OpenAI’s Groundbreaking Math Breakthroughs: Upending Decades-Old Computational Complexity Assumptions

OpenAI’s Groundbreaking Math Breakthroughs: Upending Decades-Old Computational Complexity Assumptions

OpenAI recently published full proposed solutions to three long-unresolved, high-stakes math problems that have stymied researchers for decades. The most seismic of these results is a formal, highly unanticipated refutation of the Exponential Time Hypothesis (ETH) — a decades-old foundational computer science assumption that forms a core pillar of how researchers define what types of problems are functionally impossible to compute efficiently.

As has been the case with previous unannounced cutting-edge proof releases from the lab, the global scientific community reacted with a chaotic mix of unbridled excitement and sharp frustration. Levent Alpöge, a mathematician at competing AI lab Anthropic, pointed out that OpenAI’s results appeared to build heavily on pre-existing published work, and just hours after that public comment, Anthropic released its own independent, separate refutation of ETH.

A Departure From Past "Proof Dump" Playbooks: Structured, Field-Specific Releases

Unlike OpenAI’s earlier unceremonious, one-size-fits-all drops of high-stakes proofs, the lab rolled out the three new results with tailored, field-specific structures for each relevant research community:

Mathematicians received full access to the corresponding work via a public GitHub release

Computer scientists organized a dedicated working group, negotiated to have the proofs shared fully publicly, and secured a commitment from OpenAI to fund a dedicated peer conference for the field to discuss and validate the results

Physicists formed their own independent working group, which will set the terms for how their corresponding breakthrough result is formally released to the public.

MIT’s Ryan Williams: Steering the Field Through a Paradigm-Shifting Moment

Ryan Williams, a leading MIT computer scientist whose research centers on computational complexity, is leading the computer science community’s coordinated response to the new findings. Williams previously publicly criticized OpenAI’s workflows, famously calling a mass unstructured proof dump the lab published earlier this October a “DDoS attack on mathematicians.” But he told TechCrunch he agreed to collaborate with the AI frontier labs to steward these field-altering ideas responsibly, rather than letting them create unmanaged chaos.

OpenAI tapped Williams to help interpret the ETH solution, alongside UC San Diego computer scientists Russell Impagliazzo and Ramamohan Paturi — the pair who first formally published the Exponential Time Hypothesis back in 1999. The trio’s first public action was releasing an open letter requesting broad feedback from relevant scientists, and launching a public conversation around new community norms for properly integrating AI-generated contributions to mathematical and computational research.

“Some parts [of OpenAI’s proof] are, you could say, routine, but then there are two or three alien-looking parts — and it’s exactly those sections that made me think, wow, this is something the whole community needs to work through together,” Williams shared with TechCrunch.

Context: How The ETH Refutation Connects To The Legendary P vs. NP Problem

The Exponential Time Hypothesis sits at the center of the decades-old P vs. NP problem, a long-unresolved question that has shaped computer science and adjacent fields for generations. Put simply, researchers have spent decades trying to confirm whether every problem whose solution can be easily verified (the NP class) can also be easily solved (the P class).

The classic lay example of this dynamic is Sudoku: once a full solution to a puzzle is filled in, you can verify its correctness in seconds, but deriving that full solution from a blank grid is extremely time-consuming. This dynamic underpins all modern public cryptography: when you hold a decryption key, you can instantly unlock a scrambled message, but cracking a properly designed cryptographic system from scratch is so computationally expensive that it is functionally useless for attackers.

But researchers have never formally proven that this dynamic is universally true.

“P = NP would mean that not only can you verify Sudoku solutions when they exist, you can just generate them out of thin air,” Williams explained. “P = NP asserts this is true for every single problem you can verify quickly: if you can check an answer fast, you can find an answer fast. That sounds far too good to be true.”

Virtually the entire field has long operated under the assumption that P ≠ NP, and ETH was widely accepted as strong formal supporting evidence for that position. The hypothesis broadly states that even the fastest possible algorithm will require exponentially more compute time to reach a solution for a specific class of complex logic puzzles, as the number of variables in the system grows.

OpenAI and Anthropic’s shared claims assert that their AI models derived new algorithms that solve these exact logic puzzles in sub-exponential time — meaning far faster than the theoretical speed limit ETH was supposed to enforce. If the proofs hold up to peer review, they will formally disprove ETH, and potentially open a viable path to a full proof that P = NP.

“Every computer scientist should probably assign a little more weight to the plausibility of P = NP being true, though individual confidence levels will vary,” Williams noted. “This definitely calls into question a lot of the intuitions we thought we had about how tractable hard NP problems really are. I also don’t think this is the final answer, by any stretch — I don’t believe OpenAI’s algorithm could possibly be the very best algorithm it’s possible to develop for these problems.”

Ripple Effects For The Field And Global Cybersecurity

ETH has been cited as a core foundational assumption in thousands of published papers across the last decade. If OpenAI and Anthropic’s claims hold up, all of that existing published work will need to be re-examined to adjust for the new reality.

“That’s part of why I pushed OpenAI so hard to support a dedicated conference on this work,” Williams said. “This is going to impact the research of thousands of people, and after we work to understand the algorithms themselves, we need to figure out what comes next: how do we rebuild the core foundations of our field, compared to what we operated with before?”

Before addressing purely academic questions, Williams prioritized urgent cybersecurity risks: the very first question he asked OpenAI’s research lead Dan Roberts was whether the new sub-exponential algorithms break any widely deployed modern cryptographic systems. OpenAI ran full internal audits, and confirmed they are reasonably confident the current results do not create any practical, immediate cryptographic vulnerabilities.

AI-Driven Scientific Discovery: A Work In Progress

This new era of AI-generated, field-shaking scientific breakthroughs is still very much an unregulated, unrefined process, but Williams says he remains cautiously optimistic, even with major outstanding concerns — particularly around training the next generation of scientists. As a computational complexity researcher who regularly works with high-level formal math, he notes the transition to this new AI-augmented paradigm has been far easier for him than for many pure mathematicians.

He argues AI is ultimately pushing researchers to raise the bar for their own work: “AI is forcing people to be much more ambitious about the kind of work they pursue, and think deeply about this: if all the problems I spent years working on used to be my core research focus, but are now low-hanging fruit for AI to solve, how do I level up the scientific questions I choose to go after?”