OpenAI and Anthropic recently stated that they have successfully disproved the exponential time hypothesis. Vast areas of theoretical computer science, complexity studies, and algorithm design have long relied on the assumption that solving certain computational problems requires an amount of effort that scales exponentially relative to the size of the input.
Academic Conferences Face Disruption
Jonsson anticipates that his situation will become common at the upcoming IEEE Symposium on Foundations of Computer Science, where his colleagues will present related work. Researchers will likely open their presentations by noting that the foundational hypothesis is false and acknowledging that segments of their papers require revision. While this presents a vexing hurdle, scientists view it as an opportunity to learn and advance their understanding.
Independent researchers have not yet verified the AI-generated proofs, which both organizations released on GitHub in both natural language and the machine-checkable Lean format. OpenAI reached out to the UC Santa Barbara computer scientists who originally formulated the hypothesis, alongside an MIT academic who previously challenged a variant of it, to help the broader academic community examine the proofs. This outreach involves a feedback form for volunteers and a commitment from OpenAI to sponsor an academic conference on the subject.
Daniel Lokshtanov, a professor at UC Santa Barbara, notes that if the findings hold up, his own work will require major revisions, potentially affecting around 50 of his published papers. He describes the development as a massive disruption dropped directly into the field, calling it fascinating and troubling at the same time.
Rethinking Years of Algorithms Research
Thore Husfeldt, who leads the algorithms research group at the IT University of Copenhagen, dedicated months of work this year to a single problem within his subfield. He now considers that entire line of inquiry obsolete following the recent revelation. Having initially formulated a version of the hypothesis that OpenAI's work challenges, he has spent time at his blackboard trying to determine the extent to which his past calculations are affected by the shift.
Despite the disruption, Husfeldt views the outcome with enthusiasm rather than fatigue. He compares mathematical research to navigating a dark room, noting that the new proofs act like floodlights illuminating an entire subfield for exploration. While the rapid pace of AI-driven discoveries has overwhelmed some academics and introduced psychological strain within departments, others find motivation in the newly opened avenues of study.
The sheer volume of material generated by automated systems has also presented challenges. Researchers face an influx of manuscripts and unverified papers, making it difficult to digest every new claim. Earlier releases from OpenAI, which included hundreds of manuscripts dropped simultaneously, drew criticism from academics who pointed to errors and a rushed rollout aimed at commercial milestones rather than methodical scientific progress. However, the more recent detailed paper on the hypothesis has been received as a more carefully structured document.
Victor Lagerkvist, an associate professor at Linköping University, suggests that abandoning the assumption of exponential scaling opens the door to revisiting problems previously abandoned as intractable. While the findings do not immediately compromise the encryption protocols securing web infrastructure, experts suggest that cryptosystem designers should take note of how shifting computational boundaries might influence future security models.
Researchers intend to analyze the proofs thoroughly to identify new deductions and applications. Jonsson plans to reexamine his previous studies and remains optimistic about uncovering fresh insights through the detailed analysis of the AI-generated results, emphasizing that truth in scientific progress transcends the origin of the proof.
