AI
“Physical Superintelligence's AI research system, powered by an unreleased Anthropic model, solved the 18-year-old Aaronson-Ambainis conjecture in quantum computing." (Post overlay text: "Anthropic's unreleased AI model helped solve an 18-year-old math problem in quantum computing / A proof of the Aaronson-Ambainis conjecture / The…”
Plain restatementOn 6 November 2026, Physical Superintelligence (PSI) announced that an internally built research system, running instances of a non-public Anthropic model, produced a proof of the Aaronson-Ambainis conjecture (posed 2008), that the Lean proof assistant accepted the proof, and that the system later improved the bound twice. A paper and a Lean 4 formalization were published.
This one largely holds up, with caveats worth knowing. On 6 November 2026 the company Physical Superintelligence did announce that its internal AI research system, running an unnamed non-public Anthropic model, produced a proof of the Aaronson-Ambainis conjecture, a question in quantum computing open since 2008, and that the proof-checking software Lean accepted it on the third day of the run. Unlike most company claims, this one comes with a public artifact: the paper and the Lean formalization are on GitHub and anyone can inspect or recompile them. The important caveats are that nothing here has been peer reviewed, the proof's authors themselves say readers still need to confirm that the statement Lean checked is really the original conjecture, and neither Scott Aaronson nor Andris Ambainis had publicly weighed in as of 8 November 2026. The "huge result" quote from Stanford researchers is real, but it was published by Physical Superintelligence on its own page, and that group is still reading the proof rather than having confirmed it. The phrase "the AI solved it" also simplifies a process in which a company researcher chose the problem, built and ran the system, reviewed the output and takes responsibility for its correctness. The same conjecture had a claimed proof in 2019 that was later withdrawn over an error, which is why specialists treat such announcements as provisional until they have been read.
Physical [drifted from the evidence:] Superintelligence's AI research system, [drifted from the evidence:] powered by an unreleased Anthropic model, [drifted from the evidence:] solved the 18-year-old Aaronson-Ambainis conjecture in quantum computing." (Post overlay text: "Anthropic's unreleased AI model helped solve an 18-year-old math problem in quantum computing / A proof of the Aaronson-Ambainis conjecture [drifted from the evidence:] / The [drifted from the evidence:] Aaronson-Ambainis conjecture is now a theorem." Caption adds: proof produced in three days, verified in Lean, improved twice, and [drifted from the evidence:] "Stanford researchers reviewing the work called it a [drifted from the evidence:] 'huge result.'")
[added by the neutral restatement:] On 6 November 2026, Physical [added by the neutral restatement:] Superintelligence (PSI) announced that an internally built research system, [added by the neutral restatement:] running instances of a non-public Anthropic model, [added by the neutral restatement:] produced a proof of the Aaronson-Ambainis conjecture [added by the neutral restatement:] (posed 2008), that the Lean [added by the neutral restatement:] proof assistant accepted the proof, and that the system later improved [added by the neutral restatement:] the bound twice. [added by the neutral restatement:] A paper and a [added by the neutral restatement:] Lean 4 formalization were published.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Physical Superintelligence announced on 6 November 2026 that its internal research system, running instances of a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture. The announcement is on PSI's own site and official X account.
- The conjecture does date to 2008 (Aaronson and Ambainis, written up in Theory of Computing in 2014), so "18-year-old" is correct arithmetic for 2026.
- The subject matter is described correctly: the conjecture bears on whether quantum computers can be exponentially faster than classical ones on random inputs lacking special structure, via a statement about influential variables in bounded low-degree polynomials.
- A paper and a Lean 4 formalization are genuinely public on GitHub, and the headline Lean theorem is stated against a definition named `PrintedConjecture` written in the original Aaronson-Ambainis normalisation.
- PSI states the proof was complete on the third day of the run and that Lean accepted it, consistent with the post's "within three days" and "verified using Lean."
- PSI states the system improved the result twice afterwards with the final version also Lean-accepted. Secondary reporting of the paper's appendix gives the exponent path 22 to 11 to 7.
- The "huge result" quote is real and is attributed to Stanford researchers who work on this exact problem (Blanc, Docter, Strassle, Tan, authors of arXiv:2608.19158). PSI says they have been reading the full proof.
- Every load-bearing fact in the post traces to a single interested party. PSI announced the result, PSI published the paper, PSI's two named reviewers are PSI advisors, and the outside praise quote is relayed on PSI's own page rather than published by the Stanford group. The post presents this as a settled event rather than as a company announcement with an attached checkable artifact. The post's caption does say "PSI says," but the headline overlay and the claim as circulated drop that attribution entirely.
- The post presents the proof as finished and accepted. Both the paper's authors and the repository leave an explicit task to readers, namely confirming that the formal statement Lean verified, `PrintedConjecture`, matches the conjecture as originally printed. Lean guarantees the theorem written in Lean follows from its assumptions. It cannot guarantee that the theorem written in Lean is the conjecture. At the time of this check no expert outside PSI had published that comparison.
- "AI research system solved" compresses a human-directed pipeline. Per the repository's own disclosure, a PSI staff member built the system, chose the problem, ran it, prompted the write-up, reviewed the draft and is responsible for its correctness, while AI models produced the statements, first draft and formalization. The AI contribution is substantial and disclosed, but "the AI solved it" is a stronger and cleaner story than the record supports.
