OpenAI says its models overturned a core computer science belief

OpenAI has published two new manuscripts on GitHub, one of which claims a proof that the Exponential Time Hypothesis — an assumption hundreds of research papers across databases, computational biology and geometry rest on — is false. The release comes just a few weeks after the company put out manuscripts with over 700 pages of solutions and progress on 372 major math problems.

This time the scope is wider. The new results don't just cover mathematics. They extend to computer science and physics, and OpenAI has built a different review process around each field. That process, and the question of who gets credit for machine-generated discoveries, is as much a part of this story as the results themselves.

What did OpenAI actually publish?

The two new manuscripts sit on GitHub and are accessible right now. The first establishes Sogge's local smoothing conjecture for waves in every spatial dimension. In plain terms, this is a conjecture about how solutions to the wave equation behave — a question that has sat in harmonic analysis for years and that the model now claims to settle across all dimensions, not just some.

The second manuscript is the more consequential one for everyday computing. It shows proof that the Exponential Time Hypothesis (ETH) is false. ETH is an assumption in computer science that some problems are simply too hard to solve quickly — that no matter how clever you are, certain tasks will always need time that grows exponentially with their size.

As OpenAI notes, ETH is widely believed to be true. It has served as the basis for hundreds of papers across a wide variety of fields, including databases, computational biology and geometry. If the conjecture really is false, then we could, in theory, find better algorithms to solve problems in those fields more efficiently. That is the practical stake: faster methods for workloads that researchers had written off as unavoidably slow.

There's a detail here that grounds the whole exercise. The scientists who originally proposed ETH are now part of the group OpenAI has asked to examine and explain the results its models generated. The people whose conjecture is being overturned are, in effect, the first reviewers of the overturning.

The takeaway: one manuscript closes a long-standing math conjecture, and the other attacks an assumption that hundreds of algorithms papers treat as settled ground.

Why is OpenAI using three different release strategies?

For this release, OpenAI has experimented with three different strategies — one per field — and the differences matter if you care about how AI-generated science gets validated.

For the math results, the company has simply published them. Researchers can look the results up and make their own efforts to explain them. That's the traditional open-manuscript route, the same one it used for the earlier 700-plus-page release covering 372 problems.

For the computer science results, OpenAI formed an external group of expert collaborators. It asked that group to explain the results in their own writeups and materials, while also releasing the raw AI output to the public. So there are two layers: what the model produced, and what human experts make of it, available side by side.

For the physics results, OpenAI has deferred the release of the materials entirely to the group handling them. The company has handed over control of when and how those results surface. The group examining the physics results has yet to publish anything about its work, but it has provided a contact email — YANGMILLS4D[@]gmail.com — for questions about its efforts.

OpenAI says it is welcoming feedback "on how to broaden participation and make this process more useful." That's a small sentence carrying a big implication: the company doesn't consider its current review model finished, and it's asking the research community to help shape how machine-generated results get checked.

The takeaway: math gets open publication, computer science gets parallel expert commentary, and physics gets full human control over the release.

What are the physics results?

The physics portion involves proof of interacting quantum field theories containing gauge fields and fermions in four dimensions. That's a technically dense target, and the source gives little elaboration — which is itself telling. The reviewing group hasn't published its writeup yet, so the public record currently consists of OpenAI's description plus an email address.

The choice to defer release to the external group is the interesting part of the mechanism. Rather than dropping raw model output into the world as it did with the math papers, OpenAI has let domain experts decide the pace. Whether that produces a more rigorous or simply a slower disclosure remains to be seen, and nothing has been published from this track so far.

Who gets credit for discoveries made by a model?

OpenAI has drawn a clear line on attribution, and it's the part of this announcement other research organizations will watch closely. The company says it will not seek any academic credit for its models' discoveries.

Instead, the people in each expert group will get credit for the writeups and materials they create. The human explainers own the explanation, not the machine that produced the underlying result.

The groups also carry a responsibility that usually falls to authors: making sure the AI-generated results appropriately credit all the researchers, techniques and ideas the models drew on. Training data provenance, in other words, becomes part of the formal citation record. That's an unusual arrangement — the reviewers become both interpreters and bibliographers.

The takeaway: OpenAI is testing a credit model where humans own the writeups and the citation duty, while the company steps back from authorship entirely.

Why does this matter beyond the proofs?

Three weeks separated this release from the last one, and the scope expanded from 372 math problems to results spanning three fields. The cadence alone signals that OpenAI treats AI-generated research output as a product line now, not a one-off demonstration.

The bigger stakes are epistemic. If ETH falls, hundreds of papers that assumed its truth need re-examination, and fields from databases to computational biology may have room for faster algorithms in theory. If Sogge's conjecture holds up under scrutiny, a standing question in wave behavior is closed. If the quantum field theory proof survives its reviewers, a notoriously hard physics target moves.

And the verification process OpenAI is running — open manuscripts for math, expert parallel writeups for computer science, deferred human-controlled release for physics — is itself an experiment in how the scientific community should handle results no human derived. With feedback explicitly invited on broadening participation, expect this review structure to evolve as the expert groups publish their assessments.