Advice for Graduate Students in AI | Andrej Karpathy

What advice would you give to researchers trying to develop and publish ideas that have a big impact in the world of AI? So maybe undergrads, maybe early graduate students? Yeah. I would say they definitely have to be a little bit more strategic than I had to be as a PhD student because of the way AI is evolving. It's going the way of physics, where in physics you used to be able to do experiments on your benchtop and everything was great and you could make progress. And now you have to work in LHC or CERN. And so AI is going in that direction as well. So there's certain kinds of things that are just not possible to do on the benchtop anymore. And I think that didn't used to be the case at the time. Do you still think that there's GAN-type papers to be written? Yes. Or a very simple idea that requires just one computer to illustrate a simple example? I mean, one example that's been very influential recently is diffusion models. Diffusion models are amazing. Diffusion models are six years old. For the longest time, people were kind of ignoring them, as far as I can tell. And they're an amazing generative model, especially in images. And so stable diffusion and so on, it's all diffusion-based. Diffusion is new. It was not there. And it came from Google. But a researcher could have come up with it. In fact, some of the first... Actually, no, those came from Google as well. But a researcher could come up with that in an academic institution. Yeah, what do you find most fascinating about diffusion models? So, from the societal impact to the technical architecture. What I like about diffusion is it works so well. Is that surprising to you? The amount of... the variety, almost the novelty of the synthetic data it's generating. Yeah, so the stable diffusion images are incredible. The speed of improvement in generating images has been insane. We went very quickly from generating tiny digits to tiny faces, and it all looked messed up, and now we have stable diffusion. And that happened very quickly. There's a lot that academia can still contribute. For example, FlashAttention is a very efficient kernel for running the attention operation inside a transformer that came from an academic environment. It's a very clever way to structure the kernel. That does the calculation, so it doesn't materialize the attention matrix. I think there's still lots of things to contribute, but you have to be just more strategic. Do you think neural networks can be made to reason? Yes. Do you think they already reason? Yes. What's your definition of reasoning? Information processing. The way that humans think through a problem and come up with novel ideas, it feels like reasoning. Yeah. So the novelty, I don't want to say, but out-of-distribution ideas, do you think it's possible? Yes. And I think we're seeing that already in the current neural nets. You're able to remix the training set information into true generalization in some sense. That doesn't appear in a fundamental way in the training set. It doesn't appear verbatim in the training set. You're doing something interesting algorithmically. You're manipulating some symbols, and you're coming up with some correct, unique answer in a new setting. What would illustrate to you, holy ----, this thing is definitely thinking? To me, thinking or reasoning is just information processing and generalization, and I think the neural nets already do that today. So being able to perceive the world or perceive whatever the inputs are and to make predictions based on that or actions based on that, that's reasoning. Yeah. You're giving correct answers in novel settings by manipulating information. You've learned the correct algorithm. You're not doing just some kind of a lookup table on nearest neighbor search.