OpenAI claims AI models have disproved Exponential Time Hypothesis and unveil new math‑physics breakthroughs

OpenAI announced that its latest generation of artificial‑intelligence models has produced a formal proof that the Exponential Time Hypothesis (ETH) is false, a claim that challenges a long‑standing assumption in computational complexity theory. The company also released a batch of new results addressing open problems in mathematics and physics.

The Exponential Time Hypothesis posits that certain classes of problems, notably the satisfiability of Boolean formulas with a fixed clause width, cannot be solved in sub‑exponential time. A proof that the hypothesis does not hold would reshape expectations about the ultimate limits of algorithmic efficiency and could have ripple effects across fields that rely on hardness assumptions, such as cryptography and optimization.

OpenAI said the breakthrough emerged from a workflow that combines large‑scale language models trained on scientific literature with automated theorem‑proving tools. Researchers supplied the model with a formal statement of ETH and asked it to generate a proof outline, which was then refined and checked by symbolic verification software. According to the announcement, the resulting proof passed internal consistency checks and was subsequently reviewed by external experts before public release.

The development follows a growing body of work where AI systems assist or autonomously produce mathematical insights. Earlier efforts have seen language models solve International Mathematical Olympiad‑style problems, while DeepMind’s AlphaFold transformed protein‑structure prediction. OpenAI’s claim adds a new dimension by tackling a foundational conjecture in theoretical computer science rather than a domain‑specific puzzle.

Alongside the ETH result, the organization published a technical report detailing several other discoveries. These include novel constructions in number theory, progress on combinatorial optimization conjectures, and a set of calculations that advance understanding of quantum field‑theoretic models. The report provides the formal statements and proof sketches, inviting the broader community to verify and extend the findings.

Academic reaction has been cautiously optimistic. Scholars acknowledge the potential significance but emphasize the need for rigorous peer review and independent replication. Some experts note that the verification pipeline—linking language‑model output to formal proof assistants—offers a promising path for scaling mathematical research, while others warn that premature publicity could outpace the validation process.

If the falsification of ETH is confirmed, it could open avenues for designing algorithms that run faster than previously thought possible for a range of hard problems. Such a shift would influence both theoretical investigations and practical applications, from cryptographic protocol design to large‑scale data analysis.

OpenAI indicated that future work will focus on open‑sourcing key components of the proof‑generation pipeline and establishing collaborations with university labs to test the system on additional open problems. The company also plans to host workshops aimed at integrating AI‑assisted proof techniques into standard mathematical practice.

The announcement highlights the accelerating convergence of artificial intelligence and fundamental scientific inquiry. While the claims remain subject to detailed scrutiny, the episode illustrates how advanced language models are beginning to play a substantive role in exploring questions that have long been considered the province of human mathematicians and physicists.