Aaronson-Ambainis Conjecture Solved by AI in Just 3 Days: What the Proof Really Means

On 6 November 2026, a small Cambridge startup called Physical Superintelligence — PSI for short — announced something that stopped the quantum computing world in its tracks. Its internal AI research system had produced a complete proof of the Aaronson-Ambainis conjecture, a problem that had resisted every human attempt since 2008. The system reached the Aaronson-Ambainis conjecture proof on the third day of its run, and the Lean proof assistant — software that checks every logical step by computer — accepted it.

Eighteen years is a long time for a maths problem to sit open, especially one that sits at the centre of a field as fast-moving as quantum computing. The Aaronson-Ambainis conjecture asks a deceptively simple question: when can a quantum computer be exponentially faster than an ordinary one? The new proof answers it for the hardest case — problems with no hidden structure — and the answer is a firm no.

This article explains the Aaronson-Ambainis conjecture in plain English, walks through how the AI system found the Aaronson-Ambainis conjecture proof, shows exactly what the machine-checked verification does and does not guarantee, and separates the genuine milestone from the hype. The result is real and important. It is also narrower — and less finished — than some headlines suggest.

Aaronson-Ambainis conjecture proof context — golden quantum computer chandelier in a research laboratory

Table of Contents

What Is the Aaronson-Ambainis Conjecture?

The Aaronson-Ambainis conjecture is a statement about Boolean functions — functions whose inputs and outputs are strings of bits, the 0s and 1s that all computing is built from. It was posed in 2008 by Scott Aaronson and Andris Ambainis in their paper “The Need for Structure in Quantum Speedups.” For eighteen years it stood among the central open problems of quantum complexity theory, the branch of computer science that sorts problems by how much effort quantum and ordinary computers need to solve them.

Here is the idea in plain English. Take any function that maps a string of bits to a yes-or-no answer, and suppose the function can be approximated by a polynomial of low degree — a mathematical expression built from the input bits with only a limited number of multiplications. The Aaronson-Ambainis conjecture says that such a function must have at least one input bit that matters a lot: flipping that single bit changes the output with a probability that cannot be too small. More precisely, if the polynomial has degree d and variance v (a measure of how much the output actually varies), then some variable has influence at least a fixed fraction of v raised to the power 5/2, divided by d raised to the power 6.

Why should anyone outside pure maths care? Because the acceptance probability of a quantum algorithm — the chance that it gives the right answer — is itself a low-degree polynomial in the input bits. If every low-degree polynomial has an influential variable, then a quantum algorithm cannot hide its dependence on the input. It cannot be sensitive to everything at once and nothing in particular. That single fact is what blocks an exponential quantum speedup on unstructured problems, as the next sections explain.

Aaronson-Ambainis conjecture mathematics — university chalkboard covered in advanced equations

The Aaronson-Ambainis conjecture in one sentence

If a yes-or-no question about a string of bits can be described by a low-degree polynomial, then at least one single bit must have a large influence on the answer. There is no way for the answer to depend on all the bits equally and only weakly.

Why quantum speedups need structure

Quantum computers earn their speed from interference: many computational paths are explored at once, and the wrong answers cancel out while the right ones reinforce each other. That trick works beautifully when the problem has structure to exploit — a hidden period, a symmetry, a pattern the algorithm can lock onto. Shor’s algorithm, which factors large numbers exponentially faster than any known classical method, works precisely because factoring has that kind of structure.

The Aaronson-Ambainis conjecture is about the opposite case: a random oracle, a black box whose answers look random, with no pattern to exploit. On such inputs, the Aaronson-Ambainis conjecture says, a quantum computer cannot be exponentially faster than a classical one. The speedup, if any, is at most polynomial — a real advantage, but not the dramatic, exponential leap that makes headlines. The Aaronson-Ambainis conjecture therefore draws a boundary line through the whole field: structure is not a nice bonus for quantum speedups, it is a requirement.

An AI Proved the Aaronson-Ambainis Conjecture in Three Days

On 6 November 2026, Physical Superintelligence published a short research note and a 13-page paper announcing the Aaronson-Ambainis conjecture proof, together with the complete formalisation in Lean. According to PSI, the proof came from an internal research system running instances of a non-public Anthropic model on the company’s own computing infrastructure. The system had a complete proof during the third day of the run. Lean accepted it. PSI then pointed the system at related open problems, and it improved the result twice more. Lean accepted the final version too.

