Richard Sutton warns LLM scaling enthusiasts raw compute isn't enough
turing award winner richard sutton says ai science needs new methods beyond scaling compute for true intelligence.
richard sutton, a turing award winner, has expressed concerns that simply scaling large language models with more compute is not enough to achieve true intelligence. in his 2019 essay 'the bitter lesson,' sutton emphasized that general methods scaling with compute beat sophisticated human-engineered approaches. he now advocates for agents with world models that continuously learn from their own experience, as the real world is infinitely complex and human expertise is a bottleneck. other researchers, such as jerry tworek and jayden teoh, agree that new methods are needed, but also note that the bitter lesson has been misread by some as dismissing promising ideas that don't follow its spirit. this debate has implications for the future of ai research and development.