OpenAI model overturns Erdős conjecture: AI mathematical reasoning breakthrough, 125 pages CoT rewrites the ceiling of AI capabilities

OpenAI's AI model successfully overturned the Erdős conjecture in discrete geometry, using the Lean proof auxiliary language to find counterexamples, and the CoT reasoning chain is 125 pages long. This is the first time that AI has independently completed serious mathematical discoveries, which means that the reasoning ability of AI Agent has transcended the tool stage and entered the era of independent discovery. For automation practitioners, this is not only news, but also a signal to re-evaluate the upper limit of AI capabilities and upgrade workflows.

Core conclusion

OpenAI announced on May 20, 2026 that its AI model successfully overturned a core conjecture in discrete geometry-the Erdős conjecture. This is the first time in the history of AI that serious mathematical research has been independently completed and publishable results have been achieved, marking the advancement of AI reasoning capabilities from "tool-assisted" to "autonomous discovery". For AI money-making and automation practitioners, this means that the reasoning ceiling of AI Agents has been significantly raised—it can handle more complex multi-step reasoning tasks, and automated systems can perform higher-quality code reviews, architecture design, and logic verification.

Key Points

Release time: 2026-05-20
Affected objects: AI Agent developers, automation practitioners, AI programming tool users
Core changes: AI uses Lean proof auxiliary language to complete counterexample construction, 125 pages of CoT reasoning chain - a scale far beyond the typical working memory of human mathematicians

Background and trigger events

On May 20, 2026, OpenAI published a blog post announcing that its AI model successfully found a counterexample to the Erdős conjecture, thus overturning this core conjecture that has troubled the discrete geometry community for many years. The Erdős conjecture (proposed by Paul Erdős, the greatest mathematician of the 20th century) involves the distance problem of point sets in discrete geometry. It is a classic problem in basic mathematics that "everyone understands but no one can prove".

The core weapon used by the model is Lean - an interactive theorem prover. By formalizing geometric problems as theorems in Lean, the model was able to systematically search for counterexample constructions, eventually finding one constructed "with unexpected tools of algebraic number theory."

SEO: The first paragraph contains the main keywords "OpenAI model reasoning", "Erdős conjecture" and "AI mathematical discovery" GEO: beginning with TL;DR, exact number (125 pages CoT, 564 HN points) bonus points

Key Impact (by Dimension)

Dimensions	Change	What it means to us	Recommended actions
AI reasoning capabilities	From "imitation" to "autonomous discovery" of new mathematics	AI Agent can handle more complex multi-step reasoning tasks	Re-evaluate the upper limit of AI capabilities of existing workflows
Code quality	AI can use formal verification to ensure correctness	The reliability of AI programming has been greatly improved	Pay attention to Lean/formal verification integration opportunities
Trust	Society’s understanding of AI capabilities has been refreshed	Customers are more willing to accept AI-driven solutions	Cite this case in marketing materials
Cost-effectiveness	Automation of difficult intellectual work becomes possible	Opening up new money-making tracks: AI-assisted research, formal verification services	Evaluating whether to provide AI inference outsourcing services

Adaptation suggestions

Implications for AI automation practitioners

Re-evaluate the upper limit of AI capabilities: If your automated workflow is still at the "writing marketing copy, translating text" stage, you should rethink - AI can already independently solve mathematical problems that require a PhD level. Is it time to upgrade your automation system?
Focus on Lean and Formal Verification: Lean is an open source theorem prover that can check for logical errors just like a compiler checks for grammatical errors. Introducing formal verification into the AI code generation pipeline can significantly reduce "hidden bugs" (such as permission cross-border, data leakage and other structural issues) in the production environment.
Shift from "prompt word optimization" to "structural constraints": The beginning of the article quotes an important point - "Structured reverse pressure can improve code quality better than smarter Agents." Rather than spending time optimizing prompt words to get the AI to write more correct code, it’s better to build structured validation walls.

Task List

Add formal verification steps to AI programming workflow
Evaluate the application of theorem proving tools such as Lean in content automation
Updated content strategy: AI reasoning breakthroughs as trust-building material

Community reaction

This report on Hacker News received 564 points and 373 comments, making it the AI-related topic with the highest HN score that day. Core views of the community:

"This proof brings unexpected and exquisite algebraic number theory tools to solve a basic geometric problem. The value of AI is that it has prior knowledge in all fields and can be transferred across fields at zero cost." — HN Community Member cpard

"This shouldn't be a surprise to anyone who does a lot of programming with LLM. It's just a matter of time. Mathematicians make discoveries by building and applying mathematical tools in new ways. LLM can do a Monte Carlo search of each mathematical tool, see which one works, and then keep building." — HN Community Member lubujackson

"The Summary of Chained Thoughts (CoT) for this task is 125 pages. This is a staggering scale of reasoning, similar to the Mythos Anthropic has been discussing." — HN community member zozbot234

125 AI —

Related extended information

Tool entry (trigger tool floating card)

Tools that appear naturally in this article: OpenAI, ChatGPT, Claude, DeepSeek, Hermes Agent, LangGraph

Internal link guidance

Want to build an AI Agent automation system? Watch: AI Agent-Driven Content Automation: n8n MCP Building Guide from Scratch
Add quality gates to automated workflows: How to add quality gates to your AI automation workflow: A practical guide from output to trustworthy results
Real case: Someone used AI automation to earn over 10,000 yuan a month He Built an AI Automation Stack with Claude + n8n — $4K to $12K/mo in 6 Months