The claim spread fast. Jack Clark’s Import AI newsletter led its 10 November issue with the story, and coverage followed across the quantum and AI press within days. What made the announcement land so hard was the combination: an 18-year-old problem, a three-day run, and a machine-checked proof that anyone can download and re-run.

Who is PSI?

Physical Superintelligence is a US startup based in Cambridge, Massachusetts, that describes itself as an AI-native physics laboratory. It launched in September 2026 as a venture-backed public benefit corporation — a for-profit company with a stated public mission. The Aaronson-Ambainis proof is its first major public result, and it arrived barely two months after the company’s launch.

Two things about PSI matter for understanding this story. First, the company is easy to confuse with the Paul Scherrer Institute, the Swiss research centre that shares the same initials — they are unrelated. Second, PSI is not a household name like OpenAI or Google DeepMind. Its announcement carried no big-brand weight of its own, which is exactly why the Lean formalisation matters so much: the Aaronson-Ambainis conjecture proof does not ask readers to trust PSI’s reputation. It asks them to run the code.

Inside the three-day run

PSI’s research system was built by staff member David Turturean, who scoped the problem, ran the system, reviewed and edited the paper, and takes responsibility for its correctness. The paper’s AI disclosure says copies of a non-public Anthropic model — called instances — worked inside the system “to search for existing ideas, to propose new ideas, to combine and test these ideas, and to supervise the work and the communication between other instances.”

That description is worth reading carefully. This was not one model answering one prompt. It was a system of many model instances, each with a role, coordinated by a harness that Turturean built: some instances searched the literature, some proposed approaches, some tested them, and some supervised the others. A timeline in the paper’s appendix says the first complete proof was written down on the third day — a version that proved the Aaronson-Ambainis conjecture, but weakly. The system then kept working, strengthening the bound twice, and Lean accepted each version in turn.

The proof by the numbers

The scale of the formal proof is part of the story. A formal proof in Lean is not a sketch: every step must be spelled out in code, and the Lean kernel checks each one. Here are the figures PSI published on 6 November 2026:

  • 9,454 lines of Lean code in the final formalisation.
  • 648 machine-checked theorems, with no step left as a placeholder and no added axiom — no unproved assumption smuggled in.
  • 84 Lean files (modules) holding 761 declarations, the named definitions and results.
  • Four theorems in the final version, proved with no placeholders.
  • About 11 minutes for a single-threaded replay of the whole formalisation, using 7.4 GB of memory — which puts an independent re-run within reach of anyone with an ordinary Linux machine.
  • The formal statement itself is a short definition called PrintedConjecture.

The final bound the Aaronson-Ambainis conjecture proof establishes is explicit: in any non-constant bounded polynomial of degree d with variance v, some single input bit has influence of at least (8√15 / (125π⁴)) × v^(5/2) / d^6 — roughly 0.0025 times v to the power 2.5, divided by d to the power 6. The constant is small, but the shape of the bound is what the Aaronson-Ambainis conjecture demanded: influence that cannot vanish as the problem grows.

How the Proof Was Found

The proof did not come out of nowhere. It followed a plan that other mathematicians had already laid out — and it closed the one gap in that plan. Understanding that lineage is essential, because it shows what the AI system actually did: it did not invent quantum complexity theory from scratch. It finished a route that human researchers had mapped most of the way.

The route comes from Shachar Lovett and Jiapeng Zhang, who reduced the Aaronson-Ambainis conjecture to a single open inequality. Prove that inequality, and the Aaronson-Ambainis conjecture follows. Their plan drew on earlier ideas from David Reimer and others, and it left one step unproven. PSI’s system replaced that step with a new inequality of its own — and then formalised the whole chain in Lean.

That is a pattern worth noting for anyone following AI research. The system’s job was not to dream up a new field. It was to search a well-defined space of candidate ideas, test them against a merciless checker, combine the ones that survived, and keep going until the chain closed. Lean played the role of a tireless referee: every candidate step was either accepted or rejected immediately, with no room for hand-waving. A human mathematician working alone might spend weeks checking whether a promising inequality actually holds. The system could test, discard, and retry at a pace no person can match — and it had three days to do it.

