New York Times

训练大模型的真门槛不是算力是数据 — 而好数据正变得越来越难拿

Common Crawl's 300B pages anchor every LLM. Books vanish over copyright; Reddit/Stack Overflow now charge. What you feed now beats which architecture.