ChatGPT 5.5 Pro Produces PhD-Level Math Research in One Hour

Key Takeaways

• ChatGPT 5.5 Pro completed PhD-level mathematical research in roughly one hour with minimal human guidance
• LLMs can now solve open research problems that human mathematicians missed, not just find existing answers
• The bar for what counts as a 'good first research problem' for new mathematicians has been raised

Timothy Gowers, Fields Medal winner and one of the world's most prominent mathematicians, just reported something that should make anyone paying attention to AI sit up straight. ChatGPT 5.5 Pro produced what he describes as PhD-level mathematical research in about an hour. His contribution? Essentially none.

"We are all having to keep revising upwards our assessments of the mathematical capabilities of large language models," Gowers wrote on his blog. "I have just made a fairly large revision."

From Finding Answers to Finding New Arguments

The initial reaction to LLMs solving math problems was easy to dismiss. Early "solutions" often meant the model found an existing answer in the literature or made an obvious deduction from known results. Mathematicians could comfort themselves: these systems were search engines with better prose, not thinkers.

That comfort is evaporating. LLMs have now solved several of the open Erdős problems, a collection of challenges posed by the legendary mathematician Paul Erdős that have stumped researchers for decades. The problems are tracked on Thomas Bloom's website, and the AI solutions keep coming.

Gowers describes the current state: "LLMs have got to the point where if a problem has an easy argument that for one reason or another human mathematicians have missed, then there is a good chance that the LLMs will spot it." The reason humans missed these arguments varies. Sometimes the problem just hasn't received much attention. Sometimes the solution requires combining techniques from different areas in non-obvious ways.

Testing Against Genuinely Open Problems

Gowers decided to run an experiment. In combinatorics, research papers often introduce new parameters and pose several natural questions about them. Authors can't spend weeks on every question, so some remain open despite having approachable solutions. These problems have traditionally been perfect for PhD students and early-career researchers. Solving an officially open problem builds confidence and credentials.

He fed ChatGPT 5.5 Pro a selection of problems from a paper by Mel Nathanson, titled "Diversity, Equity and Inclusion for Problems in Additive Number Theory." The results were apparently striking enough that Gowers felt compelled to write about them immediately.

What This Means for Mathematical Training

The implications hit hardest for how mathematicians are trained. Finding a first research result is a crucial step. It proves to the student, their advisor, and future employers that they can do original work. If LLMs can now clear the bar that used to define "publishable first result," that bar needs to move.

“It is no longer enough that somebody asks a problem: it needs to be hard enough for an LLM not to be able to solve it.”

— Timothy Gowers

Gowers acknowledges a counter-argument that offers limited comfort: "Quite a lot of perfectly good human mathematics consists in putting together existing knowledge and proof techniques." If that's what LLMs are doing, they're doing exactly what mathematicians do. The distinction between "synthesis" and "originality" gets blurry.

The Larger Pattern

This fits a pattern we've seen across fields. AI doesn't need to match top experts to be disruptive. It needs to handle tasks that took humans significant time, or that served as proving grounds for newcomers. Radiology residents, junior lawyers doing document review, entry-level programmers, now first-year math PhD students. The jobs that trained the next generation are becoming optional.

What remains human? Gowers doesn't speculate, but his experiment suggests the remaining territory is smaller than many mathematicians assumed. Problems that require genuinely novel approaches, questions that haven't been asked before, and the taste to know which problems matter. Whether those are enough to sustain the traditional pipeline of mathematical talent is an open question.

Frequently Asked Questions

What is ChatGPT 5.5 Pro?

ChatGPT 5.5 Pro is a newer version of OpenAI's language model that Gowers received early access to test. It appears to have significantly improved mathematical reasoning compared to previous versions.

What are Erdős problems?

These are open mathematical problems posed by Paul Erdős, one of history's most prolific mathematicians. They cover various areas of mathematics and have challenged researchers for decades. LLMs have recently solved several of them.

Can LLMs do original mathematical research?

According to Gowers, LLMs can now find arguments that human mathematicians missed. Whether this counts as "original" is debatable, since humans also build on existing knowledge. The practical distinction is becoming less meaningful.

How will this affect math education?

Problems that were once appropriate for first-time researchers may now be solvable by LLMs. This raises the difficulty bar for what counts as meaningful early research, potentially changing how graduate students are trained.

Source: Hacker News: Best