ChatGPT 5.5 Pro produces publishable LaTeX papers: a milestone in AI writing capabilities

Core conclusion

On May 8, 2026, Fields Medal winner Timothy Gowers published a blog article that shocked the academic community: ChatGPT 5.5 Pro completed the proof construction of a number theory problem in only 17 minutes of reasoning + 2 minutes of writing, and automatically output it as an academic preprint in LaTeX format.

Key conclusion: This is the first time that AI has evolved from a "problem-solving tool" to an "academic writing collaborator", and its significance to the automated production of content is far greater than the mathematics community itself.

Background: Gowers’ test design

Gowers picked an open problem posed by mathematician Mel Nathanson in his paper - proving a quadratic upper bound on the diameter of a sumset. The characteristics of this problem are:

• It has a known linear algebra construction (Nathanson's original proof), but is not optimal
• ChatGPT needs to discover more efficient Sidon sets by itself to improve the upper bound
• Proving ideas requires real combinatorial insights, not literature searches

Gowers added in the HN discussion that he deliberately chose a "less popular but publicly documented" question to ensure that ChatGPT was not replicating existing results.

"I think this test is fair: ChatGPT really needs to construct the proof itself, rather than extract it from the literature." - Timothy Gowers

Key process restoration

Step 1: Reasoning (17 minutes)

ChatGPT 5.5 Pro spent 17 minutes thinking about the question structure. Its core insight is that Nathanson uses powers of 2 as the basis for Sidon set construction, but this results in a linear increase in diameter. If you switch to the more efficient quadratic Sidon set, you can push the diameter from linear to quadratic optimal.

This is a classic "change a tool" moment in combinatorics. Gowers commented: "Nathanson did not do this, probably because he was trying to generalize his inductive construction, while ChatGPT chose the optimal path directly from scratch."

Step 2: Writing (2 minutes 23 seconds)

The initial output style of ChatGPT is "verbose and has a typical LLM accent". Gowers directly requested: "Please write this in the academic preprint style in LaTeX format." After 2 minutes and 23 seconds, ChatGPT output a fully structured LaTeX document, including introduction, lemma, proof, conclusion and reference frame.

Step 3: Verification (a few hours)

Gowers spent hours manually verifying the correctness of the proof. Final confirmation: The proof logic is complete, and ChatGPT also improved another related conclusion of Nathanson (restricted sumset problem) using the same method.

The final preprint is here: A note on sumsets and restricted sumsets

Implications for automated content production

This experiment has three key implications for the field of AI content production:

This means: The output format ceiling of AI is being broken. In the past, I could only write blogs, but now I can write essays. For content creators, this means:

• The cost of producing high-quality structured content (research reports, white papers, technical documentation) will drop dramatically
• The "verifiability" of AI output is improving - Gowers can verify mathematical proofs, which means it is not empty AI nonsense
• LaTeX/Markdown structured output will become standard for AI content production

Disagreement in academia

Two schools of thought emerged in the HN discussion:

Optimists (including famous mathematicians such as John Baez) believe:

• If the problem itself is difficult enough, AI being able to think for 17 minutes is a breakthrough in itself
• "If the value of ideas comes from their utility rather than scarcity, then AI creating more good ideas is a good thing."
• Feedback from a physics professor: Gemini can find missing imaginary units in the code. This kind of fine-grained inspection is extremely rare in traditional peer reviews.

The Worrying Party believes:

• The pool of “entry-level questions” for training PhD students has been drained by AI
• The minimum threshold for “being able to do research” is being raised by AI
• In the future, it will be increasingly difficult to judge the value of a work without stripping it of the contribution of AI

Tool entries (natural matching)

The core AI tools involved in this article: OpenAI, ChatGPT, Claude, Gemini. The API pricing and inference capabilities of these tools are rapidly iterating and are critical to automating content production pipelines.

Related extended information

• Gowers
• LaTeX PDF
• HN （639 upvotes）

Internal link guidance

• Want to use AI to output structured content? Watch: Claude Code automated writing practice: build an AI content production pipeline in 30 minutes
• How to ensure the reliability of AI writing? Watch: He earns over 10,000 per month by relying on AI code review + specification-driven development: a practical review of a freelance developer
• Want to learn about the latest tools for AI Agent content production? Watch: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes