How a reasoning model cracked an 80-year-old math problem - Episode 20

The podcast discusses an extraordinary achievement where a general-purpose AI model was used to disprove an 80-year-old conjecture related to the unit distance problem by mathematician Paul Erdős. The team from OpenAI, including Alexander Wei, Hongxun Wu, and Lijie Chen, explain how they used their model to generate a proof that overturned previous assumptions about the arrangement of points on a plane. This breakthrough has sparked excitement within both the AI and mathematical communities, highlighting the potential of AI to make high-level scientific discoveries.

The conversation details how AI models have evolved from mere question-answering tools to entities capable of complex reasoning and problem solving through increased test-time computation and strategic prompting. The models, initially tested on challenging problems, have demonstrated an ability to produce creative and unconventional solutions, such as disproving long-standing mathematical conjectures. This progress underscores the importance of general models that are not narrowly trained but are capable of broad reasoning across different domains.

The researchers describe their process of verifying the AI-generated proof, involving checking the solution with mathematical colleagues and expert reviewers. They express both astonishment and confidence in the model's ability, noting that the proof was checked and initially met with skepticism, which was alleviated after careful review. This points to a new era where AI can contribute seriously to academic research, albeit with rigorous validation.

The episode emphasizes the potential for AI to accelerate scientific discovery across disciplines beyond mathematics, including physics, astronomy, and cryptography. The speakers advocate for empowering researchers with AI tools, promoting collaborative exploration rather than competition, and broadening access to reasoning models to foster global scientific progress. They see AI as an amplifier of human ingenuity, capable of generating novel ideas and solving problems previously deemed intractable.

The researchers offer practical advice for other scientists and mathematicians interested in working with AI. They recommend starting with accessible tools like GPT Pro, asking bold questions, and trusting the models while understanding their limitations. They also discuss the importance of learning how to effectively collaborate with AI systems, viewing them as partners in discovery rather than replacements, and anticipate continued exponential improvements in AI's reasoning abilities.