Training Sand to Think: Artificial General Intelligence & Future of Physics

Large Language Models (LLMs) are fundamentally revolutionizing the conduct of theoretical physics and mathematics by evolving from simple predictive tools into autonomous research agents. Driven by scaling laws—where increased compute, data, and algorithmic ingenuity yield predictable performance gains—these models have rapidly progressed from basic logic to achieving gold-medal scores on the International Math Olympiad. Techniques like chain-of-thought prompting and iterative self-correction allow these systems to solve complex graduate-level problems and generate novel mathematical proofs, such as the unit-distance conjecture. This trajectory mirrors the evolution of chess computers, moving from specialized tools to superhuman entities. As these models become more capable and cost-effective, they are poised to usher in a golden era of scientific discovery, functioning as tireless, scalable collaborators that augment human expertise and accelerate the resolution of long-standing theoretical challenges.