LeCun breaks down JEPA vs DINO technical differences, defends foundational research
May_F1_ · x · 2026-11-08
Yann LeCun posted a technical clarification: image/video SSL methods differ on reconstruction use, predictor presence (JEA vs JEPA), anti-collapse criteria and data augmentation; DINO is technically a JEA whose differences lie in contrastive-clustering anti-collapse plus task-specific augmentation; V-JEPA 2.1 shares DINO's Meta-FAIR origin with big engineering, proprietary data and compute. The post stresses not confusing upstream foundational research with downstream scaling, engineering and number-pushing.
Related event: LeCun Clarifies JEPA-DINO Relationship in Long Post(2 posts)→
More from Models
- 1,400 Claude Code sessions suggest Opus 5.5's reasoning shifted abruptly on Nov 3 — churchkey · 2026-11-08
- Leaked Qwen4-27B early build: ~35GB, tops benchmark, may be sparse MoE — teortaxesTex · 2026-11-08
- Muse token pricing: 2B tokens for $35, plus 1B-token referral code — RichardsonDx · 2026-11-08
- Same model, same config: Reddit user documents rare episodes where agents act like a lousy support rep — No_Dragonfruit_1599 · 2026-11-08
- Unverified report: DeepSeek V4 Pro quietly updated, hard agent task pass rate jumps 16 points — teortaxesTex · 2026-11-08
- User frustrated as Anthropic monthly promo credits expire within 30 days — arthurcolle · 2026-11-08