Meta researchers trained a model on 2 million hours of video with no labels, no physics textbooks, and no supervision of any kind. The model apparently picked up physical intuitions purely from watching the world.
The approach is unsupervised throughout - no human annotations, no structured curriculum. Just raw video at scale, and a model left to figure out patterns on its own.
It’s an interesting result, though the details of what exactly the model “learned” and how that’s being measured matter a lot here. Learning to predict the next frame of a video is not the same as understanding physics, and that distinction tends to get blurry in the coverage of these announcements. Worth watching for the full paper to see what the benchmarks actually show.

Collections

Meta researchers trained a model on 2 million hours of unlabeled video using purely unsupervised learning, with no human annotations or structured curriculum. The model reportedly developed physical intuitions from raw video alone. The post notes important caveats: predicting video frames is not equivalent to understanding physics, and the actual benchmarks from the full paper will be critical to evaluating the real significance of the result.

Meta trained a model on 2 million hours of unlabeled video