OpenAI & Anthropic Refute ETH: A Breakthrough in AI

Frontier AI laboratories have reached a unprecedented milestone in theoretical computer science. OpenAI and Anthropic have announced that their advanced AI models generated proofs refuting the Exponential Time Hypothesis (ETH)—a foundational assumption that has dictated computational complexity theory for decades.

In a major development for artificial intelligence and high-level mathematics, OpenAI published solutions to three long-standing mathematical and scientific challenges. Among these findings was a surprise rejection of the Exponential Time Hypothesis (ETH). The release prompted intense discussion across academic circles, followed quickly by a parallel announcement from rival lab Anthropic, sparking both excitement and scrutiny regarding how AI systems generate complex theoretical breakthroughs.

Dual Refutations and Evolving AI Lab Disclosures

The academic response to OpenAI's proof drop reflected a combination of awe and skepticism. Following OpenAI's announcement, Anthropic mathematician Levent Alpöge suggested that OpenAI's proof appeared to draw heavily on pre-existing research. Just hours later, Anthropic released its own distinct refutation of the Exponential Time Hypothesis, demonstrating that multiple frontier models are converging on complex computational proofs simultaneously.

Addressing past criticisms over how AI proofs are shared with academia, OpenAI altered its publishing strategy across the disciplines involved:

  • Mathematics: The proof materials were uploaded directly to GitHub for public review.
  • Computer Science: A formal working group was established, requesting public proof releases and urging OpenAI to fund an academic conference to evaluate the findings.
  • Physics: A specialized working group was formed to determine the appropriate framework for releasing their specific results.
"Some parts are, maybe you could say routine, but then yes, there are two or three alien-looking parts, and it’s exactly these parts that I saw that made me think, wow, this is something that the whole community needs to work on together." — Ryan Williams, MIT Professor

Deconstructing ETH and the P vs NP Question

To understand the magnitude of this breakthrough, one must examine its relationship to the famous P versus NP problem. At its core, P vs NP asks whether every problem whose answer can be easily verified by a computer (NP) can also be easily solved by a computer (P).

A practical illustration is the game of Sudoku: verifying that a completed grid is correct takes seconds, whereas solving an empty grid from scratch is far more labor-intensive. Modern digital infrastructure and cryptographic algorithms rely heavily on this exact asymmetric principle—encoded data can be easily decrypted with a key, but cracking the code without one takes an impractical amount of computational time.

Key Complexity Principles at a Glance

P vs NP: Asks if problems with fast-verification solutions (NP) also possess fast-solving algorithms (P).

Exponential Time Hypothesis (ETH): Asserts that the fastest algorithms for tricky logic problems inherently require exponential processing time as variable counts grow.

Sub-Exponential Algorithms: Novel procedures claimed by OpenAI and Anthropic models that solve these hard logic puzzles faster than previously thought theoretically possible.

Historically, computer scientists operated under the assumption that P ≠ NP, supported by the Exponential Time Hypothesis formulated in 1999 by UC San Diego computer scientists Russell Impagliazzo and Ramamohan Paturi. ETH posited that hard logic problems fundamentally require exponential time to compute as variables scale. However, the new sub-exponential algorithms generated by OpenAI and Anthropic models challenge this long-held intuition, suggesting that these complex problem classes may be far more solvable than previously believed.

Academic Collaboration, Cybersecurity, and the Future of Research

Organizing the computer science community's response is MIT Professor Ryan Williams, who previously referred to an earlier OpenAI proof drop as a "DDOS attack on mathematicians." Recognizing the fundamental importance of the discovery, Williams joined forces with ETH co-creators Impagliazzo and Paturi to help interpret the AI-generated solutions at OpenAI's request.

The collaborative group initiated an open letter seeking feedback from the broader scientific community to evaluate the proofs and establish guidelines for AI-assisted theoretical discoveries. Beyond pure theoretical mathematics, the refutation of ETH carries significant practical implications:

  1. Revisiting Decade-Old Literature: Countless computer science papers written over the last ten years rely on ETH as a foundational premise and will now need systematic re-evaluation.
  2. Cybersecurity Safeguards: Because faster computational algorithms can threaten cryptographic standards, Williams confirmed with OpenAI science lead Dan Roberts that security audits were conducted. OpenAI expressed confidence that existing cryptographic systems remain secure for now.
  3. Redefining Scientific Inquiry: As routine theoretical proofs become automated by AI systems, researchers are prompted to elevate the complexity of the questions they pursue.

While Williams notes that the current AI-generated algorithm is unlikely to be the absolute optimal solution, the breakthrough fundamentally reframes theoretical computer science, proving that frontier AI models are transitioning from basic coding tools to active drivers of core scientific discovery.