Andrej Karpathy on how LLMs work
Karpathy sees an LLM as a lossy compression of the internet, produced by training a Transformer to predict the next token, which forces it to absorb both knowledge and general problem-solving circuits. He insists these are alien, ghost-like artifacts of imitation rather than animal minds: weights are hazy memory, the context window is working memory, and reinforcement learning makes them jagged, superhuman where answers can be verified and clumsy everywhere else.
16 points · 6 appearances · 2022–2026
How the view evolved
Trained neural nets are alien artifacts, not brain analogues → We're summoning ghosts, not building animals
In 2022 he called trained nets alien artifacts of a compression objective; by 2025 he described them more specifically as ghosts that imitate human internet data, a different kind of intelligence that could be made more animal-like over time.
Language lacks a simple reward function for self-improvement → Computers automate what you can specify; LLMs what you can verify
In 2023 he said language has no general reward function, so self-improvement was an open question outside narrow domains; by 2026 he described labs training on verification rewards, so LLMs automate what can be verified.
LLMs have jagged intelligence: superhuman yet make silly mistakes → Computers automate what you can specify; LLMs what you can verify
In 2025 he described jaggedness as an observed quirk (strawberry, 9.11 vs 9.9); by 2026 he explained it: RL rewards verifiable domains like math and code, so capability peaks there and stagnates elsewhere.
LLMs are the kernel of an emerging operating system → Neural nets may become the host process, CPUs the co-processor
In 2023 the LLM was the kernel of a new OS running on classical computers; by 2026 he imagined a full inversion where the neural net is the host process and CPUs become co-processors.
Up next
2022–2026 · changed view
“Yeah, so I had a blog post on Software 2.0, I think several years ago now. And the reason I wrote that post is because I kept— I kind of saw something remarkable happening in like software development and how a lot of code was being transitioned to be written not in sort of like C++ and so on, but it's written in the weights of a neural net. Basically just saying that neural nets are taking over software, the realm of software, and taking more and more tasks. And at the time I think not many people understood this deeply enough, that this is a big deal, it's a big transition. Neural networks were seen as one of multiple classification algorithms you might use for your data set problem on Kaggle. Like, this is not that; this is a change in how we program computers. And I saw neural nets as, this is going to take over, the way we program computers is going to change, is not going to be people writing a software in C++ or something like that and directly programming the software. It's going to be accumulating training sets and data sets and crafting these objectives by which we train these neural nets, and at some point there's going to be a compilation process from the data sets and the objective and the architecture specification into the binary, which is really just the neural net, you know, weights and the forward pass of the neural net, and then you can deploy that binary.”
2022–2026 · changed view
“One more thing that I would say is that many times, when people think about education, they think more about what I would say is a softer component of diffusing knowledge. I have something very hard and technical in mind. In my mind, education is the very difficult technical process of building ramps to knowledge. In my mind, nanochat is a ramp to knowledge because it's very simple. It's the super simplified full-stack thing. If you give this artifact to someone and they look through it, they're learning a ton of stuff. It's giving you a lot of what I call eurekas per second, which is understanding per second. That's what I want, lots of eurekas per second. So to me, this is a technical problem of how do we build these ramps to knowledge. So I almost think of Eureka as maybe not that different from some of the frontier labs or some of the work that's going on there.”
2022–2026 · changed view
“The consistent thing I've been always interested in, for me at least, is AI, and that's probably where I'm spending my rest of my life on, because I just care about a lot. And I actually care about like many other problems as well, like say aging, which I basically view as disease, and I care about that as well, but I don't think it's a good idea to go after it specifically. I don't actually think that humans will be able to come up with the answer. I think the correct thing to do is to ignore those problems and you solve AI and then use that to solve everything else. And I think there's a chance that this will work; I think it's a very high chance. And that's kind of like the way I'm betting.”
2022–2026
“Yeah, the data engine is what I call the almost biological feeling like process by which you perfect the training sets for these neural networks. So because most of the programming now is in the level of these data sets, and make sure they're large, diverse and clean. Oh, basically you have a data set that you think is good, you train your neural net, you deploy it, and then you observe how well it's performing, and you're trying to always increase the quality of your data set. So you're trying to catch scenarios basically there are basically rare, and it is in these scenarios that the neural nets will typically struggle in, because they weren't told what to do in those rare cases in the data set. But now you can close the loop, because if you can now collect all those at scale, you can then feed them back into the reconstruction process I described and reconstruct the truth in those cases and add it to the data set. And so the whole thing ends up being like a staircase of improvement of perfecting your training set. And you have to go through deployments so that you can mine the parts that are not yet represented well in the data set. So your data set is basically imperfect, it needs to be diverse, it has pockets there are missing, and you need to pad out the pockets. You can sort of think of it that way.”