Research

OpenAI and Anthropic claim AI-assisted proofs refute the Exponential Time Hypothesis

Frontier labs release solutions to long-standing computational complexity problems, sparking debate over verification and scientific norms.

Last verified:

AI Models Challenge a Cornerstone of Computer Science

OpenAI released mathematical proofs on October 27 claiming to refute the Exponential Time Hypothesis (ETH), a 25-year-old conjecture that has shaped researchers’ understanding of which computational problems are inherently hard to solve. According to TechCrunch AI, Anthropic released its own refutation within hours, signaling parallel—or derivative—findings across frontier labs. However, neither proof has undergone formal peer review, and the claims have triggered structured scrutiny from the academic community rather than immediate acceptance.

The Exponential Time Hypothesis and Its Role in Cryptography

The ETH, first published by UC San Diego computer scientists Russell Impagliazzo and Ramamohan Paturi in 1999, belongs to a broader class of problems known as P versus NP. Reductively, the P versus NP question asks whether problems whose solutions can be quickly verified can also be quickly solved. Sudoku exemplifies this dynamic: verifying a completed puzzle takes seconds, but finding the solution is computationally expensive. This asymmetry is not merely academic—it underpins modern cryptography. When you possess the decryption key, revealing the plaintext is fast, but without the key, cracking an encrypted message should be prohibitively slow. If P equals NP (meaning verification speed implies solution speed), cryptographic systems would collapse.

The ETH assumes that certain NP-complete problems cannot be solved faster than exponential time, even with optimal algorithms. This assumption has anchored decades of research into computational complexity and hardness assumptions. According to TechCrunch, OpenAI’s model produced arguments contradicting this assumption.

Staged Release and Community Governance

Rather than releasing all three proofs identically, OpenAI adopted discipline-specific dissemination strategies. Mathematics researchers received the work on GitHub immediately. Computer scientists—the primary audience for the ETH claim—formed a working group that requested public proof sharing and commitments to fund a conference for community discussion. Physicists organized separately to determine their own publication norms.

MIT Professor Ryan Williams is coordinating the computer-science response. According to TechCrunch, Williams has previously criticized OpenAI’s disclosure tactics, characterizing an earlier October proof dump as a “DDOS attack on mathematicians.” This time, however, he agreed to collaborate, collaborating alongside Russell Impagliazzo and Ramamohan Paturi—the co-authors of the original ETH paper itself. Williams told TechCrunch that OpenAI’s proof contains “routine” components alongside “two or three alien-looking parts,” suggesting genuinely novel insights requiring communal examination.

Anthropic’s Rapid Response and Attribution Questions

Anthropic mathematician Levent Alpöge raised concerns about OpenAI’s originality, suggesting the proofs drew from pre-existing work. Hours later, Anthropic released its own refutation. According to TechCrunch, this timing raises questions about whether Anthropic independently replicated the finding or whether it built on OpenAI’s disclosed ideas. The ambiguity highlights ongoing tensions within frontier labs over attribution, credit, and the norms governing AI-assisted mathematical discovery.

Why This Matters

If the proofs survive peer review, they would overturn a foundational assumption in computational complexity and force researchers to recalibrate hardness assumptions underlying cryptography and algorithm design. More immediately, the process itself signals how frontier labs are negotiating the gap between rapid AI discovery and scientific rigor—moving from unilateral proof dumps toward coordinated working groups and funding commitments. Teams managing cryptographic infrastructure or designing hardness-based protocols should monitor peer-review outcomes over the next 6–12 months; if the ETH truly falls, threat models may need revision. For the research community, the outcome will likely inform how AI contributions to fundamental mathematics are vetted and credited going forward.

Frequently Asked Questions

What is the Exponential Time Hypothesis and why does it matter?

The ETH is a foundational assumption in computer science stating that certain NP-complete problems cannot be solved faster than exponential time. It underpins the hardness assumptions behind cryptographic systems and computational complexity theory.

Did OpenAI actually prove the ETH is false?

OpenAI claims its model produced a refutation, but the proof has not yet undergone independent peer review. MIT Professor Ryan Williams noted the proof contains novel, non-routine components that require community scrutiny.

Why did Anthropic release its own refutation so quickly?

According to TechCrunch, Anthropic mathematician Levent Alpöge suggested OpenAI's work was based on pre-existing material. Anthropic's rapid follow-up release may reflect confidence in parallel findings or a response to OpenAI's move.

How does this differ from previous AI-assisted discoveries?

This time, OpenAI tailored its release process to each research discipline—mathematicians received a GitHub dump, while computer scientists and physicists formed working groups to govern how proofs are published and discussed.

#computational-complexity #mathematical-proof #frontier-labs #peer-review #ETH