Aaronson-Ambainis conjecture research — scientist writing physics equations on a transparent board

The human role has not disappeared

It would be a mistake to describe this as a proof produced by a machine alone. Turturean chose the problem, built the harness, and reviewed the output. PSI’s advisor Jim Halverson audited the Aaronson-Ambainis conjecture proof. The model instances packaged the results into a first draft of the paper, and humans reviewed the draft and remain responsible for its correctness. What changed is the division of labour: the searching, proposing, testing, and combining — the slow, grinding middle of mathematical research — was carried out largely by the system, while humans set the direction and owned the result.

That division is the honest version of “AI proved a theorem.” The system did an enormous amount of genuine mathematical work, including the creative step of finding the replacement inequality. But it worked inside a problem a human framed, on infrastructure a human built, under supervision a human provided. Readers who want the full picture should keep both halves in view.

What Lean Checks — and What It Does Not

Lean is a programming language built for mathematics, and a proof assistant in which a computer checks every step of an argument. Rather than relying on a human referee to read and check every line, you write the proof as code, and a small, trusted program called the kernel verifies that each step follows logically from the last. If the kernel accepts, the theorem as written is proved. There is no ambiguity, no “the reader can fill in the details,” and no place to hide a gap.

PSI replayed its formalisation through Lean’s core kernel and through lean4lean, an independent type checker. That is a strong guarantee — stronger, in one sense, than traditional peer review, because a computer does not get tired, miss a step, or defer to a famous author. A single-threaded replay takes about eleven minutes, so any sceptic with a Linux machine can verify the formal proof for themselves.

But the guarantee is narrow, and PSI’s own materials say so. Lean acceptance proves the theorem as it was formalised. It does not, by itself, establish three things that still need human checking:

  • That the formal statement matches the Aaronson-Ambainis conjecture. A reader must confirm that the short Lean definition called PrintedConjecture really says what Aaronson and Ambainis wrote in 2008. If the formalisation accidentally states something slightly weaker, Lean would still accept the Aaronson-Ambainis conjecture proof — of the weaker statement.
  • That the human-readable paper is correct. The paper written for human readers “is not itself machine-checked.” Its explanations, motivations, and claims about what the result means still need expert reading.
  • That every piece is formalised. One of the four final theorems depends on a hypothesis from Agarwal and Ben-David that is not formalised in the repository. That does not make the proof wrong, but it does mean one link in the chain rests on a stated assumption rather than on machine-checked code.

Lean’s kernel, compiler, the Mathlib library of definitions, and the checking hardware are also part of the trusted base. None of that is a flaw — every verification system has a trusted base — but it is why the Stanford team’s independent reading matters, and why peer review still has a job to do. Machine checking removes one class of error (a wrong step) while leaving another (a wrong statement of the problem) firmly in human hands.

The 18-Year Road to the Proof

The Aaronson-Ambainis conjecture was not ignored for eighteen years. It was attacked, and it fought back. Its history is a useful antidote to the idea that AI simply walked in and solved something humans could not touch.

  • 2008 — the Aaronson-Ambainis conjecture is posed. Aaronson and Ambainis publish “The Need for Structure in Quantum Speedups,” arguing that quantum speedups require structure in the problem. The Aaronson-Ambainis conjecture is the precise mathematical form of that argument.
  • 2019 — a proof collapses. A proof attempt by Keller and Klein falls apart over one flawed lemma. Aaronson writes about it on his blog Shtetl-Optimized in a post titled “The Aaronson-Ambainis Conjecture (2008-2019),” and the Aaronson-Ambainis conjecture returns to open status. The episode shows how easily a long, technical proof can hide a single fatal step — and why a machine-checked proof is so attractive.
  • August 2026 — the Stanford special case. Four researchers at Stanford — Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan — prove the consequence of the Aaronson-Ambainis conjecture for quantum algorithms that query in a fixed number of parallel rounds. It is a genuine advance, and it shows the Aaronson-Ambainis conjecture was already cracking at the edges.
  • 6 November 2026 — PSI announces the full proof. The complete proof arrives on day three of the run, is accepted by Lean, is improved twice, and is published on GitHub the same day with the paper and the full formalisation.

