OpenAI claims it solved an 80-year-old math problem — for real this time

OpenAI has recently made headlines for a breakthrough that moves beyond the typical capabilities of generative chatbots: the formal resolution of a mathematical problem that has remained unsolved for eight decades. The company’s latest reasoning model, o1, successfully disproved a geometry conjecture first proposed in 1946 by Paul Erdős. This development is particularly notable not just for the mathematical discovery itself, but because it marks a moment of redemption for OpenAI. Earlier this year, the company faced significant criticism from the academic community when it prematurely claimed to have solved a different conjecture, only to be corrected by vigilant mathematicians. This time, however, those same experts have validated the result, signaling a new era of reliability for AI in the hard sciences.

The problem in question belongs to the field of discrete geometry and involves the maximum number of unit distances possible among a set of points in a specific dimension. For decades, mathematicians have chipped away at various permutations of this problem, but the specific conjecture disproved by OpenAI had remained a persistent outlier. By identifying a counterexample that violates the theorized limits, o1 demonstrated a level of "reasoning" that transcends mere pattern matching. Unlike previous versions of GPT, which functioned primarily as next-token predictors, the o1 model utilizes reinforcement learning and a "chain-of-thought" process that allows it to iterate on complex logic, recognize errors in its own pathing, and pursue multifaceted proofs over extended periods.

Technically, this achievement highlights a shift in how large language models (LLMs) are being trained for specialized tasks. While traditional LLMs excel at creative writing or coding, they have historically struggled with the rigid, non-negotiable logic required for formal mathematics. OpenAI’s approach with o1 involves a system of internal rewards for correct logical steps, essentially teaching the model to "think" before it speaks. This architectural change allows the AI to navigate search spaces that are far too vast for human intuition to map efficiently. By automating the search for counterexamples in geometric configurations, the AI acts as an ultra-fast collaborator that can test hypotheses at a scale previously inconceivable.

The implications for the broader tech industry and the scientific community are profound. We are moving away from the era of "hallucinating" AI toward systems capable of verifying their own outputs against objective truths. This transition is essential if AI is to become a tool for genuine scientific discovery rather than just a productivity aid. For competitors like Google DeepMind—which has its own storied history with AlphaGeometry and G-flow networks—OpenAI’s success serves as a formidable challenge. It suggests that the gap between generalized reasoning and specialized scientific problem-solving is narrowing, potentially leading to a market where the most valuable models are those that can contribute original intellectual property to the fields of physics, chemistry, and cryptography.

As these tools integrate more deeply into academic workflows, they will likely redefine the role of the mathematician. Rather than spending years manually testing edge cases, researchers will pivot toward framing problems and verifying the sophisticated proofs generated by algorithmic partners. However, this also raises questions regarding the "black box" nature of AI. While the counterexample provided by o1 is verifiable, the process by which it arrived at the solution remains difficult to parse. If AI begins to solve problems that humans cannot explain, the nature of scientific "understanding" itself may need to be reevaluated.

In the coming months, the industry will be watching to see if this success can be replicated across other disciplines. The "o1" series is currently a prototype of what a truly analytical AI might look like, but its true test will lie in whether it can tackle "millennium prize" problems or contribute to breakthroughs in material science. For now, OpenAI has regained the trust of a skeptical scientific community, proving that while its models can still be prone to flair and fiction, they are increasingly capable of finding the cold, hard truths hidden within the numbers. If this trend continues, the next great scientific revolution may not come from a laboratory, but from a server rack.