OpenAI's Dan Roberts: Why AI Can Now Make Discoveries

Autonomous scientific discovery is becoming a reality as AI models transition from simple task execution to deep reasoning. Reinforcement learning now serves as a foundational paradigm, enabling models to navigate long, multi-step calculation paths and challenge established mathematical conjectures, such as the Erdos problems. By leveraging test-time compute, these systems generate internal thought processes that allow for iterative exploration and backtracking, mirroring human scientific inquiry. Dan Roberts, lead of the Foundations of Reinforcement Learning team at OpenAI, applies his background in theoretical physics to treat AI as a complex system, emphasizing that scaling laws rely on understanding the underlying mechanics of intelligence rather than just increasing parameter counts. This shift toward models that can autonomously reason and verify their own work marks a significant departure from traditional supervised learning, paving the way for breakthroughs in complex domains like mathematics and science.