Yann LeCun发新论文挑战Scaling Law：1.09亿参数模型胜过720亿

Yann LeCun challenges scaling laws: 109M-param model beats 72B-param model

Yann LeCun’s latest work demonstrates that a compact 109M-parameter model — built on algebraic, physics-aware representations — surpasses a 72B-parameter model on structured visual-language reasoning benchmarks. It refutes the assumption that scale alone guarantees capability gains, emphasizing architectural efficiency and grounded representation instead. For AI application founders, this is a critical cost-and-speed signal: in real-world verticals — from legal doc analysis to factory defect detection — chasing state-of-the-art parameter counts adds latency, inference cost, and deployment friction without proportional ROI. Early teams should rigorously benchmark ‘minimum viable architectures’: small models (<1B params) fine-tuned on high-signal domain data, augmented with deterministic post-processing and constrained decoding. This approach slashes per-request inference cost by 5–10x, enables sub-second API response times, supports air-gapped or edge deployment, and shortens model iteration cycles from days to hours — all while improving auditability and reducing hallucination risk. Scaling smartly beats scaling blindly — every time.