OAKWEB.AI
Home/Understanding AI/Solved by AI
12 — SOLVED BY AI

Solved by
machine.

For most of history, the hardest open problems in math and science fell to people. That's changing fast. This page tracks every problem that humans couldn't crack and AI did — what the problem is, what the AI actually did, and an honest status for each, because in this field the headline and the proof are often two different things. Updated October 31, 2026.

THE SCOREBOARD

24 problems, five kinds of “solved”

Tap a status to show only those entries. Tap it again to show everything.

Showing all 24.

  • CONFIRMED Independently verified or peer-reviewed.
  • NEW RECORD Beat the best known answer. The problem itself is still open.
  • MACHINE-CHECKED The proof passes a computer proof-checker (Lean). Experts are still reviewing it.
  • CLAIMED Announced, but not yet checked by anyone outside.
  • CONTESTED Real work, but what it actually proves is disputed.

Two things to know before you read. First, “checked in Lean” means a computer verified the proof follows from its stated assumptions — but a human still has to confirm the statement is the real problem. Some formal proofs have turned out to prove something weaker than the headline. Second, a lot of “AI solved X” headlines really mean “AI beat the best known answer to X.” Both are real progress; they aren't the same thing, and this page says which is which.

THE 2026 WAVE

The year AI started doing frontier math

Most of this list happened in the last ten months. Newest first.

26 Oct 2026MACHINE-CHECKED

The Exponential Time Hypothesis, refuted

Computer science · OpenAI internal model · Anthropic's Claude, hours later

The problem: A 27-year-old assumption that some logic puzzles (like 3-SAT) can't be solved much faster than trying every possibility. Huge parts of computer science were built on it being true.

What the AI did: An unnamed OpenAI model produced an algorithm that beats that limit, checked in Lean the next day. Hours later Anthropic posted its own algorithm for the same problem, credited to Claude with “no human input.”

Why it matters: If it holds, one of the field's load-bearing beliefs is false. It does not settle P vs NP, isn't fast in practice, and doesn't break encryption.

The catch: The hypothesis's own authors are reviewing it and say their involvement “should not be taken as certification.” It breaks a weaker version of ETH; stricter versions stand.

Source: TechCrunch ❯
26 Oct 2026MACHINE-CHECKED

Local smoothing — and Kakeya — in every dimension

Harmonic analysis · OpenAI internal model

The problem: A 1991 conjecture about how waves spread out and overlap. Humans had only managed the 2-D case (2019); the related Kakeya problem was proved in 3-D in 2025.

What the AI did: OpenAI's model claimed the result in all dimensions — which would also settle the Kakeya conjecture — and released its output with a Lean proof.

Why it matters: If correct, it closes a whole research area at once.

The catch: No expert verdict has been published yet.

Source: OpenAI math repo ❯
26 Oct 2026CLAIMED

Yang–Mills theories exist in 4-D

Mathematical physics · OpenAI models

The problem: Half of one of the seven $1M Millennium Prize problems: rigorously building the quantum theory behind the strong nuclear force. No 4-D theory of this kind had ever been constructed.

What the AI did: OpenAI says its models proved such theories exist, including QCD. A group led by Nobel laureate David Gross will write it up.

Why it matters: It would put a pillar of particle physics on solid mathematical ground for the first time.

The catch: The AI's output hasn't been released, and it's not the Millennium Prize — that also needs the “mass gap,” which isn't claimed.

Source: The Next Web ❯
8 Oct 2026CONFIRMED

Nine Erdős problems, peer-reviewed

Combinatorics · number theory · Google DeepMind, AlphaProof Nexus

The problem: Paul Erdős left hundreds of open problems; some, like #12, had been open since 1970. Separately, hundreds of conjectures sit unproved in the OEIS integer-sequence database.

What the AI did: DeepMind's language-model-plus-Lean system solved 9 of 353 open Erdős problems and 44 of 492 OEIS conjectures, every proof machine-checked — at a few hundred dollars per problem. Published in Science.

Why it matters: The first batch of AI-solved open problems to pass journal peer review.

The catch: The hit rate is about 2.5%, and some solved problems are variants of the originals.

Source: arXiv paper ❯
6 Oct 2026MACHINE-CHECKED

The quasi-Riemann hypothesis

Number theory · OpenAI internal model

The problem: Since the 1800s, nobody could prove the zeros of the zeta function — the key to how prime numbers are spaced — stay inside any strip narrower than the obvious one.

What the AI did: OpenAI's model produced a 199-page proof, verified in Lean, that they stay between 1/8 and 7/8, and a shorter one for 1/12–11/12.

