In a week when crypto markets were already on edge, OpenAI dropped a headline that sent a very specific kind of panic through Crypto Twitter: “OpenAI says its AI has proven that ETH is false.” If you are a person who spends any time watching Ethereum price charts, your heart probably skipped a beat. Did an artificial intelligence just somehow invalidate the second-largest cryptocurrency? Did Vitalik Buterin get caught off guard? The jokes came fast and furious. One user named Takumi posted, “Damn algo traders just slaughtered $ETH,” while Andrew Jiang quipped, “Vitalik has been real quiet since this came out.” It was a perfect moment of internet chaos, except for one tiny problem: nobody was talking about Ethereum at all. The ETH in question is the Exponential Time Hypothesis, a computer science conjecture from 1999 that has absolutely nothing to do with decentralized finance, smart contracts, gas fees, or Buterin. It is an abstract mathematical claim about the limits of computation, and OpenAI’s announcement was about an algorithm, not an asset. So, to all the degens who were frantically checking their portfolios, you can breathe. Your coins are fine. The real story is far nerdier, and in some ways far more exciting, than a crypto crash.
The Exponential Time Hypothesis, or ETH for short, is the kind of thing that only a theoretical computer scientist could love, but it matters a great deal behind the scenes. It was proposed in 1999 by computer scientists Russell Impagliazzo and Ramamohan Paturi, along with Francis Zane, and it makes a very specific bet about how hard a classic logic puzzle called 3-SAT really is. Imagine you are given a giant list of rules, each rule involving three true-or-false options, and you have to find a setting of true and false values that satisfies every single rule at the same time. That is 3-SAT. The obvious way to solve it is to try every possible combination of true and false values, which becomes impossibly slow as the number of variables grows. ETH says that there is no dramatically faster way to do this—no clever shortcut that lets you solve every instance of 3-SAT in what mathematicians call “subexponential time,” meaning time that grows more slowly than any exponential function of the number of variables. In other words, ETH is a map of what computers will probably never be able to do quickly, no matter how much hardware you throw at it. Hundreds of papers in graph theory, computational biology, databases, and geometry rely on ETH as an assumption. When researchers want to show that some problem cannot be solved quickly, they often prove that if it could, then 3-SAT could also be solved quickly, which would violate ETH. So ETH is not just an obscure conjecture; it is a load-bearing wall in a large part of theoretical computer science.
OpenAI’s claim, if it holds up, is that an AI model just knocked that wall down. The company published a 67-page paper describing a deterministic algorithm—one that involves no randomness—that counts every satisfying assignment of 3-SAT instances in subexponential time. That is an even harder job than merely deciding whether a satisfying assignment exists, because counting all of the solutions is strictly more demanding than finding one. If you can count them, you can certainly tell whether the number is zero or nonzero, so the easier problem comes along for free. OpenAI says its AI discovered the disproof on October 16 and then, on October 17, the result was formalized in Lean, a proof assistant software that mechanically verifies every logical step. Both the paper and the Lean formalization were posted on GitHub, in the same way OpenAI released 722 math manuscripts earlier this month, of which only 162 came with a computer-checked main result. The company’s blog post framed the achievement in bold terms: “The refutation of the ETH reopens questions about how much faster these tasks could become,” it said, “and may help guide the search for better algorithms.” If true, it would mean that a whole family of problems that researchers believed were permanently hard might actually be solvable much faster than previously thought. That would not just be a small tweak; it would be a shift in the landscape of complexity theory, the branch of computer science that studies which problems are easy, which are hard, and why.
