you know, not jobs, but tasks, kind of difficult definitions.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because the problem is like society will refactor based on the tasks that make up jobs compared to what's based on what's automatable or not.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But today, what jobs are replaceable by AI?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So a good example recently was Jeff Hinton's prediction that radiologists would not be a job anymore.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this turned out to be very wrong in a bunch of ways, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So radiologists are alive and well and growing, even though computer vision is really, really good at recognizing all the different things that they have to recognize in images.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's just messy, complicated things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

job with a lot of surfaces and dealing with patients and all this kind of stuff in the context of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess I don't actually know that by that definition, AI has made a huge amount of dent yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But some of the jobs maybe that I would be looking for have some features that I think make it very amenable to automation earlier than later.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

As an example, call center employees often come up, and I think rightly so, because call center employees are

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

have a number of simplifying properties with respect to what's automatable today.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Their jobs are pretty simple.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's a sequence of tasks and every task looks similar.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like you take a phone call with a person, it's 10 minutes of interaction or whatever it is, probably a bit longer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

In my experience, a lot longer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you complete some task in some scheme and you change some database entries around or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So you keep repeating something over and over again, and that's your job.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So basically, you do want to bring in the task horizon, how long it takes to perform a task.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then you want to also remove context, like you're not dealing with different parts of services of companies or other customers.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's just the database, you and a person you're serving.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so it's more closed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's more understandable.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's purely digital.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I would be looking for those things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But even there, I'm not actually looking at full automation yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm looking for an autonomy slider.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I almost expect that we are not going to,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

instantly replace people.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We're going to be swapping in AIs that do 80% of the volume.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They delegate 20% of the volume to humans.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And humans are supervising teams of five AIs doing the call center work that's more rote.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I would be looking for new interfaces or new companies that provide some kind of a layer that allows you to manage some of these AIs that are not yet perfect.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I would expect that across the economy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And a lot of jobs are a lot harder than call center employee.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I think that's an interesting question.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't think we're currently seeing that with radiology.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I don't have, like...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

In my understanding, but I think radiology is not a good example, basically.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know why Jeff Hinton picked on radiology because I think it's an extremely messy, complicated profession.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I would be a lot more interested in what's happening with call center employees today, for example, because I would expect a lot of the road stuff to be automatable today.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I don't have a first level access to it, but maybe I would be looking for trends of what's happening with the call center employees.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe some of the things I would also expect is maybe they are swapping in AI, but then I would still wait for a year or two because I would potentially expect them to pull back and actually rehire some of the people.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think there's an interesting point here because I do believe coding is like the perfect first thing for these LLMs and agents.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's because coding has always fundamentally worked around text.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's computer terminals and text, and everything is based around text.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And LLMs, the way they're trained on the internet, love text.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so they're perfect text processors, and there's all this data out there, and it's just a perfect fit.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And also we have a lot of infrastructure pre-built for handling code and text.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, for example, we have Visual Studio Code or, you know, your favorite IDE showing you code.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And an agent can plug into that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So for example, if an agent has a diff where it made some change, we suddenly have all this code already that shows all the differences to a code base using a diff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So it's almost like we've pre-built a lot of the infrastructure for code.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Now contrast that with some of the things that don't enjoy that at all.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So as an example, like there's people trying to build automation, not for coding, but for example, for slides.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like I saw a company doing slides.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's much, much harder.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And the reason it's much, much harder is because slides are not text.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Slides are little graphics, and they're arranged spatially, and there's visual components to it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And slides don't have this pre-built infrastructure.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, for example, if an agent is to make a different change to your slides, how does a thing show you the diff?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How do you see the diff?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's nothing that shows diffs for slides.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Someone has to build it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So it's just some of these things are not amenable to AIs as they are, which is text processors.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And code, surprisingly, is.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I think that makes sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I would say, yeah, I'm not saying that anything text is trivial, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that code is like, it's pretty structured.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Text is maybe a lot more flowery and there's a lot more like entropy in text, I would say.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know how else to put it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And also, I mean, code is hard.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so people sort of feel quite empowered by LLMs, even from like simple kind of knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I basically, I don't actually know that I have a very good answer.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, obviously like text makes it much, much easier maybe is maybe why I put it, but it doesn't mean that all text is trivial.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I guess I see it as like a progression of automation in society, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And again, like extrapolating the trend of computing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I just feel like there will be a gradual automation of a lot of things, and superintelligence will be sort of like the extrapolation of that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I do think we expect more and more autonomous entities over time that are doing a lot of the digital work, and then eventually even the physical work, probably some amount of time later.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But basically, I see it as just automation.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, but some of the things that people do is invent new things, which I would just put into the automation, if that makes sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, it is fundamentally automation, but I mean, it will be like extremely foreign.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think it will look really strange because like you mentioned, we can run all of this on a computer cluster, et cetera, and much faster and all this thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, maybe some of the scenarios, for example, that I start to get, like, nervous about with respect to when the world looks like that is this kind of, like, gradual loss of control and understanding of what's happening.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think that's actually the most likely outcome, probably, is that there will be a gradual loss of understanding of... And we'll gradually layer all this stuff everywhere, and there'll be fewer and fewer people who understand it, and that there will be a sort of this, like, scenario of a gradual loss of control and understanding of what's happening.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That, to me, seems most likely outcome of how all of this stuff will go down.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I think that's fair.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's a good pushback.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think, like, I guess I expect loss of both power.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So we're really far into a territory of... I don't know what this looks like, but if I was to write sci-fi novels, they would look along the lines of not even a single entity or something like that, that just sort of takes over everything, but actually multiple competing entities that gradually become more and more autonomous, and some of them go rogue, and the others fight them off, and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's like this hot pot of...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