Why it matters: Not the Riemann Hypothesis itself, but one expert says “the vast majority of consequences of RH already follow.” Another calls it “the most significant AI-for-math achievement so far, by quite a margin.”

The catch: Experts “largely agree that the result is correct,” per Quanta, but it isn't peer-reviewed yet.

Source: Quanta Magazine ❯
6 Oct 2026CLAIMED

722 papers in one day

All of mathematics · OpenAI

The problem: Not one problem — roughly 4,000 posed to a model at once, about 3 hours of compute each.

What the AI did: OpenAI released 722 manuscripts with claims on famous problems, including Hilbert's tenth problem over the rationals and the Unique Games Conjecture. About 42% of the headline results are Lean-checked.

Why it matters: If even a fraction hold up, it's more progress in a day than some fields see in decades. One mathematician said “they have annihilated my field of research.”

The catch: Mostly unreviewed. Three papers were withdrawn within a day over a sign error, and one expert's read was “a really nice trick rather than a fundamental breakthrough” for most.

Source: Wikipedia ❯
7 Sep 2026MACHINE-CHECKED

Blow-up for 3-D Euler and its cousins

Fluid dynamics · Buckmaster & Alpöge, with Claude and Codex

The problem: Whether the equations for ideal fluids (Euler) and two related systems can blow up in finite time with a smooth push.

What the AI did: Two mathematicians had AI write most of the work; the first blow-up proof was Lean-verified within a week. OpenAI conceded they got to Euler first.

Why it matters: Human-led, AI-heavy — the template for how a lot of math may get done now.

The catch: Not peer-reviewed. Buckmaster called the first AI-written proof “the most horrendous I have ever read.”

Source: GitHub ❯
1 Aug 2026MACHINE-CHECKED

Ten decade-old problems

Mathematics · computer science · OpenAI, “Astra”

The problem: OpenAI picked problems with no progress in at least ten years.

What the AI did: Results include the first explicit non-sofic group (open since 1999), a disproof of Connes's rigidity conjecture (1980), three Erdős problems, and the first improvement since 1978 on the upper limit for sphere-packing density.

Why it matters: The first lab batch of decade-old problems, each with a Lean certificate.

The catch: Not peer-reviewed, and OpenAI hasn't said how many problems it tried and failed.

Source: OpenAI (PDF) ❯
20 May 2026CONFIRMED

Erdős's 80-year unit-distance guess, disproved

Geometry · OpenAI internal model

The problem: In 1946 Erdős guessed how many pairs of points can sit exactly one unit apart. Nobody could prove or break it for 80 years.

What the AI did: OpenAI's model built point sets that beat the guess. Nine mathematicians, including Fields medalist Tim Gowers, verified it by hand.

Why it matters: Gowers said that from a human he'd have accepted it in the Annals “without any hesitation.”

The catch: The guess is dead, but the true answer is still unknown.

Source: arXiv paper ❯
10 Mar 2026NEW RECORD

Nine Ramsey numbers pushed up

Combinatorics · Google DeepMind, AlphaEvolve

The problem: Ramsey numbers measure when order is guaranteed to appear in chaos. Their exact values are famously hard; even the bounds move slowly.

What the AI did: AlphaEvolve raised nine lower bounds — for example R(3,13) from 60 to 61 and R(4,16) from 170 to 174 — each backed by a graph a computer can check.

Why it matters: One general-purpose AI beat many specialized human searches at once.

The catch: Better bounds, not exact answers. Preprint.

Source: arXiv paper ❯
13 Feb 2026CLAIMED

“Impossible” particle collisions aren't zero

Theoretical physics · GPT-5.2 with Strominger, Skinner et al.

The problem: Textbook arguments said a certain kind of gluon-collision amplitude always vanishes.

What the AI did: GPT-5.2 Pro spotted the general formula from human hand calculations, then a scaffolded GPT-5.2 derived and proved it in about 12 hours.

Why it matters: AI corrected a textbook assumption in particle physics.

The catch: Only in special conditions far from real experiments. Preprint, checked by its authors.

Source: arXiv paper ❯
Jan 2026CONFIRMED

Erdős #728 — the first clean AI solve

Number theory · GPT-5.2 Pro + Harmonic's Aristotle

The problem: One of Erdős's open problems, with no known solution anywhere in the literature.

What the AI did: GPT-5.2 Pro produced the argument; Harmonic's Aristotle turned it into a Lean proof. A write-up followed.

Why it matters: Widely accepted as the first Erdős problem solved autonomously by AI.

