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

Rich Sutton discusses how the concept of learning being continual is often misunderstood or misrepresented. He emphasizes that all learning is inherently ongoing—it's not a separate phase— and criticizes the field for labeling it 'continual learning' as if it were a special or unusual process. Sutton reflects on his career, his early conviction to focus on reinforcement learning, and his resilience in pursuing this path even during challenges like his illness. He advocates for a view of intelligence that incorporates high-level abstraction, sensory-motor skills, and the ability to learn from experience across various domains, aligning it with how humans and animals learn and adapt.

Sutton recaps his influential essay, 'The Bitter Lesson,' which states that long-term progress in AI stems from leveraging computation rather than human knowledge or prior assumptions. He highlights how deep learning and scaling compute have been the driving forces behind advancements like AlphaGo and language models. Sutton warns against over-reliance on human-curated data or synthetic data, emphasizing the vast complexity of the world and the importance of models that learn from raw experience. The discussion also covers how current AI systems, including large language models, are limited in their ability to learn continually and avoid catastrophic forgetting, and how new algorithms like continual backprop can help overcome these limitations.

The conversation delves into the shortcomings of present AI systems, especially in terms of continual learning and model abstraction. Sutton argues that large language models, while a major scientific breakthrough in language use, represent only a fraction of intelligence and are static once trained. He calls for AI that can form deeper models, learn from its own experience, and adapt continuously across many tasks and environments. The vision includes building a unified, self-consistent AI that maintains and updates its knowledge base, learns across a broad spectrum of scales, and can handle paradigm shifts through autonomous experience.

Sutton and Javed outline their ambitious plan with Oak Lab to develop AI systems capable of true continual learning, forming abstract models, and reasoning in a manner akin to human cognition. They aim to apply their new algorithms, such as continual backprop, to create AI agents that learn from experience without catastrophic forgetting, and that can adapt to different environments using a shared design. Their ultimate goal is to have a scalable, coherent AI capable of self-training and self-correction, leading toward a general intelligence. The discussion also touches on the challenges of energy efficiency, computational costs, and the necessity for a paradigm shift in AI research.