OpenAI claims AI models refuted the Exponential Time Hypothesis

Refuting a core computer science tenet

OpenAI announced on October 26 that its internal AI models produced a result refuting the Exponential Time Hypothesis. This conjecture has served as a cornerstone of computer science for decades, providing a framework for how quickly computers can solve specific types of logic problems. Because it underpins hundreds of studies regarding computational limits, a refutation could have wide-reaching implications for the field.

The Exponential Time Hypothesis essentially suggests that certain problems require an exponential amount of time to solve as the input size grows. By challenging this, the AI-generated result suggests that these computational boundaries might be different than previously believed. This discovery was reported by R&D World as part of a series of scientific breakthroughs the company has claimed recently.

Researchers from UC San Diego and MIT have noted that the hypothesis connects decades of research on computational difficulty. They explained that while the techniques developed around the hypothesis retain value, the interpretation of the underlying limits could change. However, these experts clarified that their involvement does not serve as a formal certification of OpenAI's specific claim.

Breakthroughs in math and physics

Beyond computer science, OpenAI reported that its models addressed several other longstanding scientific questions. In the realm of mathematics, the company claims to have proved the local smoothing conjecture in every spatial dimension. This result is significant because it resolves questions about how waves concentrate and carries implications for the Kakeya and Stein restriction conjectures.

In the field of physics, the company reported a rigorous mathematical construction of interacting quantum field theories in four dimensions. This includes quantum chromodynamics, which is the theory describing the strong force that binds quarks together inside protons and neutrons. These results represent a significant leap in using AI to handle complex, high-level theoretical constructions.

The announcement also included an existence proof for pure Yang–Mills theory. This is a component of the Clay Mathematics Institute’s Yang–Mills Millennium Prize problem. While the Millennium Prize also requires proving a mass gap, OpenAI's work specifically covers the existence proof portion of that challenge.

Formal verification and transparency

To support these claims, OpenAI published writeups for the mathematics and computer science results. They also provided formalizations in Lean, which is a specialized software used for checking the validity of formal mathematical proofs. By using Lean, the company aims to provide a way for the community to verify that the AI-generated logic holds up to rigorous standards.

However, the company chose to withhold the materials regarding the physics proofs from the public. This decision came after the independent Advisory Group on Mathematics and Artificial Intelligence called for the immediate release of the physics proofs. The group argued that scientific understanding should develop in the open so that everyone has equal access to the findings.

OpenAI stated that the external physics group chose to withhold the AI-produced materials while they prepared their own account of the work. This creates a temporary gap in public access to the physics results while the community waits for a formal presentation. The company continues to position these AI outputs as a way to improve our general understanding of the universe.

Anthropic's independent discovery

Shortly after OpenAI's announcement, Anthropic reported a similar result involving its Claude model. Anthropic claimed that its internal Claude research model found a refutation of the Exponential Time Hypothesis without any human input. This discovery was reportedly made on October 16, which is the same date OpenAI reported its own discovery.

Anthropic researchers subsequently worked to improve the presentation of the result found by the model. Unlike OpenAI, Anthropic included a Lean formalization in their release to allow for verification. They also issued a caution that this theoretical result does not mean there is a practical speedup for solving real-world problems.

The fact that two separate companies reached a similar conclusion on the same day is notable for the research community. It suggests that current AI models may be reaching a point where they can identify core algorithms for complex conjectures independently. Both companies are now navigating the balance between making these high-level claims and providing the necessary evidence for the broader scientific community.