completely autonomous activity that we've delegated to.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I kind of feel like it would have that flavor.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I basically expect there to be, I mean, a lot of these things, I mean, they will be tools to people and the people could, some of the population is like, they're acting on behalf of people or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe those people are in control, but maybe it's a loss of control overall for society in the sense of like outcomes we want or something like that, where you have entities acting on behalf of individuals that are still kind of roughly seen as out of control.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I guess what I mean is...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do, but it's business as usual because we're in an intelligence explosion already and have been for decades.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And when you look at GDP, it's basically the GDP curve that is an exponential weighted sum over so many aspects of the industry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Everything is gradually being automated, has been for hundreds of years.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Industrial revolution is automation and some of the physical components and the tool building and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Compilers are early software automation, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I kind of feel like we've been recursively self-improving and exploding for a long time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe another way to see it is,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, Earth was a pretty, I mean, if you don't look at the biomechanics and so on, it was a pretty boring place, I think, and looked very similar if you just look from space.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And Earth is spinning, and then, like, we're in the middle of this, like, firecracker event.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But we're seeing it in slow motion.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I definitely feel like this has already happened for a very long time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And, again, like, I don't see AI as, like, a distinct technology with respect to what has already been happening for a long time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This was very interesting to me because I was trying to find AI in the GDP for a while.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I thought that GDP should go up.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But then I looked at some of the other technologies that I thought were very transformative, like maybe computers or mobile phones or etc.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can't find them in GDP.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

GDP is the same exponential.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's just that even, for example, the early iPhone didn't have the App Store and it didn't have a lot of the bells and whistles that the modern iPhone has.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so even though we think of 2008, was it, when iPhone came out as like some major seismic change, it's actually not.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Everything is like so spread out and so slowly diffuses that everything ends up being averaged up into the same exponential.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's the exact same thing with computers.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can't find them in the GDP.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's like, oh, we have computers now.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's not what happened because it's such a slow progression.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And with AI, we're going to see the exact same thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's just more automation.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It allows us to write different kinds of programs that we couldn't write before.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But AI is still fundamentally a program.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's a new kind of computer and a new kind of computing system.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But it has all these problems.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's going to diffuse over time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's still going to add up to the same exponential.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And we're still going to have an exponential that's going to get extremely vertical.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's going to be very foreign to live in that kind of an environment.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Basically, I guess what I'm saying is for a while, I tried to find AI or look for AI in like the GDP curve.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of convinced myself that this is false.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that even when people talk about recursive self-improvement and labs and stuff like that, I even don't, this is business as usual.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Of course, it's going to recursively self-improve and it's been recursively self-improving.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like LLMs allow the engineers to work much more efficiently to build the next round of LLM.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And a lot more of the components are being automated and tuned and et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

all the engineers having access to Google search is sort of part of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

All the engineers having an ID, all of them having autocomplete or having cloth code, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's all just part of the same speed up of the whole thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So it's just so smooth.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, my expectation is that it stays the same pattern.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, maybe a counterpoint.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, number one, I'm actually pretty willing to be convinced one way or another on this point.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I will say, for example, computing is labor.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Computing was labor.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Computers, like, a lot of jobs disappeared because computers are automating a bunch of digital information processing that you now don't need a human for.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so computers are labor.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that has played out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And, you know, self-driving as an example is also like computers doing labor.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, like, I guess that's already been playing out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's still business as usual.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I kind of...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I see where it's coming from.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

At the same time, I do feel like people make this assumption of like, okay, we have God in the box and now it can do everything.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it just won't look like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's going to be able to do some of the things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's going to fail at some other things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's going to be gradually put into society and basically end up with the same pattern, is my prediction.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because this assumption of suddenly having a completely intelligent, fully flexible, fully general human in a box and we can dispense it at arbitrary problems in society, I don't think that we will have this

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

like discrete change.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I think we'll arrive at the same kind of gradual diffusion of this across the industry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think I understand, but I still think that you're presupposing some discrete jump, some unlock that we're waiting to claim.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And suddenly we're going to have geniuses in data centers.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I still think you're presupposing some discrete jump that I think has basically no historical precedent that I can't find in any of the statistics and that I think probably won't happen.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm a little bit suspicious.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would have to look at it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm a little bit suspicious, and I would have to take a look.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

For example, maybe some of the logs are not very good from before the Industrial Revolution or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I'm a little bit suspicious of it, but yeah, maybe you're right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't have strong opinions.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe you're saying that this was a singular event that was extremely magical, and you're saying that maybe there's going to be another event that's going to be just like that, extremely magical, it will break paradigm, and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's still some overhang that's being unlocked.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like maybe there's a new energy source.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's some unlock, in this case, some kind of a cognitive capacity.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's an overhang of cognitive work to do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you're expecting that overhang to be filled by this new technology when it crosses the threshold.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I mean, yeah, it's really hard to tell.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I understand that viewpoint.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't intuitively feel that viewpoint.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I love Nick Fling's books, by the way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, yeah, I was just listening to his podcast on the way up here.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

