IA · 28 October 2026 · 4 min read

From Complexity Theory to Physics: AI Breakthroughs Spark Clash Between Labs and Academia

In brief: OpenAI announced three major theoretical breakthroughs across mathematics, theoretical computer science, and physics—including the refutation of the long-standing Exponential Time Hypothesis (ETH)—quickly matched by a rival claim from Anthropic. To defuse criticism following earlier unverified preprint dumps, OpenAI recruited external academic 'shepherds' to vet the work. However, the move has reignited backlash from researchers who accuse private tech firms of gatekeeping scientific progress and demand open, formalized verification standards.

by Team Mocchi's

Less than a month after its mass dump of automated mathematical proofs stirred outrage across academia, the race to claim AI-driven scientific breakthroughs has reached a new boiling point. OpenAI announced three landmark discoveries: progress on partial differential equations linked to Kakeya conjectures, an existence proof in quantum field theory, and a refutation of the Exponential Time Hypothesis (ETH), a foundational pillar of theoretical computer science established in 1999.

Competition followed immediately: hours after the announcement, Anthropic mathematician Levent Alpöge questioned the originality of OpenAI's submission, before Anthropic published its own separate paper claiming to independently disprove the ETH assumption.

The stakes of computational complexity

The central point of contention is the Exponential Time Hypothesis. Formulated by Russell Impagliazzo and Ramamohan Paturi, ETH posits that foundational computational challenges, such as 3-variable Boolean satisfiability (3-SAT), inherently require exponential running time relative to input size. According to OpenAI's documentation shared on GitHub alongside verification scripts written in the Lean formal language, a new deterministic algorithm can enumerate satisfying assignments in sub-exponential time and polynomial space.

If independently confirmed, this finding would reshape the landscape of computational complexity, shifting the theoretical boundaries of computational tractability. As reported by TechCrunch, the claim has triggered equal parts fascination and skepticism across research departments, many of which remain cautious following a history of AI-generated proofs riddled with subtle inconsistencies.

The 'shepherds' model and the gatekeeping debate

Eager to avoid repeating early October's missteps—when an uncurated drop of hundreds of AI manuscripts was described by MIT researchers as an unmanageable «DDoS attack on mathematicians»—OpenAI attempted a restructured disclosure model. It appointed prominent external academics as «shepherds» to guide public dissemination, including Impagliazzo himself, MIT professor Ryan Williams, and Nobel laureate David Gross.

Yet the approach has invited fresh scrutiny. As detailed by The Verge, researchers are voicing deep frustration with what they view as corporate gatekeeping, where private AI entities unilaterally grant preferential access to chosen academics. Advisory panels such as AGMAI warned against appointing curated intermediaries, cautioning that borrowing academic prestige to validate proprietary releases distorts peer collaboration, while urging labs to deposit raw Lean-formalized proofs directly into open scholarly repositories.

Defining norms for synthetic science

Confronting these tensions, the working group led by Williams, Paturi, and Impagliazzo issued an open letter calling for international standards around AI-generated discoveries. The document addresses fundamental dilemmas: how credit should be apportioned between automated solvers and the researchers who spend months verifying, formalizing, and interpreting their outputs, and what rules must govern undisclosed access during active investigations.

As frontier systems increasingly automate hypothesis generation, scientific publishing faces a structural challenge: peer review architectures built for human pacing cannot absorb machine-speed outputs without robust formal verification methods.

Mocchi's take

The friction between frontier AI laboratories and the scientific community offers a vital lesson for businesses navigating rapid technological adoption: generating answers at computational scale creates zero real-world value without rigorous, verifiable guardrails. For software teams and enterprises embedding automated reasoning agents into core workflows, implementing formal verification is not an academic luxury but an operational safeguard against subtle hallucinations. Accelerating generation must always be paired with hardened validation systems that preserve human oversight over high-stakes outcomes.

All articles on the Mocchi's blog