What happens now that AI is good at math? - Episode 17

The speakers highlight the astonishing progress of AI in understanding and solving complex mathematical problems over recent years. They recount how, just two years ago, reasoning models capable of proving difficult theorems were non-existent, but now AI models can assist with solving research-level mathematics. Notably, ChatGPT and other models have transitioned from handling simple math problems to engaging with advanced research, even helping to solve longstanding open problems. The evolution is driven not by a single factor, but by ongoing cutting-edge research, better training techniques, and the models’ ability to use tools like calculators, as well as improvements in reasoning.

Both experts share their experiences and specific instances where AI has solved significant mathematical problems. Ernest Ryu details how he used ChatGPT to solve a 42-year-old open problem in optimization theory, illustrating AI’s potential as a creative research partner. They discuss the progress from AI's early limitations in basic math to its now capability to do literature searches and generate solutions to open Erdős problems, with some solutions being completely novel and peer-reviewed. These advances demonstrate that AI is now capable of original mathematical research, accelerating scientific discovery across disciplines.

The conversation emphasizes that AI’s capacity to handle sophisticated math will have broad impacts across all sciences, not just mathematics. AI-driven tools will enable faster experimentation, deeper interconnectedness among fields, and more efficient validation of results. Despite these advances, the speakers caution against over-reliance on AI, warning that shallow understanding and loss of expertise could be detrimental. They argue that human intuition, critical thinking, and deeper engagement remain essential. The future of science involves collaboration where AI amplifies human intelligence rather than replaces it, with AI acting as a powerful assistant to accelerate discovery and understanding.

Discussing future developments, the speakers explore the concept of an 'automated researcher'—AI systems capable of autonomous, long-term scientific investigation. Current models support iterative, short-term interactions, akin to professor-student dialogues, but the goal is to extend this to AI that can think over weeks or months, enabling true breakthroughs. They highlight the importance of increasing the 'AGI time'—the duration an AI system can simulate human-like thinking to solve complex problems—moving from seconds to weeks or months. Such systems would dramatically shorten research timelines, assist with verifying proofs, and make high-level research accessible and manageable for scientists and students alike.

The experts envisage a future where AI makes mathematics and science more fun, interconnected, and accessible. AI will help uncover connections across fields, accelerate the verification process for complex proofs, and motivate a new generation of scientists. They note that AI will not diminish the value of deep expertise but rather augment it, allowing researchers to focus on meaningful inquiry. They also warn of potential risks, such as over-trust in AI outputs and the need to maintain rigorous scientific standards. Ultimately, AI is seen as a partner in human progress—enabling faster, more profound discoveries that could revolutionize everything from medicine to materials science.