AI Learns Physics Through YouTube Videos
In this brief monologue, Tim Davies describes how large‑scale video models trained on decades of YouTube footage develop an implicit understanding of physical dynamics. By ingesting massive amounts of visual data, these models learn relationships between objects much like language models learn word relationships from text. Davies highlights emergent capabilities such as object persistence—accurately predicting where a thrown ball will land even when the motion is off‑screen—and an intuitive grasp of inertia and friction, phenomena that seem impossible without explicit physics supervision. The talk frames these findings as a natural consequence of massive video corpora, suggesting that visual AI can acquire commonsense physics simply by observing the world through video.
Chapters
More from Tim
EP 60: Why Founders Become the Bottleneck — And How AI Changes Everything | Thea Brook