The catch: The problem site calls the original wording “ambiguous.”

Source: arXiv paper ❯
2020–2025

Before the flood

The results that showed it was possible — protein shapes, matrix tricks, a buried scroll.

3 Nov 2025NEW RECORD

67 problems, with Terence Tao

Combinatorics · analysis · geometry · Google DeepMind, AlphaEvolve

The problem: A grab-bag of open optimization problems across mathematics.

What the AI did: Run with Fields medalist Terence Tao and colleagues, AlphaEvolve matched the best known answers on most of 67 problems and beat them on several.

Why it matters: Top mathematicians endorsing AI search as a working research tool.

The catch: Improved constructions, not full solutions. Preprint.

Source: arXiv paper ❯
Oct 2025CONFIRMED

A new way to make tumors visible

Cancer biology · Google & Yale, C2S-Scale

The problem: Many tumors hide from the immune system, and finding drugs that unmask them is slow.

What the AI did: An AI model predicted that the drug silmitasertib makes tumor cells more visible to the immune system — but only when an immune signal is low. Lab experiments confirmed it.

Why it matters: An AI-generated biology hypothesis that held up in the lab.

The catch: Confirmed in human cancer cell lines only — no patients yet.

Source: Google ❯
14 May 2025NEW RECORD

Beating a 1969 matrix record

Algorithms · geometry · Google DeepMind, AlphaEvolve

The problem: Since Strassen in 1969, the fastest way to multiply two 4×4 matrices of complex numbers took 49 multiplications. And the best known count for spheres touching one sphere in 11 dimensions was 592.

What the AI did: AlphaEvolve found a 48-multiplication method and an 11-D arrangement of 593 spheres. On 50+ open problems it matched the state of the art about 75% of the time and beat it about 20%.

Why it matters: One system improved many unrelated records at once — including one that stood for 56 years.

The catch: Records, not solutions: the 11-D kissing number is still unknown.

Source: Google DeepMind ❯
5 Feb 2024CONFIRMED

Reading a scroll buried by Vesuvius

Archaeology · Vesuvius Challenge winners

The problem: Scrolls carbonized in AD 79 crumble if you try to open them. For 275 years they were unreadable.

What the AI did: Machine-learning ink detection on CT scans revealed 15 columns and over 2,000 characters of text from an unopened scroll, winning a $700,000 prize.

Why it matters: The first readable text recovered from inside an unopened 2,000-year-old scroll.

The catch: About 5% of the first scroll so far, and AI was one tool alongside human-built virtual unwrapping.

Source: Vesuvius Challenge ❯
14 Dec 2023NEW RECORD

Bigger cap sets

Combinatorics · Google DeepMind, FunSearch

The problem: The cap set problem asks how many points you can pick without any three in a line. In dimension 8 the record was 496.

What the AI did: FunSearch, a language model paired with an automatic checker, found a set of 512 — and better bin-packing heuristics too.

Why it matters: The first time a language model produced a verifiable new result on an established open math problem.

The catch: A new record, not a solution: the cap set problem remains open.

Source: Nature ❯
29 Nov 2023CONTESTED

2.2 million predicted crystals

Materials science · Google DeepMind, GNoME

The problem: Only about 48,000 stable crystals were known — the raw material for better batteries and chips.

What the AI did: GNoME predicted 2.2 million new structures, about 380,000 of them the most stable. Outside labs had made 736 of them.

Why it matters: If the predictions hold, it's roughly a tenfold jump in known stable materials.

The catch: Chemists found “scant evidence” of compounds that are new, credible and useful at once. Call it a prediction catalogue, not 2.2 million discoveries.

Source: The critique (Chem. Mater.) ❯
5 Oct 2022NEW RECORD

Faster matrix multiplication

Algorithms · Google DeepMind, AlphaTensor

The problem: Matrix multiplication sits under almost all computing and AI. For 4×4 matrices, Strassen's 1969 method (49 multiplications) had stood for 50 years.

What the AI did: AlphaTensor found a 47-multiplication method for 4×4 matrices in mod-2 arithmetic, plus other records.

Why it matters: The first improvement to that case in half a century.

The catch: Only in mod-2 arithmetic, and humans improved one of its records again within a week.

Source: Nature ❯
1 Dec 2021CONFIRMED

New theorems in knot theory

Pure mathematics · Google DeepMind with Oxford & Sydney

The problem: A link between a knot's algebra and its geometry nobody had spotted, and a representation-theory conjecture stuck for nearly 40 years.

What the AI did: Machine learning pointed mathematicians at patterns; they proved a new theorem about knots and made progress on the conjecture.

