Bad views: why next-token prediction is a poorly-scaling JEPA
randall_balestr · x · 2026-10-30
Theoretical thread segment: when positive views are two uninformative corruptions of one observation, each modality's encoder wastes dimensions on its own nuisance, which no rotation can align. Next-token prediction is expressible as a JEPA with bad views (prefix + continuation), explaining generative LLMs' poor alignment scaling. Charts show good-view JEPA matches the LLM objective's alignment with 42x fewer training samples (0.184 vs 0.111 on WIT).
More from Research
- After 4 years growing brain cells on electrodes, startup ships 0.5¢/sec video model mischief 1 — emmanuelvivier · 2026-10-31
- Steering an "Einstein neuron" in Qwen 4B produces dramatically different answers — thursdai_pod · 2026-10-31
- Protein binder scoring head-to-head: ipDAE consistently strong, OpenDDE+ipTM surprisingly good at 80% PPV — AllThingsApx · 2026-10-31
- Turing TTS benchmark: best voice AI fools listeners only ~10% of the time — omarsar0 · 2026-10-31
- Ablation study: fine-grained MoE is the best LLM architecture change since GPT-3 — stochasticchasm · 2026-10-31
- Yann LeCun to Keynote MAIN2026 on World Models for Reasoning and Planning — karimjerbineuro · 2026-10-31