What's this about?
H3-World turns language understanding into world control — using the model's linguistic knowledge to act in an environment (42 upvotes). ZimaBlue evolves generalizable world action models through scalable video pre-training (40 upvotes).
What is the bridge these papers are trying to build?
Language models know an enormous amount about how the world works, absorbed from text: that glass breaks, that doors open toward you or away, that a full cup spills. What they cannot do is act on it, because knowing and controlling are different representations. Turning language understanding into world control means finding the mapping between the two — and it is attractive precisely because the language side is already trained and enormous, while embodied data is scarce and expensive.
Why is it worth noticing?
- Both belong to the world-model cluster that has led HuggingFace for three weeks.
- Video pre-training as the route to action models is the same bet as last week's agentic-game-development paper.
- The name ZimaBlue is a Black Mirror reference, which tells you something about who wrote it.