2026-09-27 · America/Los_Angeles · 社区动态 · #15

“I want to be the very best” (re-examining LLM progress on Pokemon)

It's natural to wonder how much LLM progress is "real" intelligence versus memorization or overfitting or whatever. LLMs now play Pokemon Red very well (arguably better than a human child)... but the game is famous, has been an RL/DRL target for years, and the internet has every form of training data you could name (from screenshots to detailed walkthroughs to input sequences that will automatically win). It's also possible that labs now have specialized RL environments for games - notably, Anthropic soon stopped claiming that Claude had never been explicitly trained to play Pokemon.…

热度 45.0 / 100；排名与评分保留该期记录。