Back in 2014, Elon Musk referred to AI as summoning the demon. And it wasn’t hard to see that view. Soon, Go agents would beat top humans learning from self play. By the end of 2017, the same algorithm mastered Chess and Shogi. By 2020, it didn’t even need tons of calls to the simulator, and could play Atari too.

AI looked scary. It looked like it was one FOOM away from self playing and becoming superhuman at the universe. And yet, here we are in 2023 and self driving cars still don’t work.

Does becoming superhuman at the universe make any practical sense? The universe has so many orders of magnitude more states than any Go game. And I don’t even need math to illustrate this point, just imagine tiling the universe with as many very tiny Go boards as you can fit.

That’s a lot of Go boards. So many that the difference in complexity between Go and the Universe is not a matter of number, it’s a matter of kind. Every Go program operates at the stone level. Almost nothing predicting the world operates at the atom level.

These modern self play systems like MuZero have some form of dynamics model, a function that given the current state of the world and an action, it predicts the next state. We also use these dynamics models at comma. They are world models, and contain all the knowledge of how the world works inside of them.

GPT-4 is a dynamics model also, conditioned on the prior of the action space. And there’s an even simpler way to think about it. The loss function for dynamics is compression.

Compression is prediction is intelligence, intelligence is prediction is compression. One of the coolest facts