- "an unreleased Anthropic model" is as specific as the record gets. No model name, version, or access terms are published, and no Anthropic statement about the run was found. A reader cannot attribute this to any identifiable Anthropic system, and nobody outside PSI can attempt the same run.
- Neither the research system nor the model is available to anyone outside PSI. The run is not reproducible by third parties; only the output is checkable.
- Misattribution risk in downstream coverage: at least one aggregator headline reads "Stanford team confirms" PSI's proof. The body of that piece contains no independent Stanford confirmation, only PSI's relayed quote and the statement that the group is reading the proof. Reading a work-in-progress comment as a confirmation is an error that the post's "Stanford researchers reviewing the work called it a huge result" phrasing avoids but sits close to.
- Whether the formal statement `PrintedConjecture` is a faithful encoding of Aaronson-Ambainis Conjecture 1.7. This is the single hinge of the whole result, it is acknowledged as open by the authors, and no published expert comparison exists as of 2026-11-08.
- Whether the prose paper, which is not itself machine-checked, matches the Lean development it accompanies.
- Whether Scott Aaronson or Andris Ambainis endorse the result. No public assessment by either was found as of 2026-11-08.
- Whether Anthropic agrees with PSI's characterisation of the model's role. No Anthropic statement was located.
- How much of the proof's substance came from the model versus from problem framing, prior literature supplied to it, and human review. The run is internal and the transcript is not public.
- Whether the result survives specialist scrutiny. The same conjecture had a claimed proof in 2019 that its authors withdrew over an error, which is why specialists treat announcements of this kind as provisional until read.
PSI's own research page states that a research system it developed internally, running instances of a non-public Anthropic model on PSI infrastructure, had a complete proof of the conjecture during the third day of a run, that Lean accepted it, and that the system was then set to work on related open problems. PSI's page opens with the sentence "The Aaronson-Ambainis conjecture is now a theorem." The public GitHub repository contains a paper and a Lean 4 formalization. Its README describes the result as: a bounded polynomial of low degree on the discrete cube has an influential variable, and identifies the target as Conjecture 1.7 of Aaronson and Ambainis, "The need for structure in quantum speedups," Theory of Computing 10 (2014), 133-166 (Conjecture 6 of arXiv:0911.0996v3), which asks for a bound of the form (epsilon/d)^O(1). The repository's headline theorem is `aaronson_ambainis: PrintedConjecture`, with `PrintedConjecture` stated in the normalisation of Aaronson and Ambainis at the bound (epsilon/d)^7. Reporting of the repository's audit notes states there are no `sorry` placeholders and no added axioms, that `#print axioms` lists only the three standard Mathlib axioms, and that declarations were replayed through Lean's kernel and through lean4lean, an independent implementation of Lean's type checker. The repository discloses that AI models produced the mathematical statements, the first draft and the Lean formalization, and that David Turturean built and ran the system, scoped the problem, prompted the first draft, reviewed it, and is responsible for the paper's correctness. Halverson audited the proof; Shanahan commented on the formalization. Both are PSI advisors. The explicit bound reported from the formalization is that if a degree-at-most-d real polynomial bounded in [0,1] on the discrete cube has variance at least epsilon, some coordinate has influence at least (8*sqrt(15)/(125*pi^4)) * epsilon^(5/2) / d^6, and hence at least (epsilon/d)^7. Secondary reporting of the paper's appendix describes the exponent improving within one run from 22 to 11 to 7, each version Lean-accepted. The four researchers quoted as calling it "a huge result" are the Stanford group Blanc, Docter, Strassle and Tan, whose own August 2026 preprint (arXiv:2608.19158) proved a restricted case. Their quoted words appear in PSI's announcement, reported by PSI. Per PSI, they are reading the full proof, have sent comments, and are writing a shorter exposition for experts. As of this check, multiple outlets record that no public assessment by Scott Aaronson or Andris Ambainis has appeared, and no statement from Anthropic about the model or the run was found. The conjecture has one prior claimed resolution: Aaronson's own 2019 blog post announced a Keller-Klein proof, which was subsequently withdrawn over an error in a lemma.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Did an AI really solve the Aaronson-Ambainis conjecture on its own?
Not exactly. Physical Superintelligence's research system, running a non-public Anthropic model, produced the proof, but a PSI staff member built and ran the system, chose the problem, prompted the first draft, reviewed it and is responsible for the paper's correctness. The repository itself discloses this human role.
Has the proof been independently verified by outside mathematicians?
Not fully. The proof's own authors say readers still need to confirm that the formal statement Lean checked actually matches the original conjecture as printed. As of this check, neither Scott Aaronson nor Andris Ambainis had publicly commented, and no outside peer review had occurred.
What does it mean that Lean accepted the proof?
Lean is proof-checking software, and audits found no sorry placeholders or added axioms, with results replayed through Lean's kernel and an independent checker called lean4lean. However, Lean can only confirm that the theorem as written follows logically from its assumptions, not that the theorem as written is truly equivalent to the original conjecture.
Who called this a 'huge result,' and is that endorsement independent?
The quote comes from a Stanford research group (Blanc, Docter, Strassle and Tan) who had previously proved a restricted case of the conjecture. However, the quote was published by PSI on its own announcement page, and per PSI, the group is still reading the full proof rather than having confirmed it.
Has this conjecture been claimed as solved before?
Yes. In 2019 Scott Aaronson announced a Keller-Klein proof on his blog, but it was later withdrawn due to an error in a lemma. This history is part of why specialists treat new announcements as provisional until independently reviewed.