Physical AI's Next Bottleneck Is Finding the Right Video — Rafael Levi, Bright Data

AI without data is just a box." Rafael Levi of Bright Data points out that LLMs train on trillions of words, but robotics has only about a million videos of robots doing things. Paying people to record staged actions produces biased data, because nobody opens a door naturally when told to. Meanwhile, the web holds billions of hours of real people handling objects and real physics, with gravity, motion and cause and effect. The catch is noise. Levi notes that NVIDIA discards about 96% of the video it downloads to train Cosmos, and Stable Video Diffusion discards 74%, which wastes compute, bandwidth and storage. Meta, on the other hand, trained on about a million hours of public video and needed only 62 hours of real robot data to control a robot. Bright Data's approach is "search first, collect second." It indexes more than a billion videos by the actions in them, not their titles, and returns trimmed clips with timestamps, match scores and frame counts through an API. Levi demos searches for dishwashing and clothes-folding clips and covers uses in self-driving and brand discovery. Speaker info: Bright Data: https://brightdata.com Timestamps: 00:00 Intro 0:50 The AI isn't the hard part anymore. The data is 1:15 A short history of robots learning from video 2:40 Trillions of words vs. about a million robot videos 3:15 Why staged recordings make biased data 4:35 Why simulation and teleoperation don't scale 5:14 The web as a training source 6:14 1 million hours of video, 62 hours of robot data 6:54 Learning actions from frame-to-frame motion 7:58 Throwing away 96% of the video 8:33 Search first, collect second 10:03 Demo: finding the exact action clips 11:38 The API and use cases beyond robotics 12:48 What comes back: timestamps, scores and frames 13:33 Self-driving, dashcams and physics 15:52 Wrap-up