World Models

Summary

Some elements here build upon the RL env post. Also, please read the footnotes, given a lot to cover, I moved many clarifications to the bottom. Something is happening across all major labs simultaneously and it’s not a coincidence. Yann LeCun announced he is leaving Meta to start a new lab focused entirely on World Models. He also has a technical lecture on what the world models are and what they do. Ilya Sutskever, on Dwarkesh’s podcast, described emotions as value functions, a framing that makes a lot of sense when you are moving away from pattern matching to planning and simulation. Google announced Genie 3, their approach to world simulation. In a probably throwaway line, Demis Hassabis revealed how he is spending most of his research time on World Models. Anthropic’s intepretability research shows that current models already develop internal world representations, but they are implicit, emergent and unreliable. OpenAI, when launching Sora insisted on it being a world simulator, as opposed to being just a video model. Similarly, Veo3 is also referred to as a physics model or a world model. Meta released a paper on code world model (cwm) in September last year, where a 32B model matched or outperformed larger models on execution-dependent benchmarks. When every major lab converges on the same research direction within the same window, we should pay attention. What is a World model? A world model predicts the next state or observation. The objective is to understand the causal laws of the environment where environment can be a videogame, codebase, or a market. This is distinct from what the current systems do: A transformer predicts the next token (imitation or pattern matching - what comes next in a sequence) A reasoning model (trained via RLVR) is optimizing for a reward from its training (reasoning or what the inference step follows) A world model predicts the next state (simulation i.e. what the world looks like after an intervention) For a code world model, ...