Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again

The pursuit of artificial intelligence requires moving beyond static, human-curated datasets toward systems capable of continual learning through direct experience. Reinforcement learning pioneer Rich Sutton and co-founder Khurram Javed argue that the "Bitter Lesson"—the observation that computational scaling eventually outperforms human-designed priors—necessitates a fundamental paradigm shift. Current large language models, while impressive, suffer from catastrophic forgetting and lack the ability to learn after deployment. True intelligence demands a "Big World" approach where agents continuously update their internal models through interaction with an infinitely complex environment. By integrating meta-learning for step-size optimization and generative feature discovery, this approach aims to build self-maintaining, coherent minds that evolve beyond the limitations of current pre-training methods. This vision, central to the research agenda at Oak Lab, prioritizes adaptive, experiential learning over the finite, human-bottlenecked data streams currently dominating the field.