AI Models Refute the Exponential Time Hypothesis: What Does It Mean?

In just a few days, a long-standing debate in computer science moved from research halls to the forefront of technical discussion: can AI models overturn a hypothesis that has served for more than two decades as a cornerstone of our understanding of the limits of computation? OpenAI announced that it had published solutions to three stubborn mathematical problems, the most prominent of which was a surprising refutation of what is known as the Exponential Time Hypothesis. And within hours of researchers raising objections to the methodology of the announcement, Anthropic published its own refutation of the same hypothesis.

What Is the Exponential Time Hypothesis?

Researchers Russell Impagliazzo and Ramamohan Paturi proposed this hypothesis in 1999, and it belongs to a broader family of questions known as the "P versus NP" problem. The essence of the hypothesis is that a certain class of difficult logical puzzles can only be solved by algorithms whose running time doubles exponentially as the number of variables increases — meaning that adding just a few variables can turn a computation from minutes into centuries.

To clarify the bigger picture: the P versus NP problem asks whether every problem whose solution is easy to verify is also easy to find. Sudoku is a simple example; verifying a completed grid is quick, while arriving at one is costly in time. This disparity is not an academic detail — it is the foundation on which modern encryption systems rest: whoever holds the key decrypts the message in an instant, while anyone without it needs an impractical amount of time to break it.

This is why most researchers lean toward believing that P ≠ NP, and the Exponential Time Hypothesis was among the strongest pieces of evidence supporting that belief.

What Exactly Are the Two Labs Claiming?

Both companies say their models arrived at algorithms that solve those puzzles in "sub-exponential" time — that is, faster than the limit previously thought impossible to surpass. If this proves correct, the hypothesis has been refuted, which opens the door — theoretically, not yet practically — to reassessing the possibility that P = NP.

MIT professor Ryan Williams, a specialist in computational complexity, was among the harshest critics of OpenAI's approach of publishing proofs all at once, describing an earlier announcement as closer to a "denial-of-service attack on mathematicians." Nevertheless, he agreed to collaborate with the two labs to lead the computer science community's response, with the participation of Impagliazzo and Paturi. Williams says some parts of the proof are "routine," but that it contains two or three spots that "look completely strange" — and these are precisely what call for collective effort to understand.

A Shift in How Results Are Published

Notably, OpenAI handled each discipline differently, marking an evolution in its methodology:

• Mathematics: publishing the results directly on GitHub.
• Computer science: forming a working group that demanded the proofs be published openly, along with funding for a dedicated conference to discuss them.
• Physics: a working group that will itself determine the mechanism for announcing the result.

An open letter was also issued inviting researchers to submit their feedback and to discuss the professional norms governing AI contributions to scientific research — a point no less important than the mathematical result itself.

Is Cryptography at Risk?

The first question Williams raised was not scientific but security-related: does the new algorithm affect existing encryption systems? According to his account, OpenAI confirmed that it had conducted the necessary review and was reasonably confident the result does not lead to breaking any encryption system currently in use. But the principle still stands: a fundamental improvement in this class of algorithms could render certain key lengths insecure in the future.

Broader Implications for Scientific Research

The Exponential Time Hypothesis has been used as a starting assumption in a large number of research papers over the past decade. If the refutation holds, these works will need review — not only at the level of their results but at the level of the foundations on which they were built. Williams believes every researcher should give the possibility of P = NP slightly more weight than before, while cautioning that what the labs published "is not the final answer" and that better algorithms will emerge later.

More importantly, this new pattern of "AI-assisted scientific discovery" confronts researchers with a blunt professional question: if much of what we are used to working on becomes trivial for these models, how do we raise the ceiling on the questions we ask? Answering that question may prove more valuable than any single proof.

✦ بقلم فريق دروب أيديا