OpenAI has announced three AI-generated results spanning mathematics, theoretical computer science and physics, but its decision to release different amounts of supporting material has drawn criticism from researchers who want the work opened to wider scrutiny.
The company says it is consulting outside experts to determine how best to communicate the findings. Its approach varies by discipline: the mathematics output is public, the computer science result will also receive an explanation from an external research group, and the physics materials remain unreleased while an outside group advises on disclosure.
Researchers interviewed by The Verge described the claims as potentially significant while questioning why experts selected by OpenAI should have privileged access. Their objections concern the process for evaluating the work, rather than a consensus that the results are incorrect. Establishing whether the claims hold will require substantial technical examination.
The mathematics announcement purports to resolve the Kakeya conjecture in every dimension. In its familiar geometric formulation, the problem concerns sets containing a line segment pointing in every direction and asks how small such sets can be. Brendan Hassett, a mathematics professor at Brown University, told The Verge that the claim appeared well supported, while emphasizing that he had not yet checked its Lean formalization.
Lean is a proof assistant that can check mathematical arguments expressed in a formal language. Such verification can provide strong assurance that a conclusion follows from specified definitions and assumptions. It does not eliminate the separate work of explaining the argument in terms that researchers can readily understand, assess and build upon.
The computer science result claims to disprove the exponential time hypothesis, an influential assumption about the computational difficulty of certain problems. The hypothesis concerns limits on how efficiently those problems can be solved and is connected to the broader study of computational complexity. Disproving it would not, by itself, settle the separate P versus NP question.
The physics work concerns quantum chromodynamics, the theory describing the strong nuclear force, and its relationship to Yang–Mills theory. OpenAI’s announcement does not claim to resolve the Yang–Mills mass-gap problem, one of the Millennium Prize problems. With the underlying physics materials withheld, independent researchers have less opportunity to assess the result directly.
The release follows friction over hundreds of AI-generated preprints distributed earlier in October, The Verge reported. Several researchers welcomed the smaller batch and OpenAI’s consultation with specialists, but argued that expert assistance should accompany public access rather than precede it.
The Advisory Group on Mathematics and Artificial Intelligence, which OpenAI had previously pledged to consult, had warned against giving selected researchers early access a day before the announcement. The group argued that the practice risks unequal access, selection bias and delays, while lending the reputations of chosen academics to work others cannot yet examine.
Its preferred approach is to publish outputs in scholarly repositories, include formalized proofs and explain how problems were selected. It also recommends funding and other support for researchers undertaking the work of interpreting the results.
After the announcement, the advisory group called for immediate publication of the physics result’s natural-language and Lean materials. It said OpenAI had described the formalization as complete, removing the scientific reason the company had given for delaying release. OpenAI says it is experimenting with communication approaches and acknowledges that each carries tradeoffs.
Sources: The Verge (AI)