In a quiet weekend experiment that stunned the mathematical community, software engineer and researcher Neel Somani witnessed a pivotal shift. He watched as OpenAI’s latest model, GPT-5.2, autonomously generated a complete and verifiable proof for a high-level mathematical problem over fifteen minutes. This event, occurring in late 2024, marks a significant milestone where artificial intelligence has begun to genuinely push the frontiers of human knowledge, particularly within the revered and challenging domain of pure mathematics.
AI Mathematics Reaches a New Frontier with GPT-5.2
Neel Somani’s initial goal was simple: to establish a baseline for large language model capabilities. He wanted to see where these systems still struggled with open mathematical problems. The result, however, was profoundly unexpected. After pasting a complex problem into ChatGPT and allowing its chain-of-thought reasoning to work for a quarter of an hour, Somani returned to find a full solution. He rigorously evaluated and formalized the proof using the verification tool Harmonic, and it checked out perfectly.
This was not mere pattern recognition. The model’s reasoning process invoked advanced mathematical concepts like Legendre’s formula and Bertrand’s postulate. It even located and synthesized information from a 2013 Math Overflow post by Harvard mathematician Noam Elkies. Crucially, GPT-5.2’s final proof differed from and expanded upon Elkies’ work, providing a more complete solution to a version of a problem posed by the legendary Paul Erdős. This demonstrates a move beyond data retrieval into genuine, adaptive problem-solving.
The Erdős Problem Proving Ground for AI
The problems of Paul Erdős, the prolific Hungarian mathematician, have long served as benchmarks for human intellect. His collection of over one thousand conjectures, maintained online, varies wildly in subject and difficulty. For years, they have represented the pinnacle of abstract reasoning. Now, they have become the primary testing ground for AI-driven mathematical discovery.
The pace of progress has accelerated dramatically since the release of GPT 5.2, which Somani and others describe as “anecdotally more skilled at mathematical reasoning.” Since late December 2024, the Erdős problem website has seen a notable shift:
- •15 problems have moved from “open” to “solved.”
- •11 of these solutions specifically credit AI models as instrumental in the discovery process.
- •The first autonomous solutions appeared in November from a Gemini-powered model called AlphaEvolve, but GPT-5.2 has recently shown remarkable adeptness.
Fields Medalist Terence Tao maintains a nuanced tracking of this progress on his GitHub. He notes eight distinct Erdős problems where AI models made meaningful autonomous progress, with six additional cases where AI accelerated progress by locating and building upon obscure prior research. This data underscores a collaborative, if not yet fully independent, role for AI in advanced research.
The Scalable Advantage of AI in the ‘Long Tail’
On Mastodon, Terence Tao offered a key insight into why AI is particularly effective here. He conjectured that the scalable nature of AI systems makes them “better suited for being systematically applied to the ‘long tail’ of obscure Erdős problems.” Many of these problems, while unsolved, may have relatively straightforward solutions that simply haven’t attracted sustained human attention. “As such,” Tao continued, “many of these easier Erdős problems are now more likely to be solved by purely AI-based methods than by human or hybrid means.” This represents a fundamental shift in the economics of mathematical inquiry.
The Critical Role of Formal Verification Tools
A parallel revolution enabling this progress is the rise of formalization. Formal verification involves expressing mathematical proofs in a precise, logical language that a computer can check for absolute correctness. This labor-intensive process removes ambiguity and error.
Tools like the open-source proof assistant Lean, developed at Microsoft Research, have become industry standards. Now, AI is automating formalization itself. Harmonic’s Aristotle, the tool Somani used, promises to handle much of the tedious work of translating human or AI-generated reasoning into a verifiable format. This creates a powerful synergy: AI proposes creative solutions, and automated tools rigorously verify them.
For Harmonic founder Tudor Achim, the solved problems are less important than the changing perception among experts. “I care more about the fact that math and computer science professors are using [AI tools],” Achim said. “These people have reputations to protect, so when they’re saying they use Aristotle or they use ChatGPT, that’s real evidence.” This adoption signals a transition from novelty to trusted research instrument.
Implications for the Future of Mathematical Research
The implications of this trend are profound. AI is not replacing mathematicians but augmenting them in specific, powerful ways. It acts as a tireless research assistant capable of scanning centuries of literature, generating novel conjectures, and exploring thousands of algorithmic pathways in moments.
| AI’s Role | Human’s Role |
|---|---|
| Systematic exploration of the “long tail” of obscure problems | Posing deep, fundamental questions and setting research direction |
| Automating formal verification and proof checking | Providing intuitive understanding, creativity, and conceptual breakthroughs |
| Rapid literature review and synthesis of existing knowledge | Judging significance, weaving results into broader theories, and providing context |
The field is moving toward a hybrid model. In this model, human intuition guides AI exploration, and AI productivity amplifies human insight. This partnership could dramatically accelerate progress across number theory, combinatorics, and other fields rich in well-defined but unsolved problems.
Conclusion
The breakthrough in AI mathematics, exemplified by GPT-5.2 solving Erdős problems, is a watershed moment. It moves artificial intelligence from a tool for computation and pattern recognition into a potential partner in fundamental discovery. While fully autonomous, creative AI research remains distant, the current capabilities for augmentation, formalization, and systematic problem-solving are already reshaping mathematical practice. The evidence from researchers like Neel Somani and the adoption by leading mathematicians confirm that AI has earned a permanent seat at the table of high-level inquiry. The frontier of knowledge is now being pushed forward by a new, synthetic form of intelligence.
FAQs
Q1: What exactly did GPT-5.2 solve?
GPT-5.2 generated a novel and verifiable proof for a version of a problem from the collection of conjectures by legendary mathematician Paul Erdős. It did not simply recall an answer but constructed a logical proof using advanced mathematical axioms.
Q2: Does this mean AI can now do original mathematical research?
It indicates a significant step toward autonomous research, but within constraints. AI currently excels at solving well-defined, existing problems, especially obscure ones. Formulating entirely new fields or paradigms of mathematics remains a uniquely human strength.
Q3: What is “formal verification” and why is it important?
Formal verification uses logical computer languages to write proofs that machines can check for absolute correctness. Tools like Lean and Harmonic’s Aristotle are crucial because they eliminate human error in verifying complex, AI-generated proofs, making the results trustworthy.
Q4: Are mathematicians being replaced by AI?
No. The consensus, echoed by experts like Terence Tao, is that AI is becoming a powerful augmentative tool. It handles systematic, scalable tasks like exploring minor conjectures or checking proofs, freeing human mathematicians to focus on deep creativity, intuition, and high-level direction.
Q5: What fields of mathematics is AI impacting most?
Currently, areas with many discrete, well-posed conjectures are most affected. This includes number theory, combinatorics, and graph theory. The structure of problems in these fields aligns well with the logical, stepwise reasoning of current large language models and formal verification systems.