FULL STORY

OpenAI's Claimed Scientific Discoveries Spark Debate and Doubt

OpenAI announced three major AI-made discoveries in math, CS and physics, igniting debate over whether P vs NP was solved. Brockman downplayed the hype, and analysts soon flagged mismatches between the ETH proof text and its Lean formalization.

2026-10-26 ~ 2026-10-27 · 4 episodes · 121 posts

Episode 1 · Mathematicians Urge AI Labs Not to Handpick Who Interprets AI Discoveries (2026-10-26, 3 posts)

The IAS-backed advisory group AGMAI and nine prominent mathematicians urged AI labs not to handpick researchers to first receive and interpret AI-generated mathematical discoveries, keeping human mathematical progress organic.

Episode 2 · OpenAI announces three AI-driven discoveries in math, CS and physics (2026-10-27, 110 posts)

On October 26, OpenAI announced that its AI models have produced three major scientific discoveries spanning mathematics, computer science, and physics—a rare scale of results covering harmonic analysis, complexity theory, and quantum field theory, two of which are directly tied to Millennium Prize Problems. The company says it has provided Lean formal verification and organized independent academic review, and the academic community is watching the verification process closely amid the shock.

Confirmed

  • Mathematics: OpenAI released a preprint (official repo openai/math, already with 16k stars), stating that AI invented a new mathematical technique, "local smoothing for the Euclidean wave equation," proving the 1991 local smoothness conjecture in all dimensions, and thereby simultaneously resolving the Kakeya conjecture and the Stein restriction conjecture in all dimensions, as well as a 35-year-old open problem on wave focusing. Previously, the only results in this direction were the 2019 two-dimensional work of Guth-Wang-Zhan and some 2025 three-dimensional progress.
  • Computer science: The model claims to refute the Exponential Time Hypothesis (ETH), which asserts that problems like 3-SAT admit no subexponential-time algorithms and underpins complexity theory and numerous lower-bound proofs. According to m17, theoretical computer scientist Ryan Williams announced that an internal OpenAI model produced a refutation of ETH; Andrew Curran also posted a paper link on X claiming ETH does not hold. If 3SAT has a subexponential algorithm, the implications would be profound, since 3SAT is NP-complete.
  • Physics: Per m4's account of the official OpenAI blog post "Building scientific understanding with academic communities," the model produced a proof of Yang-Mills existence (a Millennium Prize Problem) on September 9, 2026, and completed Lean formalization; m7 and m16 also mention results on quantum chromodynamics (QCD) existence and the Standard Model.
  • Former OpenAI president Greg Brockman and Polymarket, among others, publicly shared the news; on Polymarket, the market for "OpenAI announces another Millennium Prize Problem solved by year-end" was at 54%.
  • The academic community is organizing independent reviews; multiple posts show scholars including Andrew Curran and Ryan Williams have joined the discussion.

Not yet confirmed

  • None of the proofs have (within the material covered by these posts) passed full peer review; they are all "claims" awaiting independent review conclusions.
  • Posts conflict on attribution: most attribute the results to OpenAI, but m7 says the news comes from Harmonic AI via danintheory (the harmonic-related team), and m13 says there are four results rather than three. The version based on OpenAI's official release says three.
  • m8's author esrtweet observed that people in the field have begun discussing "whether ETH being false implies P=NP"; he personally still leans toward P≠NP but says his confidence dropped noticeably over the past 4 hours—this is speculation, not a conclusion.

Why it matters

  • If the proofs hold, the refutation of ETH would shake the complexity-theoretic foundations and cryptographic lower-bound proofs underlying hundreds of papers. Cryptographer Matthew Green marveled that the result arrived far earlier than expected—he had thought such a "bizarre" breakthrough wouldn't appear until around Thanksgiving—and quipped that none of those who, two weeks ago, were so certain about the models' capability limits predicted this.
  • Yang-Mills existence is a Clay Millennium Prize Problem, and the Kakeya and Stein restriction conjectures are core open problems in harmonic analysis. AI simultaneously approaching or solving top-tier problems across three fields is seen as a landmark moment in AI shifting from tool to "agent of scientific discovery"—as Brockman put it, perhaps marking the point where AI begins genuinely helping humanity deepen its understanding of the universe.

90 more related posts →

Episode 3 · Brockman on AI math hype: P probably still doesn't equal NP (2026-10-27, 6 posts)

After OpenAI claimed that AI solved several major mathematical conjectures within weeks, sparking speculation about whether P vs NP had been cracked, multiple industry figures weighed in. When asked by Burny whether "the Riemann hypothesis or P=NP will be solved first" and "whether agents have already solved P vs NP," OpenAI President Greg Brockman (@gdb) responded that "P is very likely still not equal to NP," expressing skepticism about near-term progress on this millennium problem of computational complexity.

Confirmed

  • Greg Brockman stated clearly on 10-27 that "P is still very likely not equal to NP," his core response to questions about OpenAI's math breakthrough hype (m1/m4/m5).
  • OpenAI researcher and former Microsoft VP Sébastien Bubeck, replying to gdb, took the same position, saying "p most likely still does not equal np" (m2).
  • Eric Jang responded to gdb, suggesting the framing may be "asking the wrong question": the truly interesting territory lies between "subexponential-time 3-SAT" and "P≠NP," and polynomial-time methods could yield intriguing results on average-case bounds for approximation error (m3).

Why it matters

  • P vs NP is the central unsolved problem in computational complexity theory; whether AI can crack such problems directly bears on the capability boundaries of large models and agents.
  • Notably, those speaking out are closely tied to OpenAI (Brockman is a co-founder, Bubeck an OpenAI researcher), and their move to temper the "P=NP is near" reading implied by their own company's "math breakthrough" marketing can be seen as a restrained signal against overhyping.
  • Eric Jang's addition shifts the discussion from "equal or not" to a finer-grained landscape of algorithmic complexity, offering a more concrete angle for future technical debate.

Episode 4 · AI Claimed Solution to ETH Problem Faces Verification Doubts (2026-10-27, 2 posts)

Doubts have emerged over claims that AI solved the ETH problem, as mismatches between the textual proof and its Lean formalization suggest verification could take a long time. Commentators note this does not mean the solution is wrong, but the cleanup work may create new tasks for mathematicians.