With respect to intelligence and its evolution, I do think it came fairly, I mean, it's very, very recent, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I am surprised that it evolved.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I find it fascinating to think about all the worlds out there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, say, there's a thousand planets like Earth and what they look like.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think Nick Lane was here talking about some of the early parts, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, okay, he expects basically very similar life forms, roughly speaking, and bacteria-like things in most of them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then there's a few breaks in there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would expect that the evolution of intelligence intuitively feels to me like it should be a fairly rare event.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there have been animals for... I guess maybe you should base it on how long something has existed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, for example, if bacteria have been around for 2 billion years and nothing happened, then going to your carrier is probably pretty hard because bacteria actually came up quite early in Earth's evolution or history.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I guess, how long have we had animals?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe a couple hundred million years, like multicellular animals that like run, run, crawl, et cetera, which is maybe 10% of Earth's lifespan or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe on that timescale, it's actually not too tricky.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still feel like

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's still surprising to me, I think, intuitively that it developed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I would maybe expect just a lot of like animal-like life forms doing animal-like things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The fact that you can get something that creates culture and knowledge and accumulates it, it is surprising to me.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Basically, it's so hard to tell, right, with any of this stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I guess you can base it a little bit on how long something has existed or how long it feels like something has been bottlenecked.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So Nicolain is very good about describing this, like, very apparent bottleneck in bacteria and archaea.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

For two billion years, nothing happened.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, extreme diversity of chemical, of biochemistry, and yet nothing that grows to become animals.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Two billion years.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know that we've seen exactly that kind of an equivalent with animals and intelligence to your point, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I guess maybe we could also look at it with respect to how many times we think evolution or intelligence has like individually sprung up.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's a really good thing to investigate.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe one thought on that is, I almost feel like, well, there's the hominid intelligence.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's, I would say, like the bird intelligence, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like ravens, et cetera, are extremely clever.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But their brain parts are actually quite distinct, and we don't have that much existence.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So maybe that's a slight event of, there's a slight indication of maybe intelligence springing up a few times.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so in that case, you'd maybe expect it more frequently or something.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, and just stuff to work with.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I'm guessing it would be harder to, if I was a dolphin, I mean, how do you do, you can't have fire, for example, and stuff like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, probably like the universe of things you can do in water, like inside water, is probably lower than what you can do on land.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

just chemically.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, I do agree with this viewpoint of these niches and what's being incentivized.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still find it kind of miraculous that I don't, I would have maybe expected things to get stuck on like animals with bigger muscles, you know?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like going through intelligence is actually a really fascinating breaking point.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, exactly.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You have to incentivize some kind of adaptability.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You actually want environments that are unpredictable.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So evolution can't bake your algorithms into your weights.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A lot of animals are basically pre-baked in this sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so humans have to figure it out at test time when they get born.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so maybe you actually want these kinds of environments that actually change really rapidly or something like that where you can't foresee what will work well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so you actually put all that intelligence, you create intelligence to figure it out at test time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yes and no, because LLMs don't really have the equivalent of culture.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And maybe we're giving them way too much and incentivizing not to create it or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I guess like the mention of culture and of written record and of like passing down notes between each other, I don't think there's an equivalent of that with LLMs right now.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So LLMs don't really have culture right now.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's kind of like one of the, I think, impediments, I would say.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Can you give me some sense of what LLM culture might look like?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So in the simplest case, it would be a giant scratchpad that the LLM can edit.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And as it's reading stuff or as it's helping out with work, it's editing the scratchpad for itself.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Why can't an LLM write a book for the other LLMs?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That would be cool.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, why can't other LLMs read this LLM's book and be inspired by it or shocked by it or something like that?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's no equivalence for any of this stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think there's two powerful ideas in the realm of multi-agent that have both not been like really claimed or so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The first one I would say is culture and LLM is basically a growing repertoire of knowledge for their own purposes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The second one looks a lot more like the powerful idea of self-play.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

in my mind, is extremely powerful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So evolution actually has a lot of competition, basically, driving intelligence and evolution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And in AlphaGo, more algorithmically, AlphaGo is playing against itself, and that's how it learns to get really good at Go.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's no equivalent of self-playing LLMs, but I would expect that to also exist, but no one has done it yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Why can't an LLM, for example, create a bunch of problems that another LLM is learning to solve?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then the LLM is always trying to serve more and more difficult problems.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

