Erdős Problems Explained: What AI Has Actually Solved

Quick answer: Erdős problems are mathematical questions posed or popularized by Paul Erdős (1913–1996), mostly in number theory, combinatorics and geometry. They include solved and unsolved questions, and some carry cash prizes. Thomas Bloom’s website erdosproblems.com collects more than 1,200 of them. Since late 2025, AI systems have contributed proofs, counterexamples and formal checks, but each claim must be checked for correctness, novelty and the exact question it answers.

Here is a taste. The fraction \(\tfrac45\) can be split into three unit fractions: \(\tfrac45 = \tfrac12 + \tfrac14 + \tfrac1{20}\). The Erdős–Straus conjecture asks whether \(\tfrac4n = \tfrac1a + \tfrac1b + \tfrac1c\) has a solution in positive integers for every \(n \ge 2\). One example shows what the question means; a proof must cover every \(n\). That combination, easy to state and brutal to prove, is the signature of an Erdős problem.

Who Was Paul Erdős?

Paul Erdős was one of the most prolific mathematicians who ever lived. For much of his life he had no permanent home, traveling from colleague to colleague with a suitcase, and he wrote around 1,500 papers with more than 500 coauthors. He posed problems constantly and often attached cash prizes, from 25 dollars to several thousand dollars, roughly in proportion to how hard he thought they were.

Famous Examples

ProblemThe questionStatus (checked November 21, 2026)
Unit distance problem (1946)How many pairs among \(n\) points in the plane can be exactly 1 apart?Erdős’s conjectured growth rate disproved by an AI construction (May 2026); exact growth still unknown
Erdős–Straus conjectureDoes \(\tfrac4n = \tfrac1a + \tfrac1b + \tfrac1c\) have a solution in positive integers for every \(n \ge 2\)?Open (checked by computer for very large \(n\))
Reciprocal sums and progressions (Problem 3, a 5,000-dollar prize)If the reciprocals of a set of positive integers add up to infinity, must the set contain arbitrarily long arithmetic progressions?Claimed proof in an OpenAI manuscript (September 2026) with a Lean file; not yet peer reviewed or independently confirmed
Erdős discrepancy problemFor any sequence of \(+1\)s and \(-1\)s, can the sums \(x_d + x_{2d} + \dots + x_{md}\) be kept bounded in absolute value?No; solved by Terence Tao (posted 2015, published 2016)

What Is erdosproblems.com?

Thomas Bloom launched erdosproblems.com in May 2023 to collect questions associated with Erdős, with references and discussion. In August 2026, Quanta Magazine reported that it listed 565 solved and 652 open problems. On October 6, 2026, after a flood of unexplained AI-generated proof claims, Bloom announced that the site would stop showing open/solved labels and totals, pause new comments and proof claims, and focus on human exposition of the mathematics. So today the best way to learn a problem’s state is to read its references and explanation, not a scoreboard.

How AI Has Contributed

DateExampleWhat happened
Late 2025Problem 333An AI result was first presented as new, but a 1977 paper by Erdős and Donald Newman had already settled it: a lesson in checking the literature.
January 2026Problem 728Kevin Barreto used GPT-5.2 Pro and Harmonic’s Aristotle to produce a Lean proof of the intended version, after discussion clarified the original wording.
May 2026Google DeepMind studyResearchers reported resolving 9 of 353 formalized problem statements attempted, including subproblems and variants, with proofs checked in Lean.
May 20, 2026Unit distance conjectureAn internal OpenAI model disproved it; external mathematicians wrote a verified account.
October 6, 2026OpenAI’s large releaseHundreds of manuscripts and Lean files, including claims touching 23 Erdős problems. Bloom said he had not checked them and that no problem should yet be treated as closed.

How to Read a “Solved by AI” Headline

Type of contributionWhat it means
New proof or counterexampleA previously unresolved statement is now settled.
RediscoveryThe argument may be right, but earlier work already answered the question.
Partial result or variantProgress on a special case or a related statement, not the original problem.
FormalizationThe proof is written in a language such as Lean that a computer can check. That confirms the logic of the formal statement, not that the result is new or that the statement matches the original problem.
Newly posted claimCheck the evidence and review status before treating it as established.

Terence Tao and others have stressed this point: many AI successes are on easier or already-solved problems, while a few, like the unit distance construction, are real breakthroughs. Both can be true at once.

Why Are Erdős Problems a Test Case for AI?

They are numerous, mostly short to state, and range from exercises to problems that defeated Erdős himself. Some statements still need interpretation, some problems are related to one another, and the literature is not always easy to search, which is exactly why careful human checking remains essential.

The Erdős Number

Erdős collaborated so widely that mathematicians measure their “Erdős number”: Erdős has number 0, his coauthors have 1, their coauthors have 2, and so on. It is a fun measure of how connected the mathematical community is.

Watch: Erdős Problems and AI

If you learn better by watching, these videos are worth your time.

Terence Tao On AI Solving Erdős Problems (Dwarkesh Clips). Tao on what AI is and is not doing well with these problems. Recorded before the October 2026 releases.
The truth behind AI solving the Erdos problems (Mo Bitar). A skeptical look at early claims of AI solutions.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. Which of these is an example of the Erdős–Straus conjecture working?
Found a bug? Let us know!
2. If an AI writes a correct proof of a problem that was already solved in 1977, what is it?
Found a bug? Let us know!
3. Does a Lean proof show that a result is new?
Found a bug? Let us know!
4. Does solving a special case settle the original problem?
Found a bug? Let us know!
5. What changed on erdosproblems.com on October 6, 2026?
Found a bug? Let us know!

Frequently Asked Questions

What are the Erdős problems?

They are mathematical questions posed or popularized by Paul Erdős, mostly in number theory, combinatorics and geometry. Many are simple to state and hard to solve, and some carry cash prizes. The online collection erdosproblems.com contains more than 1,200.

How many Erdős problems are still open?

There is no single current count. In August 2026, erdosproblems.com listed 565 solved and 652 open, but in October 2026 the site stopped displaying status labels and totals. Community databases use their own categories.

Has AI solved Erdős problems?

AI has contributed real results, including the May 2026 disproof of the unit distance conjecture and a Lean proof for Problem 728 in January 2026. Other claims turned out to be rediscoveries or variants, and many October 2026 claims are not yet independently checked.

Does a Lean proof mean the result is new?

No. A Lean proof shows that a formal statement follows logically. Whether that statement matches the original problem, and whether the result was already known, must be checked separately.

Who pays the Erdős prizes?

The website does not award money. Its FAQ directs eligible claims to the Combinatorics Foundation, with requirements about publication and documentation.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Erdős Problems Explained: What AI Has Actually Solved - Effortless Math"?

No one replied yet.

Leave a Reply