Here's my talk from the CVPR 2026 "Bitter Lessons" workshop earlier this summer. I've split it up into parts for the sake of discussion.
Part 1: in which I wax poetic about my youth and force the audience to look at my dissertation results.
Atlas understands both the spatial structure of the world and how it evolves over time.
This lets us turn a handful of ordinary cameras into a "bullet time" multiview capture studio, capturing dynamic moments frozen in time.
Someone please make this journal:
- submission has a large fee
- a LLM reads the paper and accepts/rejects instantly
- the LLM tunes on all accepts ~daily, and measures the info gain from each
- fees are paid back to authors proportionally to info gain, minus training costs
One issue with the autoregressive video model approach to games: gamedevs tell you to make the first level *last*, because only after the game is done do you know how to best onboard new players. The world model is forced to make the game in exactly the opposite order it should.