In a major development for computer science and mathematics, frontier artificial intelligence laboratories have announced model outputs that directly challenge one of the field's core assumptions.
OpenAI released solutions to three long-standing mathematical problems, including a surprising rejection of the Exponential Time Hypothesis (ETH). The release quickly prompted responses across academia and competitor labs, highlighting both the immense potential and operational friction of AI-driven scientific discovery.
The ETH Refutation and Lab Reactions
The Exponential Time Hypothesis (ETH) has served as a foundational assumption in computer science, shaping how researchers understand computational complexity and the threshold of what can be efficiently computed. Yesterday, OpenAI dropped solutions to three blockbuster mathematical problems that have long challenged the academic community, with one solution centered on disproving ETH.
The announcement created a immediate reaction across scientific circles, generating a combination of excitement and skepticism. Anthropic mathematician Levent Alpöge pointed out that OpenAI's solutions appeared to draw from pre-existing academic work. Only hours later, Anthropic published its own refutation of the Exponential Time Hypothesis, escalating the rivalry between the two frontier AI organizations in high-level mathematics.
Tailored Release Models Across Disciplines
Unlike previous unannounced proof drops, OpenAI adjusted its distribution process by structuring releases specifically for different scientific domains:
- Mathematicians: Received the technical proofs directly via GitHub repositories.
- Computer Scientists: Formed a specialized working group that requested public access to proofs and asked OpenAI to sponsor a formal conference to evaluate the findings.
- Physicists: Established a dedicated working group to determine the safest and most effective protocol for releasing their discipline's result.
Academic Collaboration and Community Response
MIT Professor Ryan Williams, a leading expert in computational complexity, is organizing the computer science community's evaluation of the finding. Williams has previously voiced frustration over how AI labs publish complex theoretical proofs, once describing an earlier proof release from OpenAI as a "DDOS attack on mathematicians."
Despite prior concerns, Williams agreed to collaborate directly with the frontier lab to establish standard evaluation workflows for the field. OpenAI asked Williams to help interpret the computational solution alongside UC San Diego computer scientists Russell Impagliazzo and Ramamohan Paturi—the researchers who originally formulated and published the Exponential Time Hypothesis in 1999.
As an initial step, the team published an open letter calling on researchers to evaluate the findings and start an industry-wide discussion around the norms governing AI-generated scientific contributions.
"Some parts are, maybe you could say routine, but then yes, there are two or three alien-looking parts, and it’s exactly these parts that I saw that made me think, wow, this is something that the whole community needs to work on together," Williams stated regarding OpenAI's proof.
Understanding P vs NP and Algorithm Speeds
The Exponential Time Hypothesis is intrinsically linked to P vs NP, one of the most famous unsolved problems in computer science. Broadly, P vs NP asks whether every problem whose solution can be easily verified by a computer (NP) can also be easily solved by a computer (P).
A classic illustration of this dynamic is the game of Sudoku. Verifying a completed Sudoku grid is fast and straightforward, but finding the correct numbers for an empty grid from scratch is time-consuming. This asymmetry underpins modern digital infrastructure, particularly cryptographic systems: having the decryption key allows immediate verification, but brute-forcing the key requires an impractically long amount of time.
If P equals NP, it implies that anything quickly verifiable can also be solved quickly. Because this outcome seems implausible to most computer scientists, the consensus has historically leaned toward P ≠ NP. ETH provided formal support for this consensus by asserting that even the most optimized computer algorithms require exponentially more computational time as input variables increase in complex logic puzzles.
Claims from OpenAI and Anthropic suggest their AI models generated algorithms capable of solving these fundamental logic problems in "sub-exponential" time—faster than theoretical limits previously allowed. If validated, these sub-exponential algorithms refute ETH and alter the theoretical plausibility of P vs NP relationship pathways.
Theoretical Revaluations
Refuting ETH forces researchers to reconsider foundational assumptions regarding computational hardness:
- Intuition Adjustments: Computer scientists must re-evaluate long-standing assumptions about how tractable hard NP problems truly are.
- Algorithmic Optimization: Williams noted that while the frontier models offer a breakthrough, OpenAI's generated algorithm is unlikely to be the theoretical ceiling for performance.
Broader Implications for Research and Cybersecurity
Because ETH has been relied upon as a standard premise in computer science research over the past decade, a confirmed refutation means numerous peer-reviewed papers will need to be re-examined and revised. Williams stressed that holding community-wide conferences is essential to map out how the theoretical foundations of computer science must adapt moving forward.
Beyond pure theory, there were immediate questions regarding information security. Breakthroughs in solving complex algorithms efficiently raise concerns that cryptographic systems relying on specific key sizes could be rendered vulnerable. Addressing this, Williams consulted directly with OpenAI science lead Dan Roberts to check if the new algorithms impacted cryptographic safety.
According to Roberts, OpenAI conducted security checks prior to release and felt reasonably confident that the breakthrough did not imply active security breaks for existing cryptographic implementations.
As AI-driven discovery continues to evolve, researchers are adapting to new academic dynamics. While training future generations of scientists presents new challenges, Williams highlighted that automated tools are prompting researchers to raise their ambitions and target deeper, more complex questions now that traditional problem sets are becoming accessible to artificial intelligence.