stuff like that, you know?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So like, I think there's a bunch of ways to actually organize it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think it's a realm of research.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think I haven't seen anything that convincingly like claims both of those, like multi-agent improvements.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still think we're mostly in the realm of a single individual agent.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think, I also think that will change.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And in the realm of culture also, I would bucket also organizations.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And we haven't seen anything like that convincingly either.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that's why we're still early.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Somehow remarkably, again, some of these analogies work and they shouldn't, but somehow remarkably they do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A lot of the smaller models or the dumber, like the smaller models somehow remarkably resemble like a kindergarten student or then like a elementary school student or high school student, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And somehow we still haven't like graduated enough where this stuff can take over.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like it's still mostly like my cloth coat or codex, they still kind of feel like this elementary grade student.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I know that they can take PhD quizzes, but they still cognitively feel like a kindergarten or an elementary school student.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I don't think they can create culture because they're still kids, you know, like they're savant kids.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They have perfect memory of all this stuff, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they can convincingly create all kinds of slop that looks really good.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I still think they don't really know what they're doing and they don't really have the cognition across all these little checkboxes that we still have to collect.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I would say one thing I will almost instantly also push back on is this is not even near done.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So in a bunch of ways that I'm going to get to.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that self-driving is very interesting because it's definitely like where I get a lot of my intuitions because I spent five years on it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it has this entire history where actually the first demos of self-driving go all the way to the 1980s.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can see a demo from CMU in 1986.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's a truck that's driving itself on roads.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But OK, fast forward.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think when I was joining Tesla, I had a very early demo of Waymo.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it basically gave me a perfect drive in 2014 or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So perfect Waymo drive a decade ago.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It took us around Palo Alto and so on because I had a friend who worked there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I thought it was like very close and then still took a long time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I do think that for some kinds of tasks and jobs and so on, there's a very large demo to product gap where the demo is very easy, but the product is very hard.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's especially the case in cases like self-driving where the cost of failure is too high, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Many industries, tasks, and jobs maybe don't have that property, but when you do have that property, that definitely increases the timelines.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that, for example, in software engineering, I do actually think that that property does exist.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think for a lot of vibe coding, it doesn't.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think if you're writing actual production grade code, I think that property should exist because any mistake actually leads to security vulnerability or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Millions and hundreds of millions of people's personal social security numbers, etc, get leaked or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that it is a case that in software, people should be careful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Kind of like in self-driving.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like in self-driving, if things go wrong, you might get injury.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I guess there's worse outcomes.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I guess in software, I almost feel like it's almost unbounded how terrible some things could be.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I do think that they share that property.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I think basically what takes the long amount of time and the way to think about it is that it's a march of nines, and every single nine is a constant amount of work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So every single nine is the same amount of work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So when you get a demo and something works 90% of the time, that's just the first nine.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then you need a second nine, a third nine, fourth nine, fifth nine.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And while I was at Tesla for, was it five years or so, I think we went through maybe three nines, two nines, I don't know what it is.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But like multiple nines of iteration, there's still more nines to go.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so that's why these things take so long.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so it's definitely formative for me, like seeing something that was a demo.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm very unimpressed by demos.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So whenever I see demos of anything, I'm extremely unimpressed by that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It works better if you can, if it's a demo that someone cooked up and is just showing you its worst.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If you can interact with it, it's a bit better.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But even then, you're not done.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You need actual product.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's going to face all these challenges when it comes in contact with reality and all these different pockets of behavior that need patching.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I think we're going to see all that stuff play out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's a march of nines.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Each nine is constant.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Demos are encouraging.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Still a huge amount of work to do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think it is a kind of a critical safety domain, unless you're doing vibe coding, which is all nice and fun and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so that's why I think this also enforced my timelines from that perspective.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, it's a much harder problem.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, self-driving is just one of thousands of things that people do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's almost like a single vertical, I suppose.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Whereas when we're talking about general software engineering, it's even more, there's more surface area.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, basically, I'm not 100% sure if I fully agree with that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know how much we're getting for free, and I still think there's a lot of gaps in understanding in what we are getting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, we're definitely getting more generalizable intelligence in a single entity, whereas self-driving is a very special-purpose task that requires, in some sense, building a special-purpose task is maybe even harder in a certain sense because it doesn't fall out from a more general thing that you're doing at scale, if that makes sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I still think that the analogy doesn't, I still don't know if it fully resonates because like the LLMs are still pretty fallible and I still think that they have a lot of gaps and that it still needs to be filled in.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I don't think that we're getting like magical generalization completely out of the box sort of in a certain sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And the other aspect that I want to also actually return to when I was in the beginning was self-driving cars are nowhere near done still.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Mm-hmm.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So even though, so the deployments still are pretty minimal, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So even Waymo and so on has very few cars, and they're doing that, roughly speaking, because they're not economical, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because they've built something that lives in the future.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so they had to, like, pull back future, but they had to make it uneconomical.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So they have all these, like, you know, there's all these costs, not just marginal costs for those cars and their operation and maintenance, but also the capex of the entire thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So making it economical is still going to be a slog, I think, for them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then also I think when you look at these cars and there's no one driving, I also think it's a little bit deceiving because there are actually very elaborate teleoperation centers of people actually kind of like in a loop with these cars.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I don't have the full extent of it, but I think...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

there's more human in the loop that you might expect.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's people somewhere out there basically beaming in from the sky.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I don't actually know that they're fully in the loop with the driving.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think some of the times they are, but they're certainly involved and there are people.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And in some sense, we haven't actually removed the person.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

We've like moved them to somewhere where you can't see them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still think there will be some work, as you mentioned, going from environment to environment.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I think like there's still challenges to make self-driving real.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do agree that it's definitely across the threshold where it kind of feels real, unless it's like really teleoperated.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

For example, Waymo can't go to all the different parts of the city.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

My suspicion is it's like parts of city where you don't get a good signal.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Anyway, so basically, I don't actually know anything about the stack.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I'm just making up stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Sorry, I don't know anything about the specifics of Waymo.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I feel like I talk about them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I actually, by the way, love Waymo, and I take it all the time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I don't want to say, like, I just think people, again, are sometimes a little bit too naive about some of the progress, and I still think there's a huge amount of work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think Tesla took, in my mind, a lot more scalable approach.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think the team is doing extremely well and is going to and I I'm kind of like on the record for predicting how this thing will go, which is like when we had like early start because you can package up so many sensors.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do think Tesla is taking the more scalable strategy and is going to look a lot more like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think this will have to still play out and hasn't.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But basically, like I don't want to talk about self-driving or something that took a decade because it didn't take it didn't take it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If that makes sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, the end is not near yet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because when we're talking about self-driving, usually in my mind, it's self-driving at scale.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