The Stanford team is now writing a shorter exposition of the proof for experts — the kind of human-readable account that lets the wider community check the ideas, not just the code. One Stanford researcher called it “a huge result.” As of 7 November, neither Scott Aaronson nor Andris Ambainis had published a formal assessment, although a reader had already posted a link to PSI’s note under Aaronson’s 2019 post, asking for his view. Crypto Briefing later reported Aaronson as saying he does not fully follow the proof yet but believes it anyway — a striking detail, and a reminder that even the Aaronson-Ambainis conjecture’s author is still catching up with the formal argument.

Aaronson-Ambainis conjecture history — blackboard filled with advanced mathematical formulas

What the Proof Means for Quantum Computing

If the Aaronson-Ambainis conjecture proof survives outside scrutiny, it settles a boundary question that has shaped quantum computing since the field began: on unstructured problems — questions about random inputs with no pattern to exploit — quantum computers cannot be exponentially faster than classical ones. The finding does not unlock a faster quantum algorithm. It closes a door.

That matters because it tells researchers where not to look. As one report put it, “It’s a negative result that allows people to know, okay, we’re no longer looking there.” Quantum advantage is real, but it lives where structure lives: factoring, simulation of quantum systems, certain optimisation and search problems with exploitable patterns. The proof says the advantage cannot come from nowhere.

Three clarifications keep the result honest:

  • Shor’s algorithm is unaffected. Factoring has deep mathematical structure, and the Aaronson-Ambainis conjecture proof constrains quantum advantage on unstructured inputs, not every kind of quantum advantage. The most famous quantum speedup in history stands exactly where it stood.
  • It counts queries, not total time. The result is about how many times an algorithm must read the input, not how long the whole computation takes. As Matt Swayne noted in The Quantum Insider, that is the right measure for this question — but it is not the same as saying a quantum computer could never finish such a task sooner in wall-clock terms once all costs are counted.
  • Polynomial speedups are still on the table. The proof rules out exponential advantage on random oracles for decision problems. A quadratic speedup — the kind Grover’s search algorithm offers — is a different, smaller claim, and it is not what the Aaronson-Ambainis conjecture addresses.

For the quantum industry, the practical effect is focus. Hardware builders, algorithm designers, and investors all benefit from knowing which speedups are possible in principle and which are not. A boundary proved is a boundary that no longer needs to be guessed at.

What It Means for AI Research

For AI, the Aaronson-Ambainis conjecture proof is a different kind of milestone: evidence that research systems built around frontier models can now do original work at the edge of human knowledge — not just summarise papers or solve textbook exercises, but close an open problem that experts had left open for eighteen years.

Two features of the run deserve attention. The first is the multi-instance design: the system was not a single model call but a coordinated team of model instances searching, proposing, testing, and supervising each other. That architecture — agents with roles, checked by a formal verifier — is the same pattern now appearing across AI research tools, from coding agents to laboratory automation. The second is the verifier itself. Lean gave the system something most AI research lacks: an objective, instant, incorruptible judge of whether a step was right. In mathematics, that judge exists. In most of science, it does not, and that difference will shape where AI research systems succeed first.

It is also worth keeping this result in proportion. AI systems have been producing machine-checked mathematics for several years — proving known theorems, formalising existing results, and helping mathematicians explore. What changed on 6 November is the target: a famous open problem, solved rather than re-proved. It is a step change in ambition, not a sudden arrival of machine mathematicians. The same week brought other evidence that AI is accelerating AI research itself, and the pattern is consistent: the systems are strongest where verification is cheap and unambiguous, and humans still set the questions that matter.

Aaronson-Ambainis conjecture AI research system — supercomputer data centre running the proof search

For readers of this site, the closest parallels are the AI-for-science stories we have covered before: systems that do real scientific work under human direction, with verification as the dividing line between a demo and a result. The Aaronson-Ambainis conjecture proof belongs in that company — and it raises the bar for what counts as evidence when an AI system claims a discovery. Our earlier look at Hydrogen-1.0, the AI system doing open-ended science, and our guide to how long AI research findings stay trustworthy, both circle the same question this proof forces: how do you know an AI result is real? Here, the answer is unusually clean — run the code.

The Honest Limits of the Aaronson-Ambainis Conjecture Proof

A result this striking invites over-claiming, so here are the limits, stated plainly and in one place.