But before anyone starts rewriting the textbooks, it is worth noting that the very people who invented ETH are not exactly popping champagne. Impagliazzo, Paturi, and MIT’s Ryan Williams are leading the independent examination of OpenAI’s claim, and they have been extremely careful about what their involvement means. In an open letter, the three researchers wrote, “Our involvement should not be taken as certification of the claim,” adding that “a result of this significance requires careful, independent scrutiny and an understanding that others can reconstruct, question, and build upon.” That is mathematician-speak for: we are looking at it, but you should not assume it is true just because we are in the room. OpenAI says the participants signed no non-disclosure agreements, are working unpaid, and that the company has agreed to publish all raw data and fund a conference on the result. That is all good, but the scientific process takes time. Practical payoffs, even if the result is correct, are far off. As OpenAI itself wrote, “Finding such algorithms, and turning theoretical advances into practical improvements, will require significant further research.” The paper also explicitly says that the result does not show that every NP problem can be solved in subexponential time, so the most famous open question in computer science—P versus NP—remains completely open. For the uninitiated, P is the class of problems that computers can solve efficiently, while NP is the class of problems whose proposed solutions can be checked efficiently. The question of whether every problem with an easy-to-check answer is also easy to solve has haunted computer science for half a century. OpenAI’s announcement, stunning as it may be, does not resolve that. It only chisels away at one particular corner of the broader puzzle.
Of course, OpenAI is not the only AI lab playing this game. Anthropic, not wanting to be left out of the math wars, posted a paper credited to Claude with the wonderfully technical title “A 2^o(n) Time Algorithm for 3-SAT.” The paper, dated October 26, also tackles the Exponential Time Hypothesis head-on. According to the repository, an internal research Claude model discovered the algorithm, and the authors are quick to add that “the result is purely asymptotic and says nothing about solving formulas in practice.” In plain language, that means the algorithm looks good on paper when you are talking about infinitely large problems, but it does not mean your laptop is about to become an oracle. This is a recurring pattern in the world of AI-generated mathematics: the results are theoretical, not practical, and they are more about pushing the boundaries of what can be proven than about shipping useful software. Anthropic has been on a tear with formal mathematics too. In September, Claude formalized Fermat’s Last Theorem in Lean, a historic achievement in the world of proof assistants. And the mathematician Terence Tao, who has been watching this unfold with a mix of wonder and wariness, saw this kind of convergence coming. In September, he wrote, “We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential.” That is a remarkably astute observation, and it is exactly what seems to have happened here: the moment OpenAI’s claim about ETH became public, other AI systems and research teams scrambled to produce their own results, eager to share a piece of the spotlight.
What comes next is a story about money, trust, and the slow, careful work of verification. OpenAI has made a recent $5 million contribution to the Simons Foundation, a major supporter of mathematics and theoretical computer science, to fund workshops, conferences, and special programs that will help researchers examine and build on AI-produced results. That is a meaningful gesture, because the bottleneck in AI mathematics is no longer generating claims; it is figuring out which claims are actually true. The old model of scientific discovery, where a lone genius scribbles equations in a notebook and slowly convinces the world, is being replaced by something faster, messier, and more collaborative. AI models can produce proofs and algorithms at a pace that human mathematicians cannot match, but those outputs are only as good as the verification methods behind them. That is why Lean and other proof assistants are so important: they provide a way to mechanically check every step, removing the possibility of subtle human error. But even a machine-checked proof can rest on assumptions that are wrong, or use definitions that are subtly different from what the wider community expects, which is why human researchers like Impagliazzo, Paturi, and Williams still have a vital role to play. The conference OpenAI has agreed to fund will bring together the people best equipped to stress-test this result, and the open-source publication of raw data means that anyone with the right expertise can dig in. For the rest of us, the takeaway is simple: this is an exciting moment, but not a settled one. The Exponential Time Hypothesis may be wounded, or it may survive; either way, the process of finding out will itself be a contribution to science. And for the crypto crowd, the message is even simpler. Your Ethereum is safe, the price of ETH is unaffected, and the only thing being refuted is a 1999 conjecture about logic puzzles. The real drama is happening in the world of algorithms and proof assistants, far away from the charts you are refreshing, and it is going to be fascinating to watch.