- Faster Algorithm: OpenAI says AI produced an algorithm that would refute a key assumption about how quickly certain computing problems can be solved.
- Wave Mathematics: A mathematics manuscript claims local-smoothing bounds on wave concentration over time across all spatial dimensions from two upward.
- Physics Access: OpenAI is sharing a particle-physics claim with selected experts before wider release, while the two mathematics manuscripts are public.
OpenAI says AI has produced an algorithm that would overturn the Exponential Time Hypothesis, a key assumption about how quickly certain hard computing problems can be solved. The claim is part of an announcement from October 26 spanning wave mathematics and particle physics, with two manuscripts public and the physics work initially shared with selected experts.
The October 26 release added manuscripts on the Exponential Time Hypothesis and local smoothing, together with Lean formalizations, which express mathematical arguments for computer checking. A third claim concerns the rigorous construction of quantum field theories. The announcement combines potentially consequential results with different arrangements for deciding whether their arguments hold and making them understandable to other researchers.
A Challenge to an Exponential Barrier
The Exponential Time Hypothesis, or ETH, concerns how quickly certain Boolean problems can be solved. A Boolean formula describes conditions using true-or-false variables. Satisfiability asks whether any choice of those values meets every condition; counting satisfying assignments asks how many choices work.
OpenAI’s formalization scope describes a deterministic, non-random algorithm for exactly counting solutions to fixed-width CNF formulas. A CNF formula joins clauses with AND; within each clause, variables or their negations are joined with OR. Each term is a literal, and fixed width means the maximum number of literals per clause stays constant as the input grows.
The claimed running time is subexponential in the number of variables: its variable-dependent factor eventually grows more slowly than every fixed positive exponential rate. The time bound also includes polynomial overhead in the input length in bits, an extra factor bounded by a fixed power of that length. The polynomial-space bound similarly limits memory growth to a fixed power of the input length.
For clauses with at most three literals, such a result would refute ETH’s assumed exponential-time barrier. That would change the interpretation of research using ETH to establish conditional limits on algorithms. It is an asymptotic claim about how computation scales, rather than a measured speedup for software running on an ordinary workload.
The group organizing examination of the claim includes ETH originators Russell Impagliazzo and Ramamohan Paturi of UC San Diego, alongside MIT’s Ryan Williams. Their open letter separates checking correctness and limitations from producing accessible explanations and exploring consequences for algorithms and lower bounds. They also argue that the ideas, reductions and techniques developed around ETH retain value even if the hypothesis falls.
“Our involvement should not be taken as certification of the claim,” the three researchers write. At their request, OpenAI made the materials public immediately so others could examine them. The organizers want independent approaches and criticism alongside coordinated work that can distribute tasks and avoid duplication.
What the Wave Result Would Establish
The mathematics manuscript concerns local smoothing: bounds on how strongly waves concentrate when their behavior is measured over time. OpenAI claims estimates across every spatial dimension from two upward.
Its published scope note also asserts a geometric consequence involving Kakeya sets, which contain a unit line segment pointing in every direction. The claim gives each such set full Hausdorff dimension, a geometric dimension based on coverings by arbitrarily small pieces. Here, full means matching the dimension of the surrounding space. That would extend the earlier three-dimensional Kakeya maximal result to every spatial dimension from two upward.
Physics Foundations Await Wider Inspection
OpenAI’s physics claims concern rigorous mathematical existence for certain interacting quantum field theories in four spacetime dimensions, including Yang–Mills theories and quantum chromodynamics, or QCD. Quantum field theory describes particles and their interactions through fields; QCD describes the strong interaction.
A rigorous mathematical construction would establish a well-defined theory satisfying specified requirements, giving researchers a firmer foundation for using its ideas in physics and mathematics.
The Clay Yang–Mills problem requires a non-trivial quantum Yang–Mills theory on four-dimensional space, together with a positive mass gap: a minimum excitation energy above the vacuum. OpenAI’s claimed Yang–Mills existence construction addresses the existence part of that requirement.
A group led by physicist David Gross has early access to the physics work. The Neuron reports that the group plans to prepare its own explanation before the underlying materials are released more broadly. No fixed public release date accompanies that reported arrangement.
Columbia mathematician Peter Woit’s response to the announcement centers on the difficulty of assessing a physics claim without precise public definitions and technical detail. His questions concern what the construction establishes, including its relationship to the mass-gap requirement.
Computer Checking and Human Understanding
Lean, an open-source programming language and proof assistant that enables correct, maintainable, and formally verified code, can check whether a formal argument proves a specified theorem from its definitions and assumptions. Its validation documentation distinguishes that task from determining whether the theorem statement means what its author intends. Researchers still need to establish that the formal definitions and conclusion capture the advertised scientific problem.
In OpenAI’s February gluon-formula work, human scientists supplied small-case calculations for interactions between these particles. GPT-5.2 Pro identified a general pattern, an internal model produced a proof, and research authors checked the proposed formula against established physics relations and symmetries.
A contemporary public comparison comes from Levent Alpöge and Ethan Sussman’s Yang–Mills project. Their account describes Claude generating thousands of pages, followed by Codex extracting shorter manuscripts. The authors report that abridgment can introduce errors or leave definitions implicit, and they are working toward digestible, human-written versions. They describe the claims as tentative and not satisfactorily vetted.
Who Gets to Examine the Work
The mathematicians’ group AGMAI had objected on October 25 to labs selecting particular researchers to receive results first. Its concerns included unequal access, lending researchers’ reputations to unverified claims and biased selection. It proposed scholarly repositories containing formalized proofs and information about problem selection, with support for researchers who work to understand the outputs.
The ETH organizers acknowledge that early access can help checking while giving some researchers opportunities others lack.
On October 15, the Simons Foundation, Anthropic and OpenAI announced $15 million in combined funding and resources for mathematics in the AI era. Simons will manage awards supporting workshops, evaluation of AI-assisted results, tools and community standards, with selected researchers receiving access to tools and compute.
