AI Just Disproved the Exponential Time Hypothesis. Here's Why It Matters

OpenAI announced that an AI system produced a disproof of the exponential time hypothesis (ETH), the 1999 conjecture formalized by Russell Impagliazzo and Ramamohan Paturi, with the proof built and checked in about 10 days using the Lean formal verification tool. OpenAI has brought in complexity theorist Ryan Williams along with Impagliazzo and Paturi to evaluate the claim, which, if it holds, would force a rethink of the many complexity-theory results built on the assumption that ETH is true. OpenAI president Greg Brockman noted the result does not mean P equals NP, and researchers have reacted with a mix of excitement and caution, with some calling it possibly the biggest surprise in twenty years of complexity theory.

AI Just Disproved the Exponential Time Hypothesis. Here's Why It Matters OpenAI says its AI disproved a 27-year-old computer science belief, the exponential time hypothesis, in 10 days. Here's what that means. What is the exponential time hypothesis? The exponential time hypothesis ETH is a long-standing assumption in theoretical computer science that says certain hard problems, like the classic satisfiability problem, cannot be solved meaningfully faster than brute force in the worst case. It was formalized in 1999 by computer scientists Russell Impagliazzo and Ramamohan Paturi. OpenAI says its AI system produced a proof that this hypothesis is false, and that the proof was built and checked within about 10 days. If the result holds up under scrutiny, it would overturn an assumption that hundreds of other papers in complexity theory have been built on top of for nearly three decades. TL;DR - OpenAI announced that an AI system produced a disproof of the exponential time hypothesis , a 1999 conjecture that hard combinatorial problems have no fundamentally fast solution. - The hypothesis underpins a large body of theoretical computer science, so if the result survives review, it forces a rethink of results that were built assuming ETH was true . - This does not mean P equals NP. OpenAI’s own president, Greg Brockman, noted that P still probably does not equal NP, and the new result is a narrower, technical finding about one specific hypothesis. - The proof was checked with Lean , a formal proof-verification tool that checks each logical step the way a compiler checks code, which adds confidence but is not the same as full peer review. - OpenAI has brought in Ryan Williams , a leading complexity theorist, along with Impagliazzo and Paturi themselves, the original authors of the hypothesis, to evaluate the claim. - Real-world problems like vehicle routing, scheduling, and chip layout are the kind of “padlock” problems ETH describes, so a genuine theoretical shortcut could eventually ripple into cheaper logistics and better hardware design, even if practical speedups take years to materialize. - Reaction from researchers has been a mix of excitement and caution. One described it as possibly the biggest surprise in twenty years of complexity theory , while others on forums like Hacker News flagged that these kinds of results can be fragile. Other agents start typing. Remy starts asking. Scoping, trade-offs, edge cases — the real work. Before a line of code. Why did computer scientists believe ETH for 27 years? Think of a padlock with a combination. A three-digit lock is easy to brute-force. Add more digits and the number of combinations doesn’t grow a little, it explodes multiplicatively. That explosion is what mathematicians call exponential growth, and it describes the worst-case behavior of a specific family of problems known as NP-hard problems, including the boolean satisfiability problem SAT . ETH formalized the belief that for the hardest versions of these problems, there is no clever algorithmic trick that avoids this explosion entirely. You can shave off constant factors and find marginal improvements, but you can’t escape the fundamental exponential blowup. Researchers accepted this so widely that it became a foundation: entire subfields of theoretical computer science built conditional results on top of it, essentially saying “if ETH is true, then this other problem must also be hard.” That’s what makes a disproof significant. It’s not just one answer changing. It’s the load-bearing wall underneath a large structure of other proofs. What exactly did the AI prove? According to OpenAI, its model found a method for solving the classic hard version of this problem family that beats the exponential explosion ETH says is unavoidable. This doesn’t mean the problem suddenly becomes fast or practical. It’s still slow in absolute terms. But it’s asymptotically faster than what the hypothesis claimed was the theoretical floor, and that gap between “what we thought was the floor” and “what’s actually possible” is the entire finding. The proof was reportedly verified using Lean, a formal verification tool that checks every logical step of a mathematical proof automatically, similar to how a compiler checks code for errors rather than just trusting that it looks right. That’s a meaningful layer of rigor, but it’s distinct from the human peer