OpenAI says its internal models disproved the Exponential Time Hypothesis in a new science dump, adding a claim that reaches into 3-SAT and non-polynomial-time complexity. The same release bundles two other math and physics results, but the company says the raw output still needs the math and science communities to turn it into usable knowledge.
OpenAI and 3-SAT
The claim reaches one of the hardest questions in algorithm research. OpenAI ties the Exponential Time Hypothesis to 3-SAT, the kind of problem used to test whether some computations resist any non-polynomial-time shortcut, and says its internal models have now disproved the hypothesis altogether.
That is a bigger statement than a routine benchmark win. If the claim survives scrutiny, it would cut into a foundation that shapes how researchers think about optimization and algorithms, because the hypothesis sits close to the boundary between problems that can be solved efficiently and problems that appear to resist fast exact methods.
Lean and 377 math results
The company had already dumped 377 math results on GitHub at the beginning of the month, and it says the new material was formalized or verified in Lean. Lean is the proof assistant OpenAI says it uses to reduce the need to check and recheck work by hand, which is the point here: the model output is not being pitched as polished prose, but as structured material that can be checked mechanically before humans decide what holds.
OpenAI also says it does not seek academic credit for the discoveries produced by its internal models. Instead, it says the model-generated texts are best understood as raw research data, with the math and science communities responsible for converting that data into scientific understanding. The company says it has already reached out to external experts for that job.
Kakeya and four dimensions
The latest science dump does not stop at algorithms. OpenAI also says it resolved the Kakeya conjecture by finding a novel proof for all dimensions in the local smoothing conjecture, and it says it has a proof involving certain interacting quantum field theories containing gauge fields and fermions in four dimensions.
That set includes Quantum Chromodynamics and a unified completion of the Standard Model of particle physics. The company’s plan is to have external experts act as shepherds who assemble groups around the outputs and help convert them into new human knowledge, while the blog post also acknowledges that choosing the first shepherds shapes who gets involved first and creates asymmetry of information.
The unresolved issue is not whether OpenAI can generate more material; it already has. The question is what exact proof supports the Exponential Time Hypothesis claim, and how much of these model-generated results the broader research community will accept once the raw data leaves OpenAI’s own hands.