People don't have to get a driver's license, etc.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think that's right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think if you're sticking in the realm of bits, bits are like a million times easier than anything that touches the physical world.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I definitely grant that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Bits are completely changeable, arbitrarily reshuffleable at a very rapid speed.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So you would expect a lot more faster adaptation also in the industry and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then what was the first one?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think that's roughly right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I also think that if we are talking about knowledge work at scale, there will be some latency requirements, practically speaking, because we're going to have to create a huge amount of compute and serve that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I think the last aspect that I very briefly want to also talk about is all the rest of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

just all the rest of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So what does society think about it?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What is the legal, how is it working legally?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How is it working insurance-wise?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Who's really, like, what is the, what are those layers of it and aspects of it?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What happens with, what is the equivalent of people putting a cone on a Waymo?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You know, there's going to be equivalents of all that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I do think that, I almost feel like self-driving is a very nice analogy that you can borrow things from.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, what is the equivalent of a cone on a car?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What is the equivalent of a tele-operating worker who's, like, hidden away?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And almost like all the aspects of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Kind of like what happened with railroads and all this kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

With what, sorry?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Was it railroads?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Sorry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There is like historical precedent or was it with telecommunication industry, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like prepaving the internet that only came like a decade later, you know, and creating like a whole bubble in the telecommunications industry in the late 90s kind of thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I don't know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I understand I'm sounding very pessimistic here.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm only doing that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm actually optimistic.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think this will work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think it's tractable.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm only sounding pessimistic because when I go on my Twitter timeline, I see all this stuff that makes no sense to me.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think there's a lot of reasons for why that exists.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think a lot of it is, I think, honestly, just fundraising.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's just incentive structures.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A lot of it may be fundraising.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A lot of it is just attention, you know, converting attention to money on the internet, you know, stuff like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think there's a lot of that going on, and I think I'm only reacting to that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I'm still, like, overall very bullish on technology.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think we're going to work through all this stuff, and I think there's been a rapid amount of progress.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually know that there's overbuilding.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think that we're going to be able to gobble up what, in my understanding, is being built.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because I do think that, for example, cloud code or OpenAI Codex and stuff like that, they didn't even exist a year ago, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Is that right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think it's roughly right.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This is miraculous technology that didn't exist.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think there's going to be a huge amount of demand as we see the demand in ChashiPT already and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So yeah, I don't actually know that there's overbuilding, but I guess I'm just reacting to some of the very fast timelines that people continue to say incorrectly.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I've heard many, many times over the course of my 15 years in AI where very reputable people keep getting this wrong all the time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think I want us to be properly calibrated.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think some of this also, it does have like geopolitical ramifications and things like that when, like some of these questions.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think I don't want people to make mistakes on that sphere of things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I do want us to be grounded in reality of what technology is and isn't.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I guess maybe, like, the way I would put it is...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I feel some amount of like determinism around the things that AI labs are doing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I feel like I could help out there, but I don't know that I would like uniquely... I don't know that I would like uniquely improve it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think like my personal big fear is that a lot of this stuff happens on the side of humanity and that humanity gets disempowered by it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of like...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I care not just about all the Dyson spheres that we're going to build and that AI is going to build in a fully autonomous way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I care about what happens to humans.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I want humans to be well off in this future.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I feel like that's where I can a lot more uniquely add value than like an incremental improvement in the frontier lab.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I guess I'm most afraid of something maybe like depicted in movies like WALL-E or Idiocracy or something like that, where humanity is sort of on the side of this stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

and I want humans to be much, much better in this future.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I guess, to me, this is kind of like through education that you can actually achieve this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Oh, yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So Eureka is trying to build, I think maybe the easiest way I can describe it is we're trying to build the Starfleet Academy.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know if you've watched Star Trek.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I haven't, but yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Okay, Starfleet Academy is this elite institution for frontier technology, building spaceships and graduating cadets to be the pilots of these spaceships and whatnot.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I just imagine an elite institution for technical knowledge and basically a kind of school that's very up-to-date and very like a premier institution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

With respect to Eureka, I think one thing that is very fascinating to me about education is I do think education will pretty fundamentally change with AIs on the side.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think it has to be rewired and changed to some extent.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still think that we're pretty early.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think there's going to be a lot of people who are going to try to do the obvious things, which is like, oh, have an LLM and ask it questions and do all the basic things that you would do via prompting right now.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think it's helpful, but it still feels to me a bit like slop.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'd like to do it properly, and I think the capability is not there for what I would want.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What I'd want is like an actual tutor experience.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe a prominent example in my mind is I was recently learning Korean, so language learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I went through a phase where I was learning Korean by myself on the internet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I went through a phase where I was actually part of a small class in Korea, taking Korean with a bunch of other people, which was really funny.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But we had a teacher and like 10 people or so taking Korean.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I switched to a one-on-one tutor.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I guess what was fascinating to me is I think I had a really good tutor, but I mean, just thinking through like what this tutor was doing for me and how incredible that experience was and how high the bar is for like what I actually want to build eventually.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because I mean, she was extremely, so she instantly from a very short conversation understood like where I am as a student, what I know and don't know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And she was able to like probe exactly like the kinds of questions or things to understand my world model.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

