A 23-Year-Old Amateur Used ChatGPT to Solve a 60-Year-Old Erdős Conjecture, Validated by Terence Tao
A 23-year-old amateur mathematician has used OpenAI's GPT-5.4 Pro to crack a 60-year-old open problem in combinatorial number theory — and Fields Medal winner Terence Tao has publicly validated it as a "nice achievement". The story, broken by Scientific American earlier today, is the cleanest evidence yet that frontier AI is no longer just helping mathematicians format proofs; it is now contributing genuine new mathematical ideas at the research level.
What the Erdős Conjecture Actually Was
The problem in question is one of Paul Erdős's many open conjectures from the 1960s in additive combinatorics — a corner of number theory that asks how integer sets can be built from sums and differences. The specific statement involves bounds on the maximum size of sum-free subsets within structured sequences. It had resisted attack for six decades, including from professional researchers with full institutional support.
The amateur, who has not formally trained as a mathematician, fed the problem to GPT-5.4 Pro along with carefully chosen partial results from the existing literature. Over multiple sessions, the model proposed a novel approach combining two techniques that had never been combined for this problem before. The user then verified each step manually and worked with the model to close the gaps.
Why Terence Tao's Validation Matters
Tao is one of the most respected living mathematicians. His public assessment — calling the result "a nice achievement" and noting the proof technique was novel — is not casual praise. Tao has been openly skeptical of LLM-driven mathematics for years, repeatedly demonstrating cases where models produced confident-sounding nonsense. For him to flag this case as legitimate is the bar most research mathematicians will use to decide whether to take the result seriously.
Crucially, the proof has been independently checked. This is not a "the AI says it works" claim. The actual mathematical argument has been read line by line by humans and verified to be correct. The novelty is in how the argument was constructed, not whether it stands.
The Bigger Pattern: AI as a Research Collaborator
This is part of a steady stream of frontier AI scientific results landing in 2026. Earlier this month we covered Google's 25 percent share of global AI compute — the raw resource that powers experiments like this — and Anthropic's Project Glasswing coalition on AI for security research. The pattern is consistent: when domain experts pair up with frontier models, the productivity ceiling shifts noticeably.
What makes the Erdős result distinct is the user. A 23-year-old amateur with no PhD did this. The implication is that the bottleneck for frontier mathematics may no longer be only formal training — it may now also include access to the right model and the patience to push it. That is a different kind of democratization than software has produced before.
My Take
This is genuinely impressive and a useful counterweight to the "AI is just autocomplete" narrative. A frontier model proposing a novel proof technique that survives Tao-grade scrutiny is not autocomplete. It is reasoning, applied to a problem that resisted humans for sixty years.
That said, do not over-extrapolate. One Erdős problem solved is one data point. Mathematicians have been pairing with AI tools for years, and most attempts still produce subtle errors that take experts weeks to diagnose. The honest read is: the ceiling is rising, but the floor is still uneven. The smart play is not "AI replaces mathematicians" — it is "the next generation of mathematicians will work with AI as a peer collaborator, and the ones who do not will fall behind".
Frequently Asked Questions
What is the Erdős conjecture that ChatGPT solved?
The result concerns a 60-year-old open problem in additive combinatorics related to bounds on sum-free subsets within structured integer sequences. It is one of many conjectures Paul Erdős posed during his lifetime that resisted resolution until now.
Who is the person who solved it?
A 23-year-old amateur mathematician, without formal mathematics training. The full identity has been published in the Scientific American piece breaking the story. He used GPT-5.4 Pro across multiple sessions, then verified the proof manually with the model's help.
Has Terence Tao confirmed the proof is valid?
Yes. Tao publicly described the result as "a nice achievement" and acknowledged that the proof technique was novel. The argument has also been independently checked line by line by other mathematicians.
Does this mean AI can do mathematics research now?
Not unconditionally. AI can now genuinely contribute to research-level mathematics when paired with a careful human verifier, but unsupervised AI-generated proofs still produce subtle errors. The realistic frame is "AI as a peer collaborator", not "AI replaces mathematicians".
The Bottom Line
An open problem dating to the 1960s just fell to a 23-year-old amateur and a frontier language model. Terence Tao validated it. The proof technique is novel. This is not the end of mathematics, but it is a clear data point that the frontier of what AI plus a thoughtful human can do has moved meaningfully forward in 2026. Expect more results like this — and expect the bar for what counts as a "real" AI math achievement to keep rising.