Why it matters: AI as a hunch generator for top mathematicians.

The catch: Humans did the proving, and the 40-year conjecture isn't fully proved.

Source: Nature ❯
29 Apr 2021CONFIRMED

Counterexamples on demand

Graph theory · Adam Zsolt Wagner, reinforcement learning

The problem: A set of published conjectures in graph theory, believed true.

What the AI did: A simple learning method found counterexamples — one a 19-node graph found in a few hours on an ordinary PC.

Why it matters: Showed AI can hunt down the cases that break mathematicians' guesses.

The catch: Minor conjectures, and a preprint — but the counterexamples are easy to check.

Source: arXiv paper ❯
Nov 2020CONFIRMED

Predicting protein shapes

Biology · Google DeepMind, AlphaFold 2

The problem: Predicting a protein's 3-D shape from its sequence had been an open problem for over 50 years. Lab methods took months or years per protein.

What the AI did: AlphaFold 2 hit accuracy competitive with experiments at the CASP14 contest. Its database grew to over 200 million structures, and it won half the 2024 Nobel Prize in Chemistry.

Why it matters: Over 2 million researchers in 190 countries have used it. Biology and drug discovery sped up overnight.

The catch: It solved shape prediction — not how proteins actually fold or move.

Source: Nature ❯
WHERE IT STARTED

The first one

10 Oct 1996CONFIRMED

The Robbins conjecture

Algebra · logic · EQP automated prover, Argonne National Lab

The problem: Do three simple rules describe all of Boolean algebra — the logic inside every computer? Mathematicians, including Alfred Tarski, tried and failed for about 60 years.

What the AI did: William McCune's program EQP found the proof by itself in about 8 days on a single workstation.

Why it matters: The first famous open problem settled by a machine finding the proof itself.

The catch: This was symbolic AI — rule-based reasoning — not machine learning.

Source: McCune's Robbins page ❯

Precursors that don't count as AI. Some famous problems fell to computers that searched exhaustively under human direction, with no learning involved: the Four Color Theorem (1976), the Kepler conjecture on stacking spheres, the Boolean Pythagorean triples problem (2016, a 200-terabyte proof) and Keller's conjecture in dimension 7 (2020). They belong in the story, but humans designed every step the machine took.

AI AS AUDITOR

Checking the biggest proof ever written

Not an open problem, but too important to leave out. On 26 October 2026, Axiom's AI agents finished a complete computer-checked proof of the Classification of Finite Simple Groups — a theorem whose human proof is spread over about 15,000 pages and had never been fully checked. In about 24 days they produced 43.1 million lines of Lean and found 34 notable slips in the published record, including a step in a 1986 paper that counterexamples broke. The theorem survived; the AI found another route. Read Axiom's write-up ❯

THE FINE PRINT

Claims that didn't hold up

For every real result there's been a headline that ran ahead of the math. These are the ones worth remembering.

  • Oct 2025 “GPT-5 solved 10 Erdős problems.” It had found existing solutions in the literature, not new proofs. The post was deleted; Demis Hassabis called it “embarrassing.” Source ❯
  • Late 2025 Erdős #124 solved by Aristotle. Only an easier variant — one expert called it “very easy.” The original is still open. Source ❯
  • Dec 2025 Erdős #333 solved autonomously. The answer was already in Erdős's own 1977 paper. Source ❯
  • Jan 2026 Erdős #281, “the most unambiguous instance” yet. The proof was new, but the problem had been solved in 1936 and 1966. Source ❯
  • Oct 2026 The rational Hodge conjecture for K3 surfaces. One of three OpenAI papers withdrawn within a day over a sign error. Source ❯
  • Sep 2026 “AI solved a Millennium Prize problem.” Only the version with an artificial outside force. The Clay Institute still lists Navier–Stokes as unsolved. Source ❯
  • Oct 2026 “AI broke cryptography / proved P = NP.” Neither follows from the ETH result, even if it holds up. Source ❯
WHAT IT ADDS UP TO

Three takeaways

The pace changed in 2026. For 25 years, AI's wins on open problems came one at a time, years apart. This year they arrived in batches — nine Erdős problems in one paper, 722 manuscripts in one day. Checking is now the bottleneck. Machines can produce proofs faster than experts can read them, and several of the biggest claims here are waiting on a handful of people to finish reviewing. “Solved” needs a footnote. The honest version of almost every entry has a catch — a weaker version, a record rather than an answer, an artificial force. Read the catch.

Want the background? Start with what a model actually is, or see where these results sit in the AI timeline.