Reinforcement learning pioneer Richard Sutton today reiterated his famous 'bitter lesson' in 26 words, calling for AI development to focus on general methods that scale with compute power rather than being distracted by human knowledge. The statement has reignited debates about the direction of large language models.
https://twitter.com/RichardSSutton/status/2056419165502935198
Prominent AI critic and cognitive scientist Gary Marcus pushed back directly, arguing that current large models are still heavily dependent on human knowledge as their primary fuel — even hardcoded into system prompts and over 50 custom tools. Machine learning founder Thomas G. Dietterich also disagreed, saying that studying human abstract cognitive structures helps identify AI's blind spots and should never be considered a 'distraction.'
A long post titled 'The Sweet Lesson' surfaced in the comments offering a more nuanced perspective. John Fletcher, co-founder and chief scientist at The Innovation Game, pointed out Sutton's strategic blind spot: while the endgame for AI may be pure computation, the path to that endgame cannot bypass humans when written data is running out. The way companies are pulling ahead now is by building top-tier collaborative tools to frequently capture 'tacit knowledge' from experts — knowledge that has never been written down. Skipping this phase and jumping straight to environment interaction learning is simply not practical from an engineering standpoint.