
TECH & AI
By TrendScope Daily Team · November 8, 2026
On October 16, research models built by two rival AI labs each produced something computer scientists had spent 27 years assuming was impossible: a proof chipping away at the Exponential Time Hypothesis, a foundational assumption that a huge class of hard problems can never be solved meaningfully faster than brute-force guessing. OpenAI’s internal model and Anthropic’s Claude arrived at working refutations independently, without talking to each other, on the very same day — and when both labs published their papers roughly a week and a half later, the overlap stunned a research community that has spent careers building on the assumption two chatbots had just quietly undermined.
What the Exponential Time Hypothesis Actually Says
Proposed in 1999, the Exponential Time Hypothesis (ETH) is a conjecture about 3-SAT, a classic puzzle where you flip a row of switches to satisfy a list of logical rules. Checking every combination doubles the work with each new switch, and ETH says no algorithm can ever fully escape that exponential blow-up, however clever. It isn’t an abstract curiosity: hundreds of results across theoretical computer science are “conditional,” holding only if ETH is true. Pull that foundation out, and a large wing of the field needs a second look.
That is what the two papers claim to do. OpenAI’s 67-page write-up, titled “The Exponential Time Hypothesis is False,” describes an algorithm whose running time grows roughly as n divided by the square root of log n — a share of the total variables that keeps shrinking as problems get bigger, eventually falling below any fixed floor ETH requires. Anthropic’s 49-page paper, filed under the deadpan author credit “Claude, Anthropic,” reaches a similar conclusion with its own deterministic 3-SAT algorithm, and also refutes the related Strong Exponential Time Hypothesis. Per a detailed account from OfficeChai, Claude’s core algorithm emerged from roughly 14 million tokens of output with no human steering the discovery itself; researchers only stepped in afterward to help tidy up the write-up.
Two Labs, One Discovery, Same Day
The timing is the part that has researchers most unsettled. Neither OpenAI nor Anthropic coordinated the attempt; each company’s model found a workable approach to the same 27-year-old problem on the same day, using different methods and different underlying systems. OpenAI published first, on October 26; Anthropic’s Claude followed just hours later, with its paper name-checking OpenAI’s preprint as a parallel result. Both proofs were formalized in Lean 4, a proof assistant that checks logical steps line by line — Anthropic’s formalization alone spans more than 1,050 Lean modules — but that machine-checked formality isn’t the same as a finished human review. A detailed explainer from Emergent Mind notes that no person is yet known to have fully checked either Lean proof line by line, even though automated checkers accepted both. That gap, between a machine verifying the logic and a human understanding why it’s true, is what makes this such an uncomfortable moment for a field built on peer review.
“I would have much preferred it to be discovered by a human who could explain their thought process.”
— Russell Impagliazzo, UC San Diego, co-proposer of the Exponential Time Hypothesis
Why Researchers Are Reacting So Differently
Reactions from the people whose careers are built on ETH range from vindicated to genuinely mournful. Ryan Williams of MIT, a long-time skeptic, told reporters he “did not fully believe the hypothesis” and “felt vindicated.” Ramamohan Paturi of UC San Diego, who co-proposed ETH alongside Impagliazzo in 1999, said his first reaction was simply surprise, noting the hypothesis “has shaped a great deal of research over the past three decades.” The three are now jointly organizing an independent community review of both labs’ claims — a process they stress is not the same as certifying the results correct.
Further down the research pipeline, the mood is closer to mild panic. Complexity theorist Daniel Lokshtanov reportedly called the result a “giant bomb” for the field, estimating that around 50 of his own papers could become pointless, while Thore Husfeldt called it “fantastic” even as he acknowledged it undercuts much of his own prior work. Claude’s own literature search, cited in Anthropic’s paper, turned up more than 800 existing papers whose results depend on ETH or a closely related hypothesis.
What This Doesn’t Mean
It’s worth being precise about what these results don’t change, because early coverage reportedly overstated both points. This is not a step toward proving P equals NP — the new algorithms are still far slower than any polynomial-time solution, just not quite as catastrophically slow as full brute force. Nor does it meaningfully threaten encryption: Williams has said OpenAI checked and was “reasonably confident” the result breaks no real cryptosystems, and Anthropic’s own paper is blunt that the technique “says nothing about solving formulas in practice.” The improvement only beats brute-force search once the number of variables climbs past roughly 10 to the power of 10,000 — far beyond anything that will ever run on real hardware. One preprint examining elliptic-curve cryptography found the new approach helps only at similarly impractical sizes.
This is also not the first AI-claimed scientific breakthrough to draw a skeptical second look. Weeks earlier, in September, Anthropic’s claim about a CRISPR-like gene-editing enzyme discovered with AI assistance met a similarly mixed reception, with outside scientists accusing the company of overstating its real-world significance before independent replication. The pattern emerging across the AI industry in 2026 looks consistent: frontier labs are increasingly comfortable publishing eye-catching scientific claims generated largely by their own models, while the slower work of independent human verification perpetually plays catch-up. Readers curious about the field this upends might find a complexity theory textbook useful background before the next round of claims lands.
What to Watch Next
The immediate test is verification. Impagliazzo, Paturi, and Williams’s review group is working through both papers line by line, and complexity blogger Lance Fortnow has already written that “it is safe to assume now that the exponential-time hypothesis is toast,” adding that other researchers have since posted simpler, independent proofs of their own — a reassuring sign that the result doesn’t depend entirely on trusting either company’s model. A dedicated conference on the implications is reportedly being organized, and researchers like Husfeldt and Lagerkvist are already said to be revisiting complexity problems they had shelved as hopeless under the old assumption. If the refutation survives scrutiny, the more interesting long-term story may not be the theorem itself but the method: two different AI systems, working alone, independently finding a path through a problem that had stumped human researchers for 27 years.
As an Amazon Associate, TrendScope Daily earns from qualifying purchases made through links in this article.
Sources: OfficeChai, Al Jazeera, and Emergent Mind.
Quick questions
What is the Exponential Time Hypothesis?
It’s a 1999 conjecture in computer science stating that 3-SAT, a classic logic puzzle, can never be solved by an algorithm whose running time avoids exponential growth as the problem gets bigger. Hundreds of other results in complexity theory assume it is true.
Does this mean AI solved P vs NP?
No. Researchers including MIT’s Ryan Williams have been clear that the new algorithms are still far slower than any polynomial-time solution; they just grow slightly less explosively than previously believed possible, which is a different and much narrower claim.
Could this break encryption?
No credible reports suggest so. The new approach only outperforms brute-force search once the number of variables exceeds roughly 10 to the power of 10,000, far beyond any real-world computation, and Williams has said OpenAI checked that the result breaks no existing cryptosystems.
Related on TrendScope Daily
- Anthropic Files for a Nasdaq IPO — and Its Own Paperwork Reveals a Revenue Explosion
- OpenAI’s New Image Model Makes ChatGPT Pictures Faster, Animated
- Claude Haiku 5.5 Just Got 90% Cheaper — and It’s Already Beating OpenAI’s Budget Model
- More Tech & AI stories: browse the Tech & AI section.