Richard Sutton, a Turing Award winner for reinforcement learning, expressed skepticism about large language models (LLMs) on the Dwarkesh Podcast, suggesting they primarily imitate human behavior rather than genuinely learn from the world. The podcast host offered a counterpoint, positing that imitation can provide a valuable starting point for reinforcement learning agents. This discussion raises the question of which approach—LLMs or agents trained without human data—would ultimately achieve greater capability given identical resources and environments. AI
IMPACT Raises fundamental questions about the learning paradigms of LLMs versus traditional reinforcement learning agents.
RANK_REASON Opinion piece from a notable researcher in the field.
Read on Mastodon — mastodon.social →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
