Yesterday, OpenAI unveiled solutions to three significant mathematical challenges that have puzzled academics for an extended period. Among these was an unexpected refutation of the Exponential Time Hypothesis (ETH), a long-standing postulate in computer science that significantly influences our comprehension of computational difficulty.
Consistent with prior proof disclosures from the leading research lab, scholarly communities reacted with both excitement and apprehension. Levent Alpöge, a mathematician at Anthropic, implied that OpenAI’s solutions might draw from previously established research. Within hours, Anthropic then published its own counter-argument to ETH.
Nonetheless, OpenAI’s approach is undergoing changes. On this occasion, the lab organized its announcements distinctly for each of the three participating fields: Mathematicians received the work via GitHub; computer scientists established a working group, requesting public sharing of the proofs and OpenAI’s sponsorship for a discussion conference; and physicists formed a working group to decide on the release method for their findings.
Professor Ryan Williams from MIT, a computer scientist, is coordinating his discipline’s reaction to this recent discovery. Williams, whose research primarily concerns computational complexity inquiries, has previously voiced criticism of OpenAI’s methodology, describing their proof release in early October as a “DDOS attack on mathematicians.”
Nevertheless, he consented to collaborate with the cutting-edge labs to assist in guiding crucial concepts within the field, as he informed TechCrunch. The lab enlisted his aid in interpreting the solution, alongside UC San Diego computer scientists Russell Impagliazzo and Ramamohan Paturi, who initially introduced the ETH in 1999.
As their initial action, they issued an open letter soliciting input from scientists engaged with these challenges and fostering a dialogue on the established practices concerning AI’s involvement in their research.
“Certain elements [of OpenAI’s proof] are, arguably, conventional, but then indeed, there are a couple of distinctly unfamiliar sections, and it was precisely these portions that led me to conclude that this is a matter requiring collaborative effort from the entire community,” Williams disclosed to TechCrunch.
Understanding P vs NP: A Core Computer Science Dilemma
The ETH belongs to a wider category of challenges referred to as P vs NP, a topic that has invigorated computer science and numerous other disciplines for many decades. Simply put, investigators aim to determine if any problem whose solution is readily confirmable (NP) can likewise be readily discovered (P).
Consider Sudoku as a straightforward illustration: After a puzzle’s solution is established, its correctness can be swiftly confirmed, yet arriving at that solution initially requires considerable time. This pattern applies to a multitude of significant problems and forms the basis of cryptographic systems: Possessing the key to an encrypted message allows for its decryption, but effectively breaking a robust code is too time-intensive to be practical.
Nonetheless, scholars have not yet definitively established if this holds universally. “P = NP implies that, beyond simply verifying existing Sudoku solutions, one could also directly create them,” Williams explained. “P = NP posits that this applies to every problem verifiable in short order — if quick verification is possible, then quick solving is also possible. Therefore, this proposition appears overly optimistic.”
Consequently, the majority of scientists presume that P ≠ NP. The ETH offered substantiation for this. Fundamentally, it states that even the most efficient computer algorithm will require an exponentially increasing amount of time to find a solution as the variable count expands within a particular category of complex logic puzzles.
OpenAI and Anthropic assert that their models have devised algorithms capable of solving these puzzles in “sub-exponential” time, indicating a speed exceeding prior expectations. If accurate, they might have disproven the ETH—or refuted it, scientifically speaking—potentially paving the way for algorithms that generally achieve P = NP.
“All computer scientists should likely assign slightly greater credibility to the possibility of P = NP, though individual perspectives may differ,” Williams commented. “I believe it somewhat challenges many of our preconceived notions regarding… the tractability of challenging NP problems. Furthermore, I suspect this is likely not the definitive solution. I don’t think OpenAI’s algorithm represents the absolute optimal algorithm achievable.”
Future Endeavors for Computer Scientists
The enduring nature of ETH has resulted in its incorporation into numerous scientific publications over the last ten years, all of which will require re-examination if the AI labs’ assertions of its disproof prove valid, Williams noted.
“This is partly why I advocated for and urged OpenAI to back a conference on this topic — it will impact the work of many researchers, and beyond comprehending the algorithms, we need to determine what comes next. We must ascertain how fundamental principles should adapt compared to our prior understanding?” he stated.
Prior to addressing scientific implications, cybersecurity concerns arise. Williams mentioned that an enhanced algorithm might suggest that cryptographic systems dependent on specific key sizes could be deemed compromised. Therefore, one of his initial inquiries to OpenAI science lead Dan Roberts “was whether they had verified if this impacts any crypto systems… is anything vulnerable now? And apparently, they confirmed, and were reasonably certain that this did not… indicate any vulnerabilities.”
This emerging AI model for scientific breakthroughs remains under development, yet Williams expresses optimism despite legitimate concerns, especially regarding the education of future scientists. His background as a computer scientist skilled in advanced mathematics may have facilitated this transition more for him than for pure mathematicians.
Nevertheless, he suggests AI is “compelling individuals to adopt a more ambitious approach to their work and to delve deeper into questions like, if all the tasks I previously focused on are now readily handled by AI, how can I elevate the caliber of scientific inquiries I pursue?”