Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?
Original reporting by TechCrunch
The Exponential Time Hypothesis (ETH) refers to a foundational computer science assumption asserting that certain computationally hard problems require exponentially increasing time to solve as they grow in size. This durable premise, which underpins our understanding of what is difficult to compute and relates directly to the enduring P vs NP problem, was recently challenged when OpenAI released solutions to three blockbuster math problems. One of these was a surprising rejection of ETH itself. The announcement immediately sent ripples through scientific communities, sparking a mix of thrill and dismay. Rival lab Anthropic quickly joined the fray, suggesting OpenAI’s work might build on pre-existing research, and hours later released its own refutation of ETH.
Evolving AI Science
OpenAI’s process for engaging with the scientific community, however, is evolving. For these complex findings, the lab structured its releases differently, most notably for computer scientists. This involved forming a working group led by MIT Professor Ryan Williams—a prominent computer scientist who had previously criticized OpenAI’s "proof dumps." Williams, along with the original authors of ETH, is now collaborating with OpenAI to interpret the proof's "alien-looking parts" and help the field navigate its profound implications. Their first step was an open letter calling for community feedback and a discussion of new norms around AI contributions to science. The potential refutation of ETH challenges decades of foundational assumptions about computational complexity, demanding a collective effort to re-evaluate established research and chart a path forward in this new era of AI-driven scientific discovery.
The potential refutation of the Exponential Time Hypothesis, spearheaded by OpenAI and corroborated by Anthropic, marks a pivotal moment in computer science. If these claims hold, they demand a painstaking re-evaluation of foundational assumptions that underpin decades of research into computational limits. The immediate task for the scientific community involves rigorous validation of these "alien-looking" proofs and a collective effort to understand their full ramifications, particularly for fields whose models rely on ETH's durability. This breakthrough not only challenges specific mathematical theories but also casts a new light on the plausibility of P = NP, potentially revolutionizing our understanding of what can be "easily" computed.
New Scientific Paradigms
Beyond the specific mathematical breakthrough, this episode highlights a profound shift in the nature of scientific discovery itself. AI models are emerging not merely as sophisticated tools for analysis, but as independent generators of foundational insights, forcing a redefinition of scientific collaboration. The nascent working groups and calls for open discourse, though born from initial friction, signal a necessary evolution in how AI-driven discoveries are vetted, shared, and integrated into established scientific norms. This new paradigm challenges human researchers to elevate their ambition, pushing the boundaries of inquiry beyond problems an AI might readily solve, while also necessitating new frameworks for training the next generation of scientists in an era of AI co-discovery. While immediate cybersecurity concerns appear to be mitigated, the broader implications for cryptography and our fundamental understanding of computational hardness suggest that this is merely the opening chapter in a rapidly accelerating story of AI's transformative impact on science, inviting both cautious optimism and urgent collaboration.