Researcher: JEPA-titled models are usually bad in practice, V-JEPA 2.1 the only exception
serrjoa · x · 2026-11-07
Researcher gabriberton publicly calls out that models with JEPA in their title usually perform poorly in practice—LeJepa is much worse than DINO—despite pretty papers and fancy math, with Meta's V-JEPA 2.1 likely the only exception. The quote-post urges the community to be braver about openly criticizing papers, arguing silence lets bad ideas spread fast, and thanks @giffmana for saving people an hour of reading time.
More from Research
- ThinkJev: threshold-gated System 1/2 reasoning that stops thinking once confident — RobertTLange · 2026-11-07
- UCLA and Cedars-Sinai build a glioblastoma mouse model that resists immunotherapy like patients — anne_churchland · 2026-11-07
- Reorganization effect can reverse technology effect on AI-era wages, paper shows — soumitrashukla9 · 2026-11-07
- Meta's real-time avatars: distillation recipe cuts 120 model evals to 2, a 60x speedup for infinite video — altryne · 2026-11-07
- Free primer book derives math foundations from VAEs and diffusion to WGANs — kalyan_kpl · 2026-11-07
- Kenneth Stanley at ARC Prize summit: why open-endedness is essential to AGI but brutal to benchmark — GregKamradt · 2026-11-07