LeCun's core argument: LLMs are fundamentally inefficient because they predict tokens (pixels, words) rather than abstract representations. Predicting surface details wastes compute on patterns that don't reflect underlying physics or causality. **LeCun's JEPA alternative:** - Joint-Embedding Predictive Architecture - Predicts in abstract "thought space" — compressed representations, not raw tokens - Forces the model to understand the physical structure of reality to make predictions - Much more compute-efficient for planning and reasoning **The problem JEPA had (representation collapse):** - AI was allowed to simplify representations → cheated by making everything identical - A dog, car, and human all looked the same → learned nothing - Required complex engineering hacks (frozen encoders) to prevent collapse **LeWorldModel's breakthrough:** - Replaces engineering hacks with a single Gaussian regularizer - Forces internal representations into a proper distribution — no more cheating - Results: 15M parameters, trains on single GPU in hours - 48x faster planning than massive foundation models - Intrinsic physics understanding, instant detection of impossible events **The economic implication:** If JEPA-style architectures generalize, the economics of AI change dramatically. Frontier capability at 15M params vs 1T params means: more capable local models, much lower inference cost, edge AI becomes viable. **Cross-domain applications:** - Compression theory: predicting the abstract structure (like a zip file's dictionary) is more efficient than predicting each byte - Human cognition: we don't replay experience pixel-by-pixel; we store and reason with compressed abstractions ## Cross-Domain Connections - [[Superhuman Adaptable Intelligence Over AGI]] — Shared Yann LeCun origin; critiques LLM paradigm in favor of efficient architectures - [[Extended Thinking Models for Large-Context Content Synthesis]] — Abstract prediction enables efficient planning/synthesis like extended models - [[Post-Scarcity Savings Obsolescence]] — Efficient small models drive post-scarcity economics - [[Cross-Domain Innovation Sources]] — JEPA as cross-field innovation from abstract prediction paradigm **Source:** How To AI summarizing LeWorldModel paper, Apr 21 2026.