Sharing AI progress in mathematics
TL;DR Highlight
OpenAI posted AI-made math proofs on GitHub; some were later retracted
Who Should Read
A developer using LLM APIs who wants to gauge the real strengths and limits of reasoning models
Core Mechanics
- No article text was provided
- The points below come from comments, not the original post
- Facts from the original body cannot be confirmed
- Repo published on GitHub, per comments
- Comments say openai/math holds preprints and reasoning_traces
- Number of solved problems has no source
- Figures and model names are not in the original, so they are omitted
Community Reactions
- Scale and compute claims
- One comment quoted the repo saying the average result used about three hours of ChatGPT Pro thinking compute
- A comment said proofatlas.ai lists 90 of the top 500 open problems as fully solved
- Unique Games Conjecture claim
- Several comments called it a big deal since many inapproximability results rest on it
- Retractions reported
- A comment said three Hodge-related manuscripts were withdrawn over a sign error
- It suggested more verification is needed
- Readability experience
- A comment said the proofs read like human writing
- It contrasted this with ChatGPT text that is overly defensive
- Self-built tool example
- A commenter said their autonomous math research agent solved a proof complexity problem and led to a paper
How to Apply
- Skim the reasoning traces
- Today, open the reasoning_traces folder in openai/math and read one initial prompt
- Note its structure and compare it with your own prompts
- Add a verification step to AI outputs
- Have a human or a formal checker recheck model-made proofs or analyses in your pipeline
- Hidden sign errors like the retracted cases can slip through
Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.