Rich Sutton on AI creativity and discovery (twitter.com)

In a recent presentation, AI pioneer Rich Sutton argues that current generative AI systems, which rely heavily on supervised learning, significantly lack the ability to make genuine discoveries. He highlights that while these systems can produce outputs that seem novel or useful, they often fail to deliver both novelty and quality at the same time. This limitation stems from their reliance on vast amounts of training data, resulting in outputs that mimic existing content rather than truly innovate. Sutton illustrates this with the analogy of a researcher whose work is labeled as either "good" or "novel," but never both, emphasizing that generative AI tends to produce variations without proper evaluation. Sutton proposes that true creativity and discovery in AI require a structured process that involves variation, evaluation, and selective retention—principles he connects to historical concepts like the scientific method and natural selection. Unlike current generative models, which lack an effective evaluation mechanism during runtime, he notes that successful AI systems in fields like chess or molecular biology demonstrate the importance of evaluating options to fuel genuine discovery. He encourages the AI community to pursue methods that integrate these principles to fully unleash AI's potential in scientific discovery, calling for a collaborative effort to automate creativity and discovery processes in AI.