No LLM will do that for you 100% right now, not even close, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But a tutor will do that if they're good.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Once she understands, she actually like really served me all the things that I needed at my current sliver of capability.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I need to be always appropriately challenged.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I can't be faced with something too hard or too trivial.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And a tutor is really good at serving you just the right stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so basically I felt like I was the only constraint to learning, like my own.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I was the only constraint.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I was always given the perfect information.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm the only constraint.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I felt good because I'm the only impediment that exists.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's not that I can't find knowledge or that it's not properly explained or et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like it's just my ability to memorize and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this is what I want for people.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How do you automate that?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So a very good question about the current capability, you don't.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do think that with, and that's why I think it's not actually the right time to actually build this kind of an AI tutor.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I still think it's a useful product, and lots of people will build it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I still feel like the bar is so high, and the capability is not there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I mean, even today, I would say ChachiBT is an extremely valuable educational product.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think for me, it was so fascinating to see how high the bar is.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And when I was with her, I almost felt like, there's no way I can build this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Anyone who's had a really good tutor is like, how are you going to build this?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess I'm waiting for that capability.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that in a lot of ways in the industry, for example, I did some AI consulting for computer vision.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

A lot of my times, the value that I brought to the company was telling them not to use AI.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It wasn't like, I was the AI expert, and they described a problem, and I said, don't use AI.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This was my value add.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I feel like it's the same in education right now, where I kind of feel like, for what I have in mind, it's not yet the time, but the time will come.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But for now, I'm building something that looks maybe a bit more conventional, that has a physical and digital component and so on.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think there's obvious, it's obvious how this should look like in the future.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, so I'm building the first course, and I want to have a really, really good course.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

State-of-the-art, obvious state-of-the-art destination you go to learn AI, in this case, because that's just what I'm familiar with, so I think it's a really good first product to get to be really good.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so that's what I'm building, and NanoChat, which you briefly mentioned, is a capstone project of LLM101N, which is a class that I'm building.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So that's a really big piece of it, but now I have to build out a lot of the intermediates, and then I have to actually, like, hire a small team of, you know, TAs and so on, and actually, like, build the entire course.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And maybe one more thing that I would say is, like, many times when people think about education, they think about sort of, like, the more...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

what I would say is like kind of a softer component of like diffusing knowledge or like, but I actually have something very hard and technical in mind.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so in my mind, education is kind of like the very difficult technical like process of building ramps to knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So in my mind, NanoChat is a ramp to knowledge because it's a very simple, it's like the super simplified full stack thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If you give this artifact to someone and they like look through it, they're learning a ton of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so it's giving you a lot of what I call Eurekas per second, which is like understanding per second.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's what I want.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Lots of Eurekas per second.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so to me, this is a technical problem of how do we build these ramps to knowledge.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I always think of Eureka as almost like a, it's not like maybe that different, maybe through some of the frontier labs or some of the work that's going to be going on, because I want to figure out how to build these frontier, these ramps very efficiently so that people are never stuck.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And everything is always not too hard or not too trivial.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you have just the right material to actually progress.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, I think you always have to be calibrated to what the capability, what capability exists in the industry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think a lot of people are going to pursue like, oh, just ask Chachi PT, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think like right now, for example, if you go to Chachi PT and you say, oh, teach me AI, there's no way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, it's going to give you some slop, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like when I, AI is never going to write nano chat right now, but nano chat is a really useful, I think, intermediate point.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I still, I'm collaborating with AI to create all this material.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So AI is still fundamentally very helpful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Earlier on, I built a CS231 at Stanford, which was one of the earlier, actually, sorry, I think it was the first deep learning class at Stanford, which became very popular.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And the difference in building out 231N and LLM 101N now is quite stark because I feel really empowered by the LLMs as they exist right now, but I'm very much in the loop.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So they're helping me build little materials.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I go much faster.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They're doing a lot of the boring stuff, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I feel like I'm developing the course much faster and those LLM infused in it, but it's not yet at a place where I can creatively create the content.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm still there to do that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think the trickiness is always calibrating yourself to what exists.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think it would change over time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think right now it would be hiring faculty to help work hand-in-hand with AI and a team of people probably to build state-of-the-art courses.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I think over time it can, maybe some of the TAs can actually become AIs because some of the TAs like, okay, you just take all the course materials and then I think you could serve a very good like automated TA for the student when they have more basic questions or something like that, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I think you'll need faculty for the overall architecture of a course and making sure that it fits.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I kind of see a progression of how this will evolve and maybe at some future point, you know, I'm not even that useful and AI is doing most of the design much better than I could.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I still think that that's going to take some time to play out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

No, I will hire faculty, I think, because there are domains in which I'm not an expert.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think that's the only way to offer the state-of-the-art experience for the student, ultimately.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, yeah, I do expect that I would hire faculty, but I will probably stick around in AI for some time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do have something, I think, more conventional in mind for the current capability, I think, than what people would probably anticipate.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And when I'm building Starfleet Academy, I do probably imagine a physical institution and maybe a tier below that, a digital offering that is not the state-of-the-art experience you would get when someone comes in physically full-time and we work through material from start to end and make sure you understand it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's the physical offering.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The digital offering is, yeah, a bunch of stuff on the internet and maybe some LLM assistant and it's a bit more gimmicky and a tier below, but at least it's accessible to like eight billion people.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think there's going to have to be a lot of not just education, but also re-education.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I would love to help out there because I think the jobs will probably change quite a bit.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so, for example, today, a lot of people are trying to upskill in AI specifically.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think it's a really good course to teach in this respect.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And yeah, I think the motivation-wise, before AGI, motivation is very simple to solve because people want to make money.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And this is how you make money in the industry today.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think post-AGI is a lot more interesting, possibly, because, yeah, if everything is automated and there's nothing to do for anyone, why would anyone go to a school, et cetera?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think, I guess, like, I often say that pre-AGI education is useful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Post-AGI education is fun.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And in a similar way, as people, for example, people go to gym today, but we don't need their physical strength to manipulate heavy objects because we have machines that do that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

