Erdős Problems Freezes Comments: The AI Proof Flood Pushed Out the Humans
The site that became the public scoreboard for AI in mathematics just took the scoreboard down. Thomas Bloom, who runs erdosproblems.com, posted on Monday that the site is freezing new problem comments and proof claims, removing every problem's open/solved status and the solved-percentage counter, and dropping credit language for future solutions, whether human or AI. HN titled it "Erdősproblems.com succumbs to the AI onslaught".
The numbers explain why. The site holds 1,221 problems, over 9,000 comments, almost 2,000 registered users and 10,000 to 25,000 unique visitors a day. After comments launched in August 2025 a real community formed. Then AI went from useless on these problems to solving hard ones in under a year, and the main way people now interact with the site is to advertise AI-generated proofs, often with no attempt to explain them, as a way to record what Bloom calls an increasingly meaningless priority claim. Some mathematicians have stopped working on Erdős problems because they believe they cannot compete.
Bloom's framing is careful. The post says repositories of AI-generated proofs should exist, so others do not waste tokens regenerating the same proof, and that there are now several candidates. Bloom just does not want to run one. The line: one does not open a restaurant in an abattoir. The site will keep recording results and will promote high-quality expositions, including of AI proofs, as long as the writeup is human and AI use is honestly disclosed. Plagiarism or misrepresented AI use gets declined even if the proof is correct.
This is the fourth case in four days of the same cost shift. Simon Willison's hard caps, Google's paused bug bounty, Wikimedia's cleanup bill, and now a one-person math site. Agents made producing a submission nearly free. Checking it, contextualizing it and living with the community effects stayed expensive, and the person holding that cost closed the door.
For the agent-math crowd the practical change is small: email the proof, wait for a manual update. The cultural change is larger. The best-known open benchmark for AI mathematics just said that a formal proof is not where the value of a question ends.
Link: erdosproblems.com/forum/thread/blog:9
← Back to all articles
The numbers explain why. The site holds 1,221 problems, over 9,000 comments, almost 2,000 registered users and 10,000 to 25,000 unique visitors a day. After comments launched in August 2025 a real community formed. Then AI went from useless on these problems to solving hard ones in under a year, and the main way people now interact with the site is to advertise AI-generated proofs, often with no attempt to explain them, as a way to record what Bloom calls an increasingly meaningless priority claim. Some mathematicians have stopped working on Erdős problems because they believe they cannot compete.
Bloom's framing is careful. The post says repositories of AI-generated proofs should exist, so others do not waste tokens regenerating the same proof, and that there are now several candidates. Bloom just does not want to run one. The line: one does not open a restaurant in an abattoir. The site will keep recording results and will promote high-quality expositions, including of AI proofs, as long as the writeup is human and AI use is honestly disclosed. Plagiarism or misrepresented AI use gets declined even if the proof is correct.
This is the fourth case in four days of the same cost shift. Simon Willison's hard caps, Google's paused bug bounty, Wikimedia's cleanup bill, and now a one-person math site. Agents made producing a submission nearly free. Checking it, contextualizing it and living with the community effects stayed expensive, and the person holding that cost closed the door.
For the agent-math crowd the practical change is small: email the proof, wait for a manual update. The cultural change is larger. The best-known open benchmark for AI mathematics just said that a formal proof is not where the value of a question ends.
Link: erdosproblems.com/forum/thread/blog:9
Comments