Erdős Problems Freezes Proof Submissions and Drops Its Solved Count After AI Flood
A.I. / news
Erdős Problems Freezes Proof Submissions and Drops Its Solved Count After AI Flood
Site creator Thomas Bloom says AI-written proofs had displaced discussion on the 1,221-problem database, so credit and open-or-solved labels are gone as of Oct. 6.
Thomas Bloom, who runs erdosproblems.com, froze new comments and proof submissions on the site on Oct. 6 and removed the labels that marked each problem open or solved. He wrote that AI-generated proofs without explanation had displaced the mathematical discussion the site was built for.
The changes are listed in Bloom's blog post, dated Oct. 6. Bloom is at the University of Manchester, according to Quanta Magazine, and started the site in early 2023.
What changed on the site
Four changes took effect, per the post. General discussion stays open, and proofs can now be sent by email instead.
| Change | Before | After |
|---|---|---|
| New comments and proof claims | Open to registered users | Frozen |
| Problem status | Open or solved labels, solved percentage | Neutral colours, no counter |
| Credit | Results tied to named people or AI systems | Ownership language removed |
| Focus | Claims | Written exposition of proofs |
The post gives the site's size as 1,221 problems, more than 9,000 comments and nearly 2,000 registered users, with 10,000 to 25,000 visitors a day. Bloom wrote that the site should help the Erdős community of humans flourish, and described a "gamified glory-seeking attitude" around AI results.
Users reacted in both directions on the thread. Some welcomed the focus on explanation, and others said hiding proof status works against documenting progress.
How the AI results got there
The labs have been solving problems from the list for most of 2026. Quanta, in an Aug. 3 piece by contributing writer Konstantin Kakaes, dated the milestones.
| Date | Event, per Quanta |
|---|---|
| Jan. 4, 2026 | Kevin Barreto and Liam Price solve Erdős Problem 728 |
| Jan. 2026 | Google DeepMind solves 4 problems and finds forgotten solutions to 9 more |
| May 20, 2026 | OpenAI solves the unit distance problem |
| Aug. 1, 2026 | OpenAI's Astra makes 10 advances, 3 of them Erdős problems |
Quanta quoted Bloom on the volume. "We're seeing a lot more of these 100- to 200-page papers," he said, adding that "no human has read it, and no human is going to read it."
What OpenAI claimed, and what it left out
Simon Willison summarised OpenAI's Aug. 1 post. An internal version of Astra, which OpenAI calls its next major model, worked on ten problems with no progress on the main result for at least a decade. OpenAI said it spent under $2,000 per solved problem at GPT-5.6 Sol token prices.
Willison wrote that OpenAI did not say how many problems consumed $2,000 without a solution. The results are formalised in Lean 4 in the openai/ten-proofs repository, which a reader can check without trusting the prose. Willison also wanted the prompts, which were not in the post.
The same week
On Oct. 6, OpenAI's post titled "Sharing AI progress in mathematics" reached 942 points on Hacker News. The Terminal received a 403 error when it tried to read that page, so nothing in this article relies on it. Willison's post quotes Terence Tao, a UCLA professor, as expecting "large-scale, decentralized collaborations between humans and machines."
Related coverage on this site includes Vals AI's Opus 5.5 agents naming two magnetic semiconductor candidates and OpenAI's GPT-6.1 Sol pricing.
The post as read does not say whether the freeze is temporary. Until it does, email is the only route for a new proof.
Sources
More in A.I.
- 01Qwen3.8-27B Is Apache 2.0 With a 262,144-Token Window and Already Has 482 Finetunes, Including Cloudflare's ClefAlibaba's dense 27B model claims 61.7 on SWE-bench Pro against 53.4 for Opus 4.6 Max, and the model card says nothing about training data.
- 02A GitHub Repo Claims a Lean Proof That Walter Trump's 1979 Packing of 11 Squares Is Optimal, Checked Across 7,920 ModulesThe repository does not say what produced the proof, and no paper or peer review accompanies it, but the formal checker has accepted every module.
- 03Claude Haiku 5.5 Costs $0.10 Per Million Input Tokens, a Tenth of Haiku 4.5, and Scores 39.2% on Terminal-Bench 4.0Anthropic's smallest model is also its first Haiku with an effort setting, and its own table shows Sonnet 5.5 still ahead on every listed test.
- 04ChatGPT's Chat Tab Gets GPT-6 Sol and Interactive Answers Wednesday, With Luna for Free Users Due October 8OpenAI's Intelligent UI replaces some text replies with buttons, forms and charts. The speed and safety figures are OpenAI's own internal evaluations.