They still go to gym.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Why do they go to gym?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, because it's fun, it's healthy, and you look hot when you have a six-pack.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I guess like...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I guess what I'm saying is it's attractive for people to do that in a certain like very deep psychological evolutionary sense for humanity.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I kind of think that education will kind of play out in the same way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like you'll go to school, like you go to gym.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think that right now, I think not that many people learn because learning is hard.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You bounce from material.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And some people overcome that barrier, but for most people, it's hard.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I do think that it's a technical problem to solve.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's a technical problem to do what my tutor did for me when I was learning Korean.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think it's tractable and buildable, and someone should build it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think it's going to make learning anything trivial and desirable, and people will do it for fun because it's trivial.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

If I had a tutor like that for any arbitrary piece of, like, knowledge, I think it's going to be so much easier to learn anything.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And people will do it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they'll do it for the same reasons they go to gym.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think this, so I do definitely feel like people will be, I do think like eventually it's a bit of a losing game, if that makes sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that it is in long term.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Long term, which I think is longer than I think maybe most people in the industry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's a losing game.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that people can go so far and that we barely scratch the surface of how much a person can go.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And that's just because people are bouncing off of material that's too easy or too hard.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I actually kind of feel that people will be able to go much further.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like anyone speaks five languages, because why not?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because it's so trivial.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Anyone knows, you know, all the basic curriculum of undergrad, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I kind of feel like I am betting a little bit implicitly on some of the timelessness of human nature.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think it will be desirable to do all these things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think people will look up to it as they have for millennia.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think this will continue to be true.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And actually also maybe there's some evidence of that historically because if you look at, for example, aristocrats or you look at maybe ancient Greece or something like that, whenever you had little pocket environments that were post-AGI in a certain sense,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do feel like people have spent a lot of their time flourishing in a certain way, either physically or cognitively.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I think I feel okay about the prospects of that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think if this is false and I'm wrong and we end up in like, you know, WALL-E or idiocracy future, then I think it's very, I don't even care if there's like Dyson spheres.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This is a terrible outcome.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like I actually really do care about humanity.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like,

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Everyone has to just be superhuman in a certain sense.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, maybe.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually think that... I think there will be a transitionary period where we are going to be able to be in the loop and advance things if we actually understand a lot of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do think that long-term, that probably goes away, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But maybe it's going to even become a sport.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like right now, you have powerlifters who go extreme on this direction.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So what is powerlifting in a cognitive era?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe it's people who are really trying to make Olympics out of knowing stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And if you have a perfect AI tutor, maybe you can get extremely far.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I almost feel like we're just barely... The geniuses of today are barely scratching the surface of what a human mind can do, I think.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'm similar for that matter.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I mean, a lot of people, for example, hate school and want to get out of it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I really liked school.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I loved learning things, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I wanted to stay in school.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I stayed all the way until PhD, and then they wouldn't let me stay longer, so I went to the industry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I mean, basically, roughly speaking, I love learning, even for the sake of learning, but I also love learning because it's a form of empowerment and being useful and productive.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because it feels bad to bounce from material.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It feels bad.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You get negative reward from sinking an amount of time in something and this doesn't pan out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Or like being completely bored because what you're getting is too easy or too hard.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think, yeah, I think when you actually do it properly, learning feels good.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think it's a technical problem to get there.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think for a while it's going to be AI plus human collab, and at some point maybe it's just AI.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think that's a pretty broad topic.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do feel like there's basically, I almost feel like there are 10, 20 tips and tricks that I kind of semi-consciously probably do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I guess like on a high level, I always try to, I think a lot of this comes from my physics background.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I really, really did enjoy my physics background.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I have a whole rant when I think how everyone should learn physics in early school education, because I think early school education is not about

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

crumbling knowledge or memory for tasks later in the industry.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's about booting up a brain.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think physics uniquely boots up the brain the best because some of the things that they get you to do in your brain during physics is extremely valuable later.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The idea of building models and abstractions and understanding that there's a first order of approximation that describes most of the system, but then there's a second order, third order, first order terms that may or may not be present.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And the idea that you're observing like a very noisy system, but actually there's like these fundamental frequencies that you can abstract away.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like when a physicist walks into the class and they say, assume there's a spherical cow and dot, dot, dot.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And everyone laughs at that, but actually it's brilliant.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's brilliant thinking.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's very generalizable across the industry because, yeah, cows can be approximated as a sphere, I guess, in a bunch of ways.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's a really good book, for example, Scale.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's basically from a physicist talking about biology.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And maybe this is also a book I would recommend reading.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But you can actually get a lot of really interesting approximations and chart scaling laws of animals.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you can look at their heartbeats and things like that, and they actually line up with the size of the animal and things like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can talk about an animal as a volume, and you can actually derive a lot of...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can talk about the heat dissipation of that because your heat dissipation grows as the surface area, which is growing a square, but your heat creation or generation is growing as a cube.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I just feel like physicists have all the right cognitive tools to approach problem solving in the world.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I think because of that training, I always try to find the first order terms or the second order terms of everything.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

When I'm observing a system or a thing, I have a tangle of a web of ideas or knowledge in my world, in my mind.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I'm trying to find what is the thing that actually matters?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

