AI labs challenge long-standing computer science math hypothesis

The challenge to computational limits

OpenAI recently released solutions to three major mathematical problems that have long stumped researchers. Among these was a surprising rejection of the Exponential Time Hypothesis, or ETH. This hypothesis is a durable assumption in computer science that helps researchers understand which types of problems are truly difficult to compute.

The ETH essentially claims that even the fastest possible computer algorithms need exponentially more time to find a solution as the number of variables in a specific class of logic puzzles grows. For decades, this has served as a foundation for how scientists categorize the limits of what computers can achieve in a reasonable timeframe.

However, OpenAI and Anthropic are now claiming their models have produced algorithms that solve these puzzles in "sub-exponential" time. This means the models found ways to solve the problems faster than what was previously thought possible, potentially proving the ETH to be false.

Understanding P vs NP

To understand why this matters, one must look at the broader class of problems known as P vs NP. This is a famous question asking if any problem whose solution can be easily verified (NP) can also be easily solved (P). A common example is a Sudoku puzzle, where checking a completed grid is easy, but finding the solution from scratch is time-consuming.

Most scientists assume that P does not equal NP because it seems too good to be true. If it were true, it would mean that any problem we can quickly check could also be quickly solved. The ETH provided evidence for this belief by suggesting that certain hard problems would always remain computationally expensive.

By refuting the ETH, the AI labs are suggesting that the gap between verifying and solving might be smaller than we thought. MIT Professor Ryan Williams noted that while the results might not be the final answer, they call into question the intuitions scientists have had about how tractable hard problems actually are.

Scientific review and collaboration

The response from the scientific community has been a mix of excitement and concern regarding the methodology. Anthropic mathematician Levent Alpöge suggested that OpenAI might be basing its solutions on pre-existing work, leading Anthropic to release its own refutation of the ETH shortly after.

OpenAI has structured its release differently depending on the field. Mathematicians received the work on GitHub, while computer scientists were asked to form a working group to discuss the proofs publicly. Physicists have also formed a working group to determine how their specific results will be shared with the wider community.

Professor Ryan Williams, who has criticized OpenAI's previous "proof dumps," has agreed to collaborate with the labs to help interpret the findings. He is working with the original authors of the ETH to call for feedback and establish norms for how AI contributions should be integrated into scientific research.

Security and future implications

One of the most immediate concerns regarding these new algorithms involves cybersecurity. Many modern cryptographic systems rely on the assumption that certain types of problems are too time-consuming to solve. If a new algorithm can solve these problems significantly faster, it could potentially render current encryption methods insecure.

OpenAI's science lead, Dan Roberts, stated that the lab checked whether these new algorithms affect any current crypto systems. According to the report, they were reasonably confident that the findings do not imply any immediate breaks in existing security protocols.

Despite this, the durability of the ETH means it has been included in many scientific papers over the last decade. If the AI labs' claims hold up, many researchers will need to revisit their work to understand what the new foundations of the field should be. Williams noted that the lab is even looking to fund a conference to help the community navigate these changes.

A new paradigm for discovery

This situation highlights a shift in how scientific discovery is happening. AI is now being used to tackle high-level math and logic problems that were previously the exclusive domain of human experts. This is forcing researchers to rethink what kind of work they should be doing as AI begins to handle more routine tasks.

Williams observes that AI is pushing people to be more ambitious about the types of questions they ask. If AI can solve the "low-hanging fruit" of complex math, human scientists may need to raise the bar for the types of scientific questions they prioritize in the future.

While there are concerns about how the next generation of scientists will be trained, there is also optimism about the potential for collaboration. The goal is to figure out how to use these AI-generated tools to push the boundaries of human knowledge while maintaining rigorous scientific standards.