OpenAI and Anthropic Claim AI Models Refute the Exponential Time Hypothesis

OpenAI and Anthropic announced that their latest language‑model‑driven systems have produced algorithms that solve a class of problems previously thought to require exponential time. The claim, if verified, would overturn the Exponential Time Hypothesis (ETH) – a cornerstone of computational complexity theory – and force researchers to rethink many results that rely on it.

---

What did the frontier labs actually release?

According to the TechCrunch report, OpenAI disclosed three breakthrough solutions on GitHub, each targeting a different discipline: mathematics, computer science, and physics. In the computer‑science branch, the lab presented an algorithm that allegedly solves instances of the ETH‑defined logic puzzles in sub‑exponential time. Anthropic quickly followed with its own independent refutation, suggesting the discovery was not an isolated fluke. Both labs framed the releases as collaborative, inviting the broader scientific community to examine the proofs, reproduce the results, and discuss norms for AI‑generated research.

---

What is the Exponential Time Hypothesis?

The ETH, first formalized by Impagliazzo and Paturi in 1999, posits that certain families of constraint‑satisfaction problems cannot be solved faster than in time (2^{\Omega(n)}), where (n) is the number of variables. It is a stronger statement than the famous P vs NP question, providing a concrete lower bound for a wide range of NP‑hard problems. Over the past two decades, ETH has been used as a benchmark for algorithmic optimality; many papers cite it to argue that their proposed improvements are essentially the best possible.

---

Why does refuting ETH matter?

If a genuine sub‑exponential algorithm exists for the problem class covered by ETH, several downstream assumptions crumble:

• Algorithmic optimality – Researchers who have spent years proving "optimal up to ETH" will need to revisit their proofs.
• Cryptographic security – Many cryptographic schemes rely on the belief that certain problems are intractable for any algorithm running in less than exponential time. A faster algorithm could shrink the security margin of key sizes.
• Complexity theory – The relationship between P, NP, and ETH would need to be re‑examined, potentially nudging the community closer to resolving the P = NP question.

MIT professor Ryan Williams, a leading expert on computational complexity, cautioned that while the discovery is exciting, it is unlikely to be the final word on the problem. He noted that the AI‑generated algorithm may not be the absolute best possible and that further scrutiny is essential.

---

How are researchers reacting?

The response has been a mix of enthusiasm and skepticism. Williams, along with UC San Diego colleagues Russell Impagliazzo and Ramamohan Paturi – the original authors of ETH – organized an open letter calling for community feedback and a discussion of norms around AI‑produced proofs. They also asked OpenAI to fund a dedicated conference, arguing that the potential impact on decades of literature warrants a coordinated effort.

Anthropic’s mathematician Levent Alpöge suggested that OpenAI may have built on existing, unpublished work, a claim that underscores the need for transparent provenance. Both labs have pledged to keep their proofs publicly accessible, but the academic community remains wary of “proof‑dump” tactics that can overwhelm reviewers, a concern Williams previously described as a "DDOS attack on mathematicians."

---

What could this mean for security and practice?

Williams raised a concrete security question: whether the new algorithm compromises any current cryptographic schemes. OpenAI’s science lead, Dan Roberts, reportedly checked the impact and expressed confidence that no widely deployed systems are immediately broken. Nonetheless, the episode has sparked renewed interest in post‑quantum cryptography and the need to future‑proof protocols against algorithmic advances that may arise from AI‑assisted research.

Beyond cryptography, developers may soon see AI‑generated optimizations for hard combinatorial problems in logistics, scheduling, and verification. If the sub‑exponential techniques can be generalized, they could lead to faster solvers for real‑world NP‑hard tasks, reshaping industries that rely on exact optimization.

---

How can Grafy users explore this development?

Grafy offers several ways to dig deeper into the ETH debate without leaving the platform:

• Reader – Upload the PDF versions of the OpenAI and Anthropic proofs and use the AI‑powered search to locate the "alien‑looking" sections Williams mentioned. The tool can answer specific questions about the algorithmic steps, making the dense technical material more approachable.
• Chat – Engage with multiple large‑language models side‑by‑side to compare explanations of ETH, its implications, and the new sub‑exponential claims. Prompt the models to generate analogies or to outline a simple example of an ETH‑type problem.
• Graph – Build a branching diagram that tracks the logical flow from the original ETH statement, through the AI‑generated algorithm, to the potential security impacts. This visual map can serve as a quick reference for team discussions.

While Grafy does not yet host a dedicated "complexity‑theory" sandbox, the combination of Reader, Chat, and Graph gives creators, developers, and researchers a practical workflow to examine, annotate, and share insights on the emerging proofs. For those interested in reproducing the algorithms, the Code studio can be used to prototype the sub‑exponential solver in Python, leveraging the same cloud models that powered the original discovery.

---

Where can I find the original announcement?

The full technical details and the surrounding commentary are available in the TechCrunch article titled "Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?" – https://techcrunch.com/2026/10/27/frontier-ai-labs-say-their-models-refuted-the-exponential-time-hypothesis-what-does-that-mean/.

---

Frequently asked questions

What exactly does "sub‑exponential" mean in this context?

Sub‑exponential time refers to algorithms whose running time grows slower than (2^{cn}) for any constant (c>0). In practice, it means that as the problem size (n) increases, the solver scales more gently than the exponential curve predicted by ETH, potentially enabling solutions for larger instances than previously feasible.

Should I worry that my encrypted data is now insecure?

OpenAI’s internal review, as reported by Williams, suggests that no widely used cryptographic protocols are immediately compromised. However, the discovery encourages a review of security margins and may accelerate the adoption of post‑quantum or otherwise hardened schemes.

How can I verify the AI‑generated proofs myself?

Upload the proof PDFs into Grafy Reader, then ask the assistant to walk through each lemma and its justification. You can also export the annotated document to Code to experiment with the algorithmic implementation, comparing runtime results against known benchmarks.

---

Source: TechCrunch — AI — Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?