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

Why watch

Rafael Levi discusses how physical AI and robotics face a critical data shortage—while LLMs train on trillions of words, robotics has only ~1 million training videos. He explains why staged recordings produce biased data, why simulation and teleoperation don't scale, and how the web contains billions of hours of real human actions but with significant noise (NVIDIA discards 96% of video for Cosmos). The talk covers Bright Data's solution: indexing over a billion videos by actions rather than titles and returning trimmed, timestamped clips via API for training robotics models.