Turing Award winner Richard Sutton, author of the 2019 "Bitter Lesson" essay, argues that raw compute scaling alone is insufficient for achieving intelligence, asserting AI science remains "far from finished." In a new post sparking significant discussion, Sutton reiterates that the real world's infinite complexity renders synthetic data a "big mistake" and human expertise a bottleneck preventing true scaling. He advocates for agents possessing world models capable of continuous learning from experience, contrasting with current LLMs that cease learning post-training. AI researcher Jerry Tworek contextualizes this, noting that while the Bitter Lesson holds, algorithm choice becomes critical as compute increases; he warns that transient transformer success has "reward-hacked" a generation of scientists into ignoring unexplored scaling dimensions. Jayden Teoh adds that a rigid interpretation of the lesson has stifled academic research by dismissing promising non-conforming ideas. This tension fuels a wave of new labs—45 tracked in Frontier Radar #5, including Tworek's Core Automation and Sutton's Oak Lab—pursuing learning algorithms beyond pre-training and RL. Conversely, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei maintain LLM scaling is far from exhausted, with Altman citing a model disproving a mathematical conjecture as evidence, though their positions align with their companies' technical bets.