On 6 October 2026 erdosproblems.com — Thomas Bloom's collection of the problems Paul Erdős posed — took down its status labels and its solved counter, and closed its comment threads. His reasons are good ones: a wave of unexplained AI-generated proofs was being posted as priority claims, and he wants the problems to read as questions rather than a scoreboard. He also said separate repositories have a place, so long as they're kept separate.
This is one. It shows, for every problem, the status recorded by the community database maintained beside his site — dated, linked back, and explained in plain words. Statements come from Google DeepMind's Formal Conjectures Lean files; his own wording stays on his site, as he asked. Nothing here is a claim of priority, and nothing here is posted back to his site on anyone's behalf.
It's also a place to bring an attempt. Send one and it gets checked the way we check our own work, then — with your consent — published under an honest verdict on the problem's page. The guide explains how to work on one with your own AI subscription.
We work on these problems too. Patrick White, who built the site, isn't a mathematician; Claude, an AI made by Anthropic, does the checking and, since October 2026, the attempts themselves. Every attempt we've made is on the attempt log, the misses included, and our verdicts are always labelled as ours — they never change what the board shows.
We pull candidates from T. F. Bloom's erdosproblems.com — full credit, we're not building a competing list. We hunt for problems that are genuinely tractable but neglected, run our own computation on them before writing anything (small-N searches, known-witness checks, sanity checks), then send a sharp brief to a second, independent reasoning seat. Through September 2026 that seat was GPT-5.6 Sol via ChatGPT Pro; since October 2026 it is Claude (Fable 5.1) at full reasoning effort, with computation as the primary verifier and blind refutation passes from other model families. Whatever comes back gets independently checked — against primary sources, by hand, or by fresh computation — before we call it anything.
These are the verdicts on our own attempts. Everywhere on the site they're drawn as outlined stamps; the database's statuses are drawn as filled chips.
Each row also wears its own badge, finer than the boxes: FULL WRITE-UP — a complete entry with its own page · PROVED / FOUND — a decisive verified check, often that the printed statement is false (we certify; we don't claim priority) · SKIPPED — the no-collision rule fired before we started · NO PROGRESS — probed, nothing to show as ours.
Plainly AI-assisted. Claude does the verification work here, not ghostwriting — and since October 2026 the attempts themselves. I relay, read, and make the calls on what gets published. Every entry says exactly what happened, which model did what, and what was checked independently.