The discussion around an AI-generated refutation of the Exponential Time Hypothesis now includes a second laboratory’s research artifact. TechCrunch’s October 27 report says Anthropic released its own claimed refutation after OpenAI’s announcement. That is a substantive addition to the earlier story, not an independent verdict that either argument has been accepted.
The distinction is important: two public claims create more material to examine, but do not automatically provide two independent checks. Researchers still need to understand the statements, inspect the arguments and determine which consequences follow.
Anthropic releases a separate counting-algorithm claim
Anthropic’s public ETH repository attributes its result to an internal research Claude model. Its description claims a deterministic algorithm for counting satisfying assignments of Boolean formulas with clauses of at most three literals in subexponential time. It explicitly says the result is asymptotic and makes no claim about solving formulas practically.
The repository includes a paper and a Lean 4 formalisation. Its README directs readers first to a short statement file, then to the proofs and checking process. It stresses a crucial boundary: a machine can check a formal proof without establishing that the encoded statement faithfully represents the natural-language claim.
It also describes pinned dependencies and a verification process using an additional kernel. Those are the project’s descriptions of its checking arrangement. This article has not built the repository, rerun those checks or independently validated the mathematics.
The public release is nevertheless meaningful because it gives critics a more concrete object than a headline or an unpublished claim. Disagreements can concern an explicit theorem statement, a proof step or the interpretation of a bound, rather than only whether one trusts the organisation announcing it.
The two releases need statement-level comparison
OpenAI’s previously released formalisation scope document describes exact counting for fixed-width Boolean formulas under a specified encoding, with subexponential time and polynomial space. It also distinguishes the formalised bound from sharper bounds stated in an accompanying manuscript.
That is a reason to compare the objects carefully before declaring that the laboratories have proved precisely the same theorem in precisely the same way. Counting versus deciding, encoding assumptions, resource bounds and the correspondence between a manuscript and its formal statement can all be relevant to what has actually been established.
TechCrunch adds new comments from MIT computer scientist Ryan Williams, who is helping organise scrutiny with Russell Impagliazzo and Ramamohan Paturi. Williams describes parts of the OpenAI argument as unusual enough to require collective study and discusses support for a conference to examine the implications.
His participation is evidence that a verification effort is being organised, not a certificate of correctness. Equally, Anthropic’s competing artifact should not be treated as independent replication simply because it comes from another company. That conclusion would require examining how the arguments relate and what has been checked separately.
Refuting ETH would not by itself settle P versus NP
The Clay Mathematics Institute’s explanation of P versus NP centres on the difference between efficiently checking a proposed answer and efficiently finding one. That question is broader than whether a particular exponential-time assumption survives.
A subexponential algorithm is not automatically a polynomial-time algorithm. Consequently, the announced ETH claims should not be rewritten as a proof that P equals NP. Nor does an asymptotic improvement alone establish a practical method for solving every difficult instance or breaking deployed cryptographic systems.
This is not merely cautious wording around an otherwise settled result. Complexity statements concern how resource requirements scale, and their interpretation depends on the exact problem and bound. Practical performance and broad consequences require additional arguments beyond a laboratory’s description of a theorem.
The new development is therefore a larger and more contested public research record. Anthropic has supplied a separate claimed result, while academic researchers are working out how to examine potentially disruptive arguments. The next meaningful advance will be a specific checking result, correction or independently explained consequence—not the number of laboratories repeating that a long-standing hypothesis is false.










