LeCun team proves Platonic Representation Hypothesis: good views beat scale by 42x
galilai-group · hf · 2026-10-30
Randall Balestriero and Yann LeCun (Brown/NYU/AMI Labs) turn the Platonic Representation Hypothesis into a theorem: separately trained JEPAs converge to shared representations when positive pairs are two views of one scene, while next-token prediction acts as a JEPA with bad views. Good views need 42x fewer training samples for the same alignment, and a 109M-parameter text JEPA aligns with DINOv2 and CLIP image encoders better than LLMs up to 72B—plus one rotation turns encoders into an image captioner (CIDEr .838 on COCO). Positive view design, not scale, is the lever.
More from Multimodal
- Gracia demos World Reconstruction Model: free-viewpoint basketball replays from 8 cameras, near real-time — janusch_patas · 2026-10-31
- Element Forge: open-source editable AI VFX for real footage in ComfyUI — zitroffell · 2026-10-30
- Claymation body horror on 8GB: three wrong settings, not the GPU, made local renders slow — roro9- · 2026-10-30
- AI-generated Haribo Starmix bean bag product shot looks like a real catalog page — Delahuntagram · 2026-10-30
- MiniMax LoRA creator ships open-source ComfyUI alternative to Higgsfield's Influencer Studio — infroy28 · 2026-10-30
- Scenario open-sources 3D OS: 80 agent skills taking sketches to game-ready assets — Scobleizer · 2026-10-30