review process that complexity theory results typically go through, which involves experts stress-testing the proof’s assumptions and implications over months or years. OpenAI has reportedly reached out to Ryan Williams, a well-known complexity theorist, who is forming a review group that includes Impagliazzo and Paturi, the two researchers who originally proposed ETH in 1999. Having the hypothesis’s own authors examine a claimed disproof is about as credible a review process as this kind of result could get. Does this mean P equals NP? No. P versus NP is a separate and even more famous open question that asks whether every problem whose solution can be quickly checked can also be quickly solved. If P equaled NP, it would mean all of these “padlock” problems have fast solutions, with enormous implications for cryptography and computing broadly. Remy is new. The platform isn't. Remy is the latest expression of years of platform work. Not a hastily wrapped LLM. ETH is a stronger, more specific assumption than “P does not equal NP.” Disproving ETH doesn’t touch the P vs NP question. OpenAI’s own leadership has been explicit that P still probably does not equal NP. Researchers on forums like Hacker News have also pointed out that results near this boundary can be fragile, meaning a narrow technical disproof of ETH wouldn’t suddenly compromise internet encryption or the cryptographic systems that depend on certain problems staying hard. If you were worried this breaks online banking security, it doesn’t. Why should anyone outside theoretical computer science care? Because the family of problems ETH describes shows up constantly in ordinary infrastructure. Routing delivery trucks efficiently, building school or hospital schedules without conflicts, packing shipping containers, laying out circuits on computer chips, assigning gates and crews at airports, all of these are versions of the same underlying hard combinatorial problem. A genuine theoretical advance in how fast these problems can be solved doesn’t immediately change anything you’d notice. Nobody’s delivery app gets faster overnight. But theoretical breakthroughs are often the seed that, years later, turns into better algorithms embedded in the software running logistics networks, power grids, and chip design tools. The value shows up quietly, as slightly lower costs and slightly better efficiency, rather than as a dramatic headline. Is this result actually confirmed? Not yet, and that distinction matters. What exists right now is a claim from OpenAI, backed by a Lean-verified proof, being handed to independent experts for review. That’s a meaningfully strong starting position formal verification catches a lot of errors that slip past human readers but it is not the same as the result surviving months of scrutiny from the broader theoretical computer science community. Reaction so far has split into two camps. Some researchers, describing it as potentially one of the biggest complexity theory results in decades, are genuinely excited. Others are urging caution, noting that boundary-pushing results in this area have a history of containing subtle errors that only surface once enough experts have picked them apart. The fact that OpenAI looped in the original authors of ETH, rather than only using internal review, suggests an awareness that this claim needs serious outside validation before it’s treated as settled. Frequently Asked Questions What is the exponential time hypothesis in simple terms? It’s the belief, formalized in 1999, that certain hard computational problems like boolean satisfiability cannot be solved significantly faster than brute-force search in the worst case. It’s an assumption, not a proven law, but it’s been treated as reliable enough that many other research results depend on it being true. Did AI prove P equals NP? No. That is a separate, even harder open problem. OpenAI’s reported result concerns the narrower exponential time hypothesis, and OpenAI’s own president has stated that P still probably does not equal NP. How was the proof checked? OpenAI says the proof was verified using Lean, a formal proof-assistant tool that checks each step of a mathematical argument for logical validity, similar to how a compiler checks code. This adds rigor but is separate from the full human peer review process the broader research community still needs to conduct. Does this affect internet security or encryption? Not based on what’s been described. Cryptographic security relies on specific hard problems remaining hard, and researchers have been clear that this result does not touch P vs NP or undermine the assumptions behind current encryption methods. Why does this matter if I’m not a computer scientist? Remy doesn't build the plumbing. It inherits it. Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something. Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want. The class of problems involved shows up in everyday infrastructure: delivery routing, scheduling, chip design, and logistics. Theoretical advances in how these problems can be solved don’t change anything immediately, but they can eventually filter into better, cheaper algorithms used in real systems.