Platonic Representation Hypothesis is now a theorem: JEPAs align up to one rotation
randall_balestr · x · 2026-10-30
A team working with Yann LeCun turns the Platonic Representation Hypothesis into a theorem: train JEPAs separately on any modalities and, with good views (two captions of one photo, two frames of one video), they provably learn the same embedding up to one rotation. This yields a 109M text-JEPA that aligns with vision models better than 72B LLMs, plus an alignment scaling law showing good-view JEPAs scale far better than generative LLMs. Paper, code, and a live demo (try your own photos/videos) are public.
More from Research
- Forethought: first movers could lock rivals out of space for $49-73B premium — willmacaskill · 2026-10-31
- Correction: newer model gets worse on the extreme-hard tail, pass@5 crossover predicted — teortaxesTex · 2026-10-31
- Meta demos titanRL, TorchTitan's RL library for LLMs, teaching Qwen Connect 4 — ezyang · 2026-10-31
- Biohub's Remo-ID spots sepsis from blood cell images, 43/47 correct in early study — alex_peys · 2026-10-31
- 240-Year-Old Voting Paradox Gets a Shape: It's a Klein Bottle — stevenstrogatz · 2026-10-31
- OpenAI releases 700+ math manuscripts in one batch; Fortnow says it beats 30 years of progress — kimmonismus · 2026-10-31