AI Labs Challenge Core Assumptions of Computational Complexity

The Exponential Time Hypothesis (ETH) has long served as a bedrock for understanding computational difficulty, suggesting that solving certain logic puzzles requires exponentially more time as variables increase. If the claims from OpenAI and Anthropic hold, their models have developed algorithms capable of sub-exponential performance, effectively challenging the assumption that these problems are inherently intractable. MIT professor Ryan Williams, who is coordinating the academic response, noted that while some segments of the proofs appear routine, others contain genuinely novel, alien-looking logic that warrants urgent collaborative study.

This development has created friction regarding the methodology of scientific discovery. Williams, who previously criticized OpenAI’s approach as a disruptive dump of information, is now leading a working group to interpret the findings alongside UC San Diego’s Russell Impagliazzo and Ramamohan Paturi—the researchers who originally defined the ETH in 1999. The immediate priority is determining whether these new algorithms compromise existing cryptographic systems. While OpenAI science lead Dan Roberts has expressed confidence that current security standards remain intact, the broader scientific community faces the prospect of revisiting a decade of research that relied on the now-contested hypothesis.

As AI models increasingly encroach on high-level mathematical inquiry, the role of the human researcher is shifting. Williams suggests that these tools are compelling scientists to raise the bar for their own work, moving away from problems that AI can now solve as low-hanging fruit. For now, the focus remains on peer review and establishing new norms for how frontier labs interact with the academic establishment, ensuring that these high-stakes proofs are subjected to the necessary rigor rather than released in a vacuum.