What is the first order component?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How can I simplify it?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

How can I have a simple thing that actually shows that thing, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It shows an action.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I can tack on the other terms.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe an example from one of my repos that I think illustrates it well is called micrograd.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't know if you're familiar with this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So micrograd is 100 lines of code that shows backpropagation.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can create neural networks out of simple operations like plus and times, et cetera, Lego blocks of neural networks.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you build up a computational graph, and you do a forward pass and a backward pass to get the gradients.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Now, this is at the heart of all neural network learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So MicroGrad is a 100 lines of pre-interpretable Python code, and it can do forward and backward arbitrary neural networks, but not efficiently.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So MicroGrad, these 100 lines of Python, are everything you need to understand how neural networks train.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Everything else is just efficiency.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Everything else is efficiency.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And there's a huge amount of work to do efficiency.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You need your tensors, you lay them out, you stride them, you make sure your kernels are orchestrating memory movement correctly, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's all just efficiency, roughly speaking.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But the core intellectual sort of piece of neural network training is micrograph.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's 100 lines.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You can easily understand it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're chaining.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It's a recursive application of chain rule to derive the gradient, which allows you to optimize any arbitrary differential function.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So I love finding these, like, you know, the smaller terms and serving them on a platter and discovering them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I feel like education is, like, the most intellectually interesting thing because you have a tangle of understanding, and you're trying to lay it out in a way that creates a ramp

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

where everything only depends on the thing before it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I find that this, like, you know, untangling of knowledge is just so intellectually interesting as a cognitive task.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so I love doing it personally, but I just have fascination with trying to lay things out in a certain way.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe that helps me.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah, you're presenting the pain before you present a solution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And how clever is that?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you want to take the student through that progression.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So there's a lot of other small things like that that I think make it nice and engaging and interesting.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And always prompting the student.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

There's a lot of small things like that that I think are important and a lot of good educators will do.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, how would you solve this?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, I'm not going to present a solution before you're going to guess.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That would be wasteful.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's a little bit of a...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't want to swear, but, like, it's a dick move towards you to present you with the solution before I give you a shot to try to come up with it yourself.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Well, you have a chance to try yourself, and you have an appreciation when I give you the solution.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it maximizes the amount of knowledge per new fact added.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

That's right, yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

well as the curse of knowledge and expertise yeah this is a real phenomenon and i actually suffered from it myself as much as i try to not not suffer from it but you take certain things for granted and you can't put yourself in the shoes of new of people who are just starting out and this is pervasive it happens to me as well one thing that i actually think is extremely helpful as an example someone was trying to show me a paper in biology recently and i just had instantly so many terrible questions so what i did was i used chat gpt to ask the questions with the with the paper in the context window

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then it worked through some of the simple things.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then I actually shared the thread to the person who shared it, who actually like wrote that paper or like worked on that work.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I almost feel like it was like, like if they can see the dumb questions I had, it might help them explain better in the future or something like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Because, so for example, for my material, I would love if people shared their dumb conversations with Chachi PT about the stuff that I've created, because it really helps me put myself again in the shoes of someone who's starting out.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

To say the thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Yeah.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Actually, I saw that tweet.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I thought it was really good.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I shared it with a bunch of people, actually.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think it was really good.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I noticed this many, many times.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Maybe the most prominent example is I remember back in my PhD days doing research, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You read someone's paper, right?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you work to understand what it's doing, et cetera.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then you catch them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're having beers at the conference later.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you ask them, so, like, this paper, like, so, what were you doing?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, what is the paper about?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they will just tell you these, like, three sentences that, like, perfectly capture the essence of that paper and totally give you the idea.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And you didn't have to read the paper.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And, like, it's only when you're sitting at the table with a beer or something like that and, like, oh, yeah, the paper is just, oh, you take this idea, you take that idea, and you try this experiment, and you try this thing.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And they have a way of just putting it conversationally.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And just, like, perfectly, like, why isn't that the abstract?

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I don't actually know that I have unique tips and tricks, to be honest.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Basically, it's kind of a painful process.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But, you know, redraft one.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think one thing that has always helped me quite a bit is...

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I had a small tweet about this, actually.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

So, like, learning things on demand is pretty nice.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Learning depth-wise.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I do feel like you need a bit of alternation of learning depth-wise on demand.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You're trying to achieve a certain project that you're going to get a reward from.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And learning breadth-wise, which is just, oh, let's do whatever one-on-one.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And here's all the things you might need.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Which is a lot of school does a lot of breadth-wise learning.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, oh, trust me, you'll need this later.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

You know, that kind of stuff.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

Like, okay, I trust you.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I'll learn it because I guess I need it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

But I love the kind of learning where you'll actually get a reward out of doing something and you're learning on demand.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

The other thing that I've found is extremely helpful is maybe this is an aspect where education is a bit more selfless because explaining things to people is a beautiful way to learn something more deeply.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

This happens to me all the time.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think it probably happens to other people too because

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I realize if I don't really understand something, I can't explain it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I'm trying and I'm like, actually, I don't understand this.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And it's so annoying to come to terms with that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And then you can go back and make sure you understood it.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And so it fills these gaps of your understanding.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

It forces you to come to terms with them and to reconcile them.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I love to re-explain and things like that.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

And I think people should be doing that more as well.

Dwarkesh Podcast
Andrej Karpathy — AGI is still a decade away

I think that forces you to manipulate knowledge and make sure that you know what you're talking about when you're explaining it.