A New Era of Mathematical Discovery
In a groundbreaking development, an amateur mathematician, utilizing OpenAI's GPT-5.4 Pro, has cracked a 60-year-old combinatorics problem, a feat that has eluded specialists for decades. This achievement signifies a pivotal moment for artificial intelligence, demonstrating its ability to contribute to frontier discovery in mathematics in a truly collaborative fashion. The significance lies not just in the problem's resolution, but in the *method* of its resolution, showcasing AI as an active research partner capable of exploring novel angles and surfacing connections previously unconsidered by human experts.
The problem, a research-level conjecture about prime numbers posed by mathematicians Paul Erdős, Andras Sárközy, and Endre Szemerédi in 1966, had resisted conventional approaches. Historically, mathematicians working on this problem adopted a probabilistic approach, a method so ingrained it was rarely questioned. However, GPT-5.4 Pro diverged from this established path, employing a different mathematical language and utilizing the von Mangoldt function, a tool often considered unusual or arbitrary by human mathematicians.
AI's Unique Problem-Solving Approach
The AI's solution path was entirely original, a departure from all previous human attempts. While the initial output from ChatGPT was described as "quite poor" and required expert interpretation, it contained a novel and potentially crucial insight that human experts had missed. This suggests that AI can overcome human aesthetic biases that might lead researchers to overlook unconventional but effective solutions.
This breakthrough is not an isolated incident. Since October, AI tools have contributed to solving approximately 100 of Paul Erdős' mathematical problems, which he left behind as 1,179 unsolved conjectures. Large language models are proving to be powerful research assistants, capable of finding and combining existing mathematical results in innovative ways. In some instances, language models have even constructed entirely new and valid proofs with minimal human input.
Beyond Literature Search: AI as a True Collaborator
The capabilities of current AI models extend far beyond mere literature search. For example, Google's Gemini discovered a remark in a 1981 paper that unknowingly solved Erdős problem number 1089. Mathematicians like Andrew Sutherland at the Massachusetts Institute of Technology describe these models as invaluable research assistants, noting that those unfamiliar with newer versions may not fully grasp their advanced capabilities.
The collaborative dynamic between humans and AI is redefining mathematical discovery. The process now involves the AI proposing solutions, and human researchers rigorously testing and validating the outputs. This iterative approach transforms the human role from sole problem-solver to one focused on asking better questions, pruning unproductive directions, and verifying AI-generated insights.
The Future of AI in Mathematics
The impact of AI on mathematical research is rapidly expanding. In January, Ravi Vakil, president of the American Mathematical Society, co-authored a preprint documenting how Google's language model assisted in reaching a proof. Several mathematicians anticipate that 2026 will be the year when AI-contributed results begin to pass peer review in major mathematics journals.
Competitions like First Proof are actively challenging AI with unpublished proofs from top mathematicians, with results currently under review. Furthermore, a research team led by Peking University developed a dual-agent AI system that autonomously solved a 12-year-old conjecture in commutative algebra and formalized the proof with minimal human intervention. This system, called Rethlas, explores problem-solving strategies like a human mathematician and uses another system, Archon, to transform potential proofs for verification.