  • It is not peer-reviewed. The paper is a preprint, posted without journal review. PSI’s materials do not mention a journal submission, and no journal has reviewed the paper yet. Preprints are normal in maths and physics, but they are not the same as a refereed publication.
  • The audit was internal. Advisor Jim Halverson audited the Aaronson-Ambainis conjecture proof, but that audit happened inside PSI’s orbit. The independent check so far is the Stanford team’s reading and their forthcoming shorter exposition — valuable, but not yet published.
  • The model is non-public. The proof used instances of a non-public Anthropic model whose name PSI has not given, and Anthropic has not said it took part in or endorsed the work. Outside researchers cannot rerun the discovery — only the verification. Nobody can yet say how much of the Aaronson-Ambainis conjecture proof belongs to the model, the harness, or the humans, because PSI’s announcement does not disclose enough to assign precise shares.
  • One hypothesis is not formalised. As noted above, one of the four final theorems depends on a hypothesis from Agarwal and Ben-David that the repository does not formalise. The gap is disclosed, not hidden — but it is a gap, and readers should know it is there.
  • The formal statement still needs a human check. Lean proves PrintedConjecture. Humans must confirm that PrintedConjecture is the Aaronson-Ambainis conjecture. That check is the single most important piece of unfinished verification.
  • It proves a limit, not a capability. The result says quantum computers cannot do something. It does not make quantum computers faster, cheaper, or closer to practical use. Anyone selling this as a breakthrough in quantum hardware is misreading it.

None of these limits cancels the result. A machine-checked proof of an 18-year-old conjecture is a genuine event. But the limits are the difference between understanding the event and merely repeating the headline, and they are the reason the careful coverage of this story — from Import AI to the independent explainers that followed — keeps coming back to the same words: claim, announced, says. Until peer review and independent exposition land, those verbs are the right ones.

What Happens Next

The next steps are already visible. The Stanford team’s shorter exposition for experts will be the first independent human account of how the Aaronson-Ambainis conjecture proof works — the version working mathematicians will actually read. Peer review, if PSI submits the paper to a journal, will follow on a slower clock: months, not days. And the community will keep probing the formalisation itself, checking that PrintedConjecture matches the 2008 statement and that the Agarwal and Ben-David hypothesis is sound.

There is also a second, quieter question: what does PSI do next? The company pointed the same system at related open problems immediately after the proof and improved the result twice. If that pipeline keeps producing — if the system that closed one 18-year problem can close another — then the story stops being about a single conjecture and becomes about a method. Watch for a second result. One proof is a milestone; two is a programme.

For the wider field, the proof will sharpen a debate that is already running: how much of mathematical research can be handed to systems like this, and what remains essentially human? The Aaronson-Ambainis conjecture proof suggests the answer is shifting — not because machines now do maths alone, but because the combination of a research system, a formal verifier, and human direction is now powerful enough to settle questions that stood open for a generation.

Aaronson-Ambainis Conjecture: Key Facts at a Glance

Fact Detail
Conjecture posed 2008, by Scott Aaronson and Andris Ambainis
Original paper “The Need for Structure in Quantum Speedups”
Proof announced 6 November 2026, by Physical Superintelligence (PSI)
How long it took the system Complete proof on day 3 of the run; improved twice afterwards
Model used Instances of a non-public Anthropic model (name not disclosed)
Verification Lean proof assistant accepted every version; replayed through the Lean kernel and lean4lean
Formal proof size 9,454 lines; 648 theorems; 84 Lean files; 761 declarations; no added axioms, no placeholders
Replay time About 11 minutes single-threaded; 7.4 GB of memory
Proved bound Some variable has influence ≥ (8√15 / (125π⁴)) · v^(5/2) / d^6
Key prior step Lovett–Zhang plan, leaving one open inequality; Stanford special case, August 2026
Peer review Not yet; preprint only, no journal submission disclosed
Where to read it PSI paper and Lean formalisation on GitHub

Frequently Asked Questions

Did an AI really prove the Aaronson-Ambainis conjecture?

PSI says its internal research system, running instances of a non-public Anthropic model, produced the Aaronson-Ambainis conjecture proof on the third day of a run, and the Lean proof assistant accepted it. The formal proof is