Bloom says the Erdos problems site will freeze proof claims

On 6 October 2026, Terence Tao’s blog published Changes to the Erdős problems web site, a guest post by Thomas Bloom, who created erdosproblems.com. Tao’s note says the post is crossposted from the Erdős problems forum, that it was initially written in a different file format, and that it was converted using AI. Bloom asks readers not to copy the site’s text without permission. This article paraphrases the post and uses two short quotations.

A request to the site’s home page on 6 October returned HTTP 403 from a Cloudflare challenge, so this article cannot say the changes are already visible there. What follows is what the post says Bloom is doing.

What he says will change

Bloom writes that he will make the following changes, and that, “as ever, these are somewhat experimental, and may themselves change and be clarified further in the next few weeks.”

He will freeze new problem comments and proof claims. General threads and the blog stay open. Until comments return, he asks people to email suggestions, and he will update the site himself when he thinks that is appropriate. The site will stop showing statuses such as open or solved, and problems will appear in one neutral colour. “The count of currently solved problems and the solved percentage will no longer be shown.” It will still record results, and it will stop using credit-giving language for a future result, whether a person or an AI produced it.

He will favor high-quality expositions of old or new proofs, human or AI, and he mentions an online seminar. He will try to link quickly to correct Lean formalisations verified on a project he names, Palomar. Poorly explained proofs with no formalisation are ones he is unlikely to add.

Why he gives those reasons

Bloom says comments, added in August 2025, first brought the discussion he wanted, and that the activity did not last. Some of the drop, he writes, was going to happen anyway. He says AI sped it up, because AI can solve some of the problems with little human intervention. The main public interaction now, in his account, is people advertising AI-generated proofs, often with no explanation, as a priority claim. He does not want to run a proof repository. He says a dump of formal proofs, including ones no person understands, can be useful if someone else manages it, and he uses a restaurant-and-slaughterhouse image for why that dump should stay off a site meant to explain the questions. He says a formal proof of the truth value is not the end of a question’s value.

He also writes that a quieter audience learned mathematics from the site and never commented, and that he does not know how long the hiatus will last. He will bring comments and proof claims back, perhaps in a different form, when he thinks they will do more good than harm.

His counts, as he tells them: launched 28 May 2023 with just over 200 problems; now 1,221 problems, over 9,000 comments, almost 2,000 registered users, and typically 10,000 to 25,000 unique visitors a day.

Why the solved count was being watched

Bloom writes that the site has often been treated as a front line of these changes and as a barometer of AI capabilities, and that this was not his aim. He wanted a human audience and a reference for work already done. He also says a lot of recent progress has come from people returning to the problems, not only from AI solutions, and that some mathematicians have stepped back because they think they cannot compete with AI. That barometer is his description. If he does what the post describes, a solved count and a solved percentage will no longer be on the site to cite, and he calls the changes experimental.