Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

OpenAI says its model disproved the ETH hypothesis and solved two more problems

OpenAI says its internal model solved three open scientific problems, the main one being a disproof of the Exponential Time Hypothesis (ETH). The hypothesis holds that the Boolean satisfiability problem (SAT) cannot be solved much faster than by full brute-force search, much like a lock with a huge number of switches. Hundreds of papers rely on ETH, and their "cannot be faster" estimates are now in question. For now these are the company's words; the authors of ETH will check them.

AI-processed from TG: @data_secrets (Data Secrets); edited by Hamidun News
OpenAI says its model disproved the ETH hypothesis and solved two more problems
Source: TG: @data_secrets (Data Secrets). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @data_secrets (Data Secrets) ↗
Editor-in-Chief
Key metric #1ETH will be verified by a group led by Russell Impagliazzo and Ramamohan Paturi, the authors of the hypothesis.
Key metric #2The Kakeya conjecture is proven for any dimension; previously it was only known for dimensions 2 and 3.
Key metric #3The Yang-Mills problem will be reviewed by a group led by Nobel laureate David Gross.

OpenAI said its internal, not yet released model solved three open problems from mathematics, theoretical computer science and physics. The main one is a disproof of the ETH hypothesis, on which estimates of how fast a computer can solve hard problems have rested for years. If the result holds up, it would be one of the most high-profile cases of AI advancing fundamental science rather than restating other people's papers. But so far this is the company's own claim, and there has been no independent verification.

What ETH is and why disproving it would be an event

The Exponential Time Hypothesis sounds dry, but the idea is simple. Take the Boolean satisfiability problem (SAT): there is a long formula of "yes/no" variables, and you need to find an assignment that makes it true. Picture a lock with a huge number of switches. ETH claims there is no short path to the right combination: the problem cannot be solved much faster than trying every option.

For decades this assumption served as a foundation. The logic goes like this: if a problem from graph theory, databases, geometry or computational biology could be solved faster than a certain limit, that would also yield a fast way to solve SAT, and nobody believed that was possible. So for hundreds of papers, ETH worked as a reference that said "it can't be faster, we checked." According to OpenAI, the model disproved this hypothesis. That reopens all questions of the form "how much faster can it still be?"

Practitioners should avoid reading too much into this. It does not follow that SAT solvers will become a thousand times faster tomorrow. The original announcement does not say how the disproof works or whether it yields a concrete fast algorithm. But if developers of solvers, optimizers or verifiers relied on ETH when estimating limits, this review is worth following.

The other two problems: waves and particle physics

The second problem concerns local smoothing and the Kakeya conjecture. It is about waves that focus at a single point. The Sogge conjecture from 1991 asserts that the strongest focusing cannot persist for long. From it follow the Kakeya conjecture and the Stein restriction conjecture. This is one of the central problems of harmonic analysis, and strong mathematicians have worked on it for decades, but proofs were only obtained for low dimensions: two or three. The OpenAI model, according to the announcement, produced a proof for any dimension and automatically closed the Kakeya and Stein conjectures for all dimensions.

The third problem is the existence of Yang-Mills theory, part of one of the seven Millennium Prize Problems. The model, as claimed, rigorously proved the existence of this theory, as well as of full quantum chromodynamics (QCD), where quarks are added to gluons, and of a unified theory with the particles of the Standard Model. The original post does not specify whether the result closes the Millennium Prize problem in full or only part of it. Experts will need to weigh in on this in their reviews.

How verification will work and what to do about it

The most interesting part of the story is the verification scheme. ETH will be checked by a group led by Russell Impagliazzo and Ramamohan Paturi, the scientists who formulated the hypothesis. The Yang-Mills result will be reviewed by a group led by Nobel laureate David Gross. OpenAI also pledges to fund a conference devoted to making sense of the results. The company published details on its page about working with academic communities.

For developers, product managers and founders, several practical takeaways follow.

  • Don't treat the claim as fact until it has been verified. Proofs of this complexity take months to study, and errors have been found even in work by professional mathematicians.
  • Pay attention to the format. The model delivers not a black-box answer but a proof that can be read, disputed and checked. That kind of output is the easiest to fold into scientific work.
  • Assess availability realistically. The model is internal, it is not in the API, and how much compute went into each solution is unknown. For now this is a showcase of the lab's capabilities, not a tool for your stack.
  • If your product depends on algorithmic complexity estimates, track the ETH review: the boundaries of what was considered impossible may shift as a result.

If verification confirms at least one of the three problems, the conversation about AI will shift. The question "can the model write code" will give way to "can it prove what humans could not." If it does not hold up, it will still be a useful lesson in telling a striking demonstration from a verified result. Either way, the coming weeks will show how the claim stands up to expert scrutiny.

Story

Key Questions Answered

What ETH is and why disproving it would be an event?

The Exponential Time Hypothesis sounds dry, but the idea is simple. Take the Boolean satisfiability problem (SAT): there is a long formula of "yes/no" variables, and you need to find an assignment that makes it true. Picture a lock with a huge number of switches. ETH claims there is no short path to the right combination: the problem cannot be solved much faster than trying every option. For decades this assumption served as a foundation.

What are the key details on The other two problems: waves and particle physics?

The second problem concerns local smoothing and the Kakeya conjecture. It is about waves that focus at a single point. The Sogge conjecture from 1991 asserts that the strongest focusing cannot persist for long. From it follow the Kakeya conjecture and the Stein restriction conjecture. This is one of the central problems of harmonic analysis, and strong mathematicians have worked on it for decades, but proofs were only obtained for low dimensions: two or three.

How verification will work and what to do about it?

The most interesting part of the story is the verification scheme. ETH will be checked by a group led by Russell Impagliazzo and Ramamohan Paturi, the scientists who formulated the hypothesis. The Yang-Mills result will be reviewed by a group led by Nobel laureate David Gross. OpenAI also pledges to fund a conference devoted to making sense of the results. The company published details on its page about working with academic communities.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…