OpenAI Solves 25 Year Math Problem in Days

Francis Johnson at University College London spent 25 years working on a specific problem in mathematics. He wrote two books on the subject. He eventually gave up. Now, an internal frontier model at OpenAI solved it in a matter of days. This single anecdote captures the strange tension in the current moment.

OpenAI has just released 722 mathematical papers. They prove and disprove a wide range of problems. This is not a minor update. It is a massive shift in scale. The company did not name the specific model. It only described it as an internal frontier model. The academic world is reacting with a mix of shock and skepticism.

The sheer volume of these publications marks a turning point for the field. For decades, mathematical discovery was a slow, incremental process driven by individual genius or small collaborative teams. Now, the barrier to entry for complex proof generation has collapsed. The speed at which these results were produced dwarfs traditional academic timelines, forcing a reevaluation of what constitutes a viable research strategy in the modern era.

From depth to breadth

For months, the focus has been on depth. In 2019, AI models struggled with basic math. Last month, OpenAI solved a problem related to the Navier-Stokes equations. That was a breakthrough in difficulty. This new release is a breakthrough in volume. The company has opted for breadth, releasing hundreds of discoveries at once.

Johnson says he is not surprised that AI solved the problem. He is surprised by how quickly it happened. He worked with recent models and was astonished by the clarity of the analysis. He describes the situation as the genie being out of the bottle. He believes humans must adapt to this new reality.

The shift from tackling single, hard problems to generating vast libraries of results changes the dynamic entirely. It suggests that the model is not just solving isolated puzzles but navigating the entire landscape of known mathematics with ease. This breadth allows for cross-pollination between different subfields, potentially uncovering connections that human experts, often siloed in their specializations, might miss.

The verification bottleneck

Kevin Buzzard at Imperial College London offers a more cautious view. He looked at the 30 papers in his field of number theory. Only seven seemed impressive to him. Only one was formally verified in Lean. Lean is a type of computer analysis that can prove results beyond reasonable doubt. This is a critical distinction.

Buzzard warns that acceptance by the community will take time. The six unformalized results must wait. An expert must read and check the text. Or a Lean formalization must be produced. Journalists will have to wait while mathematicians do their job. This is not a rejection of the technology. It is a defense of scientific rigor.

The reliance on Lean formalization highlights a crucial gap in the current workflow. While natural language proofs can be persuasive, they remain open to subtle logical errors that are hard to spot. The formal verification process acts as a final gatekeeper, ensuring that the logical steps are airtight. This requirement creates a significant backlog, as the volume of AI-generated proofs far exceeds the capacity of the few experts skilled in formal verification.

A strange new position

Johnson says we are in a very strange position. He notes that human beings are supposed to be adaptable. He suggests we stand back and be astonished. This is a candid admission of a new era. The pace of discovery is outpacing the pace of verification.

Buzzard acknowledges the upward trend in AI mathematics. He says it is astonishing. He assumes the new results are mostly true. He expects a significant shift in how mathematics is practiced. The tools are changing. The methods are changing. The role of the human mathematician is being redefined.

This redefinition challenges the traditional prestige associated with original discovery. If a machine can generate the proof, the human role shifts towards curation, verification, and identifying which results are meaningful. It is a move from being the primary creator of knowledge to being the primary editor and judge of that knowledge. This transition is uncomfortable but necessary for the survival of rigorous mathematical practice.

What comes next

This release is a test case. It is not just about the 722 papers. It is about the process. How will the community verify them? How will the models be improved? How will the workflow change? These are the real questions.

The genie is out of the bottle. The tools are powerful. The results are impressive. But the work is just beginning. The next few months will be critical. We will see how the community handles this influx of new knowledge. We will see how the models evolve. We will see how the field adapts.

The coming months will likely see the development of new automated verification tools to handle the sheer volume of new proofs. Universities may need to adjust their curricula to include formal methods and AI literacy as core components of mathematical training. The ultimate goal is not to replace human mathematicians but to create a new hybrid discipline where human intuition and machine precision work in tandem to push the boundaries of what is known.