OpenAI Shakes Complexity Theory — and Claims Two More Scientific Breakthroughs

The Exponential Time Hypothesis was proposed by Russell Impagliazzo and Ravi Paturi in the late 1990s. Its core claim is that certain satisfiability problems — specifically 3-SAT — cannot be solved in subexponential time. ETH is not merely a statement about one problem; it underpins a large family of complexity lower bounds across algorithm design, providing the theoretical justification that specific algorithms are essentially optimal. A preprint from OpenAI's math team now claims that hypothesis is false.

If the result survives peer review, roughly two decades of complexity theory literature faces a difficult reckoning — not invalidation, but the evaporation of the foundational justification for many published lower bounds. The 242-comment HN thread is itself a live stress test: complexity theorists with academic affiliations are pointing to specific steps they believe require more justification, while others argue the approach is novel enough to deserve serious engagement before dismissal. The honest assessment, as of publication, is that the community has not had time to fully verify the proof.

OpenAI simultaneously published two additional discoveries. A materials science result involves predicting crystal structures with a model architecture that reportedly outperforms existing methods on a standard benchmark by a margin that would normally represent several years of research progress. A number theory result touches on patterns in prime distributions. All three releases were framed as collaborative with academic communities rather than unilateral corporate announcements — a deliberate positioning of frontier AI labs as scientific organizations. Scientific credibility at this level also drives talent acquisition: if a for-profit lab can point to Nature-class discoveries, it can recruit researchers who might otherwise decline to consider it.

Sebastian Raschka's from-scratch course on building reasoning models — the architecture class underpinning OpenAI's o-series and the chain-of-thought reasoning central to the mathematical discovery claims — drew 727 points and more than a hundred comments on HN. Praise focused specifically on the pedagogical rigor: no hand-waving, literal walk-throughs of code from the ground up. The number of people who genuinely understand how reasoning models work at an implementation level is, by most estimates, a few hundred globally. Expanding that denominator matters directly for the field's ability to audit and critique results like the ETH claim.

Fred Brooks' fifty-year-old Mythical Man-Month received a fresh reading in this context, with a new essay asking whether Brooks' central law — that adding engineers to a late software project makes it later — still holds when the new team member is an AI coding assistant. The author's answer is nuanced: AI assistance may reduce coordination costs, but it doesn't reduce the conceptual complexity of understanding what is being built and why. The essay argues that AI tooling may tempt teams to take on more complexity than they can cognitively manage, because initial velocity feels deceptively high.