OpenAI released solutions to three major mathematical problems on Tuesday, including a surprising rejection of the Exponential Time Hypothesis. This durable assumption in computer science has long shaped how researchers understand the limits of computation. The release triggered a mixed response from the scientific community, ranging from excitement to skepticism.
Collaborative Review of AI Proofs
Unlike previous releases where proofs were simply published online, OpenAI structured this release to engage specific scientific disciplines. Mathematicians received the work via GitHub, while computer scientists formed a working group to review the findings. OpenAI agreed to fund a conference to discuss the results, a move welcomed by MIT Professor Ryan Williams, who organizes the field’s response.
Williams, who previously criticized OpenAI’s earlier proof drops as a “DDOS attack on mathematicians,” agreed to collaborate to help interpret the new findings. He joined forces with UC San Diego computer scientists Russell Impagliazzo and Ramamohan Paturi, who originally published the Exponential Time Hypothesis in 1999. Their first step was to release an open letter seeking feedback from scientists and discussing norms around AI contributions to research.
Williams noted that while some parts of the proof appeared routine, other sections looked “alien.” He stated that these unique elements are exactly what require the entire community to work together to verify.
Implications for Computational Complexity
The Exponential Time Hypothesis is closely tied to the broader P versus NP problem, a question that has animated computer science for decades. The P versus NP problem asks whether every problem whose solution can be quickly verified can also be quickly solved. A common example is Sudoku. Verifying a completed puzzle is fast, but finding the solution from scratch takes significant time.
Most scientists assume that P does not equal NP. The Exponential Time Hypothesis provided evidence for this view by suggesting that even the fastest algorithms require exponentially more time to solve certain logic puzzles as variables increase. However, the claims from OpenAI and Anthropic suggest their models have found algorithms that solve these puzzles in sub-exponential time. If accurate, this would refute the hypothesis and potentially open a path toward solving P versus NP.
Williams remarked that every computer scientist should now weigh the plausibility of P equaling NP more heavily. He added that the findings call into question many intuitions about how tractable hard problems are, though he does not believe OpenAI’s algorithm is the final or best possible answer.
Cybersecurity and Future Research
The durability of the Exponential Time Hypothesis means it has been cited in many scientific papers over the past decade. If the AI labs’ claims hold up, those papers will need to be revisited. Williams emphasized the need for a conference to determine how the foundations of the field should change.
Beyond theoretical implications, there are immediate cybersecurity concerns. Improved algorithms could potentially break cryptographic systems that rely on specific key sizes. Williams asked OpenAI science lead Dan Roberts if the new algorithms compromised any existing security systems. Roberts confirmed that OpenAI checked and was reasonably confident that no current cryptographic systems were broken.
Williams remains optimistic about this new paradigm of scientific discovery, despite concerns about training the next generation of scientists. He believes AI is forcing researchers to be more ambitious and to raise the bar for the scientific questions they ask.
Source: TechCrunch
