Groundbreaking Developments in Theoretical Computer Science
Professor Peter Jonsson from Linköping University is set to present his research in two weeks at one of the most esteemed conferences in theoretical computer science. However, he has expressed that a significant portion of his work became outdated earlier this week when artificial intelligence demonstrated that a long-standing hypothesis underpinning it, the Exponential Time Hypothesis (ETH), was incorrect. This hypothesis has been a cornerstone in the realm of theoretical research regarding complexity and algorithm design for the past 27 years.
This week, both OpenAI and Anthropic announced their findings that effectively refute the ETH. The established notion that the effort required to solve certain problems grows exponentially with the size of the problem has been challenged, marking a pivotal shift in the academic landscape. Jonsson anticipates that many attendees at the IEEE Symposium on Foundations of Computer Science, where one of his co-authors will present their work, will start their talks with statements like, “the ETH is false, half of my paper does not hold.” While this situation is undoubtedly frustrating, he acknowledges the valuable lessons that can be gleaned from such developments, stating, “that is what science is all about.”
Collaborative Efforts to Understand New Findings
Researchers have yet to independently verify, let alone fully comprehend, the proofs generated by AI, which have been made publicly available on GitHub in both natural language and machine-verifiable formats. OpenAI has reached out to the original proposers of the hypothesis from the University of California, Santa Barbara, alongside a scholar from MIT who has previously challenged a version of the ETH, encouraging collaboration to analyse these proofs. Currently, this initiative includes a Google form for volunteers and feedback, as well as OpenAI’s commitment to fund a conference focused on this topic.
Professor Daniel Lokshtanov from UC Santa Barbara intends to participate, acknowledging that, if confirmed, the outcome would render him “completely wrong” and necessitate revisions to approximately 50 of his published papers. He likens the situation to a massive disruption in the field, describing it as both exhilarating and daunting.
Excitement Amidst Uncertainty
Thore Husfeldt, head of the algorithms research group at the IT University of Copenhagen, has spent much of the year contemplating a singular problem within his subfield. “From what I see, everything is now irrelevant,” he remarked. Having initially formulated a version of the hypothesis that OpenAI has now refuted, he has been calculating on his whiteboard to assess what other aspects of his work may also be rendered obsolete due to this “once-in-a-generation result.” Unlike the disillusionment experienced by some mathematicians in the face of similarly revolutionary AI-generated proofs, Husfeldt expressed his excitement about the new findings.
He describes the discovery as “fantastic,” explaining that being a mathematician feels akin to navigating a completely dark room. “It’s as if someone has illuminated an entire subfield with a beam of knowledge, and now we are like excited rabbits exploring all the corners that we can suddenly uncover. It is a profoundly exhilarating moment.”
The Psychological Impact on Researchers
Science has historically progressed through the formulation of hypotheses and their subsequent falsification. It is not uncommon for one academic’s work to be overshadowed by another’s findings. However, the rapid pace of AI advancements has intensified this process, obliterating entire research agendas and fostering anxiety among researchers in specific mathematical and scientific domains, as they grapple with their relevance in an era where machines can outperform them.
“The past few months have been particularly challenging. There is a pervasive sense of pessimism, as results announced by AI seem to annihilate ongoing research projects regularly,” Husfeldt remarked. He added that his department has even contemplated hiring an occupational psychologist due to the emotional toll, as “some individuals are profoundly depressed and despondent.”
Navigating the Flood of AI-Generated Content
Researchers are also overwhelmed by the sheer volume of content being produced by AI, which often includes results of questionable quality. During an interview, Husfeldt received a 17-page document generated by AI regarding the implications of the ETH result, which he deemed unworthy of his time to read, stating, “reading mathematics requires a significant cognitive investment.”
OpenAI has directly contributed to this influx, having published 722 manuscripts addressing 400 problems on GitHub earlier this month, much to the chagrin of academics who argue that it complicates the comprehension of findings. The frustration was compounded by numerous documents containing errors, which some mathematicians perceived as a rushed release intended to bolster the company’s IPO rather than to advance the field. However, the recent 60-page article by OpenAI on the ETH appears to be of better quality than prior publications, according to academic feedback.
Implications for Future Research
Should the Exponential Time Hypothesis be proven false, suggesting that resolving complex problems does not necessarily require exponentially more time, this would be an “exciting” turn of events, asserts Victor Lagerkvist, an associate professor at Linköping. He plans to revisit problems he had previously abandoned under the assumption that the ETH was correct.
The implications of OpenAI’s findings may also extend to the field of cryptography, which underpins web security based on the assumption that certain mathematical problems demand excessive time to solve. Even though this particular result does not directly impact cryptography, Lokshtanov cautions that “it should certainly concern some cryptographic system designers.”
The Evolving Landscape of Theoretical Computer Science
Consequently, it is vital to continue advancing theoretical computer science, leveraging both human and artificial intelligence, Husfeldt emphasises: “We want the open scientific process to help identify our misunderstandings about computation, particularly since much of the modern digital infrastructure is based on assumptions regarding computation.”
As for Jonsson from Linköping, he plans to reassess both his own work and that of others reliant on the ETH’s validity, believing that his conference presentation might still be salvaged through alternative methods. He is eager to delve into OpenAI’s proof and explore what insights he can glean and apply to new challenges. Although this development invalidates the foundational premises of his research, he considers obtaining a verdict on the ETH something he could only have previously “dreamed” of. “These are thrilling times,” he concludes. “Most people will never experience anything like this in their careers.”
“I don’t care whether it’s a proof written by humans, by AI, or if it was transmitted to us by an extraterrestrial. If it’s true, we simply need to use it moving forward.”