AI models challenge a core assumption in computer science complexity theory

On October 27, 2026, OpenAI published solutions to three major mathematical problems, including a surprising refutation of the Exponential Time Hypothesis (ETH). This long-standing assumption in computer science defines the lower bound of time required to solve certain complex logical puzzles. The release triggered an immediate response from the scientific community, with Anthropic releasing its own counter-proof hours later.

What happened

OpenAI structured its release differently for each affected discipline. Mathematicians received the work via a GitHub drop, while physicists formed a working group to decide on their publication strategy. For computer scientists, OpenAI established a working group and agreed to fund a conference to discuss the findings publicly. This approach marked a shift from previous interactions, which MIT professor Ryan Williams had criticized as a "DDOS attack on mathematicians" due to the overwhelming volume of unstructured data.

Williams, who specializes in computational complexity, agreed to help interpret the results alongside UC San Diego researchers Russell Impagliazzo and Ramamohan Paturi. Impagliazzo and Paturi originally published the ETH in 1999. The group issued an open letter inviting feedback from the broader scientific community and calling for a discussion on norms for AI contributions to research. Anthropic mathematician Levent Alpöge initially suggested OpenAI relied on pre-existing work, but Anthropic subsequently released its own refutation of the hypothesis.

Williams noted that while parts of OpenAI’s proof appeared routine, other sections were "alien-looking." These unusual components convinced him that the entire community needed to collaborate on understanding the implications. The goal is not just to verify the algorithm, but to determine how foundational theories must adapt if the ETH is indeed false.

How it works

The Exponential Time Hypothesis is closely tied to the P versus NP problem, a central question in computer science. P refers to problems that can be solved quickly, while NP refers to problems where a solution can be verified quickly. A common analogy is Sudoku: verifying a completed grid is fast, but finding the solution from scratch is computationally expensive. Most scientists assume P does not equal NP, meaning some problems are inherently hard to solve even if easy to check.

ETH supports this assumption by stating that for a specific class of logic puzzles, the time required to find a solution grows exponentially as the number of variables increases. OpenAI and Anthropic claim their models discovered algorithms that solve these puzzles in "sub-exponential" time. This means the computational cost grows much slower than previously thought possible as the problem size expands. If valid, this refutes ETH and suggests that some problems considered intractable may be more solvable than anticipated.

Key details

• OpenAI released solutions to three major math problems on October 27, 2026, including a refutation of the Exponential Time Hypothesis.
• Anthropic released its own refutation of ETH hours after OpenAI’s initial drop, following comments from mathematician Levent Alpöge.
• MIT professor Ryan Williams, along with Russell Impagliazzo and Ramamohan Paturi, is leading a working group to interpret the findings.
• The ETH, published in 1999, posits that specific logic puzzles require exponentially more time to solve as variables increase.
• OpenAI agreed to fund a conference for computer scientists to discuss the implications, a change from previous unstructured releases.
• OpenAI stated they checked for cryptographic impacts and are confident no current systems are broken by these new algorithms.

Why it matters

For software engineers and security architects, the immediate concern is cryptography. Many encryption systems rely on the assumption that certain mathematical problems are too hard to solve in a reasonable timeframe. If algorithms can solve these problems in sub-exponential time, key sizes currently deemed secure might become vulnerable. However, OpenAI science lead Dan Roberts confirmed that the lab checked for immediate cryptographic breaks and found none. This suggests that while the theoretical landscape is shifting, practical security infrastructure remains intact for now.

Beyond security, this development challenges the intuitions developers hold about computational tractability. If AI can routinely find sub-exponential solutions to NP-hard problems, the definition of what is "computable" in practice may expand. This forces a reevaluation of algorithm design and optimization strategies. It also raises questions about the future of scientific discovery, as AI begins to outpace human intuition in identifying novel mathematical paths.

What you can do

• Monitor the open letter and working group updates led by Ryan Williams to track the community’s verification process.
• Review your cryptographic dependencies to ensure you are using recommended key sizes, even though no immediate break was reported.
• Engage with the upcoming conference funded by OpenAI to understand how foundational complexity assumptions may change.
• Re-evaluate internal assumptions about problem hardness in your system architecture, particularly for optimization tasks.
• Follow the public GitHub repositories where OpenAI released the mathematical proofs for peer review.
• Prepare for a shift in hiring and training, as AI tools may handle routine mathematical verification, requiring engineers to focus on higher-level problem formulation.