OpenAI released solutions to three major mathematical problems, including a rejection of the Exponential Time Hypothesis, a long-standing assumption about what is difficult to compute. Anthropic mathematician Levent Alpoge suggested the work drew on pre-existing research, then Anthropic released its own refutation hours later. MIT professor Ryan Williams is now organizing the computer science response with Russell Impagliazzo and Ramamohan Paturi, who first published ETH in 1999. For business, the episode matters because it tests how AI-generated breakthroughs get verified and absorbed.
How the ETH refutation was released
OpenAI changed its release format for each of the three disciplines involved in the announcement. Mathematicians received the work as a drop on GitHub, while computer scientists formed a working group and asked for public sharing of proofs plus funding for a conference. Physicists formed a separate working group that will decide how their result should be released. Williams, who earlier called an October proof dump a DDoS attack on mathematicians, agreed to help interpret the solution. The labs asked him to work alongside Impagliazzo and Paturi on assessment.
The computer science group began with an open letter inviting feedback from interested researchers and discussion of norms around AI contributions. Williams said parts of the OpenAI proof look routine, but two or three parts look alien and require joint work by the whole community. The claimed advance is an algorithm that solves a specific class of tricky logic puzzles in sub-exponential time rather than exponential time. If verified, that result would refute ETH and could open a path toward algorithms consistent with P equals NP. Williams added that the OpenAI algorithm is probably not the best possible version.
ETH belongs to the broader P versus NP area, which asks whether problems with quickly verifiable solutions can also be solved quickly. Sudoku illustrates the distinction: checking a completed puzzle is fast, while finding the solution takes much longer. The same asymmetry underpins cryptographic systems, where possessing the key reveals a message but cracking an effective code takes too long to be useful. Most researchers assume P does not equal NP, and ETH supported that view by stating that solution time grows exponentially with variables. The new claims challenge that intuition without yet providing a final answer.
What this means for companies using AI
For companies that rely on optimization, logistics, scheduling, chip design or drug discovery, faster algorithms for hard combinatorial problems could eventually reduce compute time and cost. Large organizations with in-house research teams and heavy workloads would feel such gains first, since they can test new solvers on real instances and integrate them into existing pipelines. Smaller firms would benefit later, through updated solvers, cloud services and software libraries that embed the improved methods. The practical signal is broader availability of tools for problems previously considered too expensive to solve at scale.
The limits are substantial, and the result remains unverified pending community review and the proposed conference. Williams said many papers from the past decade that relied on ETH will need revisiting if the refutation holds, which creates uncertainty for roadmaps built on old hardness assumptions. He also asked OpenAI science lead Dan Roberts whether cryptographic systems were affected, and the lab said it had checked and was reasonably confident that nothing was broken. Buyers should not treat the announcement as proof that encryption with certain key sizes is insecure or that P equals NP has been established.
The marker to watch is whether the working group confirms the sub-exponential algorithm and OpenAI funds the conference to define new foundations. Publication of checked proofs, independent reproduction and clarity on affected problem classes would show the claim has moved from announcement to accepted result. That outcome would prompt vendors and security teams to update guidance on algorithm choice and key sizes. Absence of those steps would leave the episode as an unresolved claim.
