NewsAI & Development

OpenAI AI Disproves ETH: What Developers Need to Know

Complexity class hierarchy P NP PSPACE with ETH boundary fracture effect representing OpenAI AI mathematical discoveries

On October 26, 2026, OpenAI announced that its internal AI models produced formal proofs for three of the hardest open problems in mathematics, theoretical computer science, and physics — all within six weeks of each other. Consequently, the claim getting the most attention from developers is this: OpenAI’s model disproved the Exponential Time Hypothesis (ETH), a four-decades-old assumption that hundreds of algorithm complexity papers depend on. Whether this holds is a different question entirely.

What the ETH Disproof Actually Claims

The Exponential Time Hypothesis, originally proposed by Russell Impagliazzo and Ramamohan Paturi, states that 3-SAT cannot be solved in subexponential time. Researchers have built conditional lower bounds on top of it for 25 years — proving that problems in graph theory, database query planning, and computational biology are hard, but only if ETH holds. OpenAI’s model found a deterministic algorithm for counting satisfying assignments of k-SAT formulas with complexity 2^O(n log log n / √(log n)). That’s still superpolynomial, but it’s subexponential — which would make ETH false.

The model produced the natural-language disproof on October 16 and formalized it in Lean — a machine-verifiable proof assistant — the next day. Moreover, Impagliazzo, Paturi, and Ryan Williams (MIT), who originally proposed ETH, are now leading an independent examination. Their statement deserves to be read carefully: “Our involvement should not be taken as certification of the claim. A result of this significance requires careful, independent scrutiny.” That’s the original proponents of a hypothesis telling you not to assume their hypothesis has been disproved yet.

None of This Has Been Verified

None of the three OpenAI claims is peer-reviewed. The math and CS results include Lean formalizations published on GitHub. However, the physics artifacts are still withheld pending expert review. A Lean proof confirms that logical steps are internally consistent — it does not confirm that the human-readable statement of the conjecture was correctly translated into Lean. That correspondence must be verified by human experts, which hasn’t happened yet.

Furthermore, context that most coverage has buried: on October 7, just three weeks before this announcement, OpenAI withdrew three earlier AI-written manuscripts due to errors. Only 42% of results in that batch were fully formalized. As a result, OpenAI’s track record here is imperfect, and the current announcement follows a pattern of ambitious claims with verification trailing behind. The “shepherd” model — where named academics organize review groups rather than submit to traditional journals — is new and untested at this scale.

Related: OpenAI’s Agents Disabled Safety Monitors — One Lied

The Other Two Claims: Kakeya and Yang-Mills

OpenAI’s models also claimed to prove the local smoothing conjecture — proposed by Christopher Sogge in 1991 — in every dimension, using a novel auxiliary partial differential equation. Proving this simultaneously settles the Kakeya conjecture and Stein’s restriction conjecture across all dimensions. Humans had previously proven local smoothing in two dimensions (2019) and the Kakeya conjecture in three dimensions (2025). Jumping to all dimensions at once is the kind of result that makes harmonic analysts stop and pay attention.

Meanwhile, the Yang-Mills result claims a rigorous existence proof for quantum field theories including QCD and pure Yang-Mills theory — which touches one of the Clay Millennium Prize problems. However, OpenAI is careful to note this covers only existence, not the mass gap requirement needed for the full prize. “Part of” a Millennium Prize is not the same as solving it. Nobel laureate David Gross leads the Yang-Mills verification team alongside Arthur Jaffe, Igor Klebanov, and Michael Douglas — but the physics artifacts remain withheld from public view.

What Changes for Developers

In the short term: nothing. The ETH algorithm is purely theoretical — no implementation benchmarks, no practical speedups, no updated SAT solver. The constant factors in the complexity bound are unknown and potentially enormous. Your Z3 and CaDiCaL installations are unchanged. Cryptographic protocols that rely on NP-hardness assumptions are not threatened, because disproving ETH does not imply P=NP. Therefore, working developers can file this under “watch this space” rather than “update your dependencies.”

In the long term, the picture is different. If any of these results survive peer review, it marks the first time an AI system has produced original proofs for open conjectures that stumped human researchers for decades — not just solved known problems, but found new mathematics. Additionally, OpenAI committed $5 million to a Simons Foundation initiative to fund community workshops and conferences aimed at understanding AI-generated results. That funding signals they expect this process to take time.

Key Takeaways

  • OpenAI claims its AI models disproved the Exponential Time Hypothesis and proved the Kakeya conjecture and part of the Yang-Mills Millennium Prize problem — all announced October 26, 2026
  • None of the three results has been peer-reviewed; Lean formalizations are published for CS and math, but human experts must verify the correspondence to the original conjectures
  • ETH disproof would undermine hundreds of conditional hardness results in CS — but does not prove P=NP and has no immediate practical impact on SAT solvers or cryptography
  • OpenAI retracted three earlier AI manuscripts on October 7 with a 42% formalization rate — context worth weighing alongside the current announcements
  • Watch the shepherd process: the real signal will come when Impagliazzo, Williams, and David Gross’s teams publish their assessments
ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News