What's this about?
Open data and scalable training for long-horizon video world models. 18 upvotes.
Why does "open data" lead the title?
Because in world modeling the data is the moat. Everyone knows roughly how to train these; almost nobody has the video corpora with the properties required — long, continuous, camera-controlled, with enough scene diversity. Releasing the data changes who can participate far more than releasing another architecture would. Long-horizon is the specific hard part: staying coherent over minutes rather than seconds, which is the problem LayerRecall, Ring Forcing and Matrix-Game were all attacking from the memory side this week.
Why is it worth noticing?
- Open data releases are rarer and more consequential than open weights in this subfield.
- Fourth long-horizon video paper this week, which makes it a clear active front.
- Modest upvotes for something that may matter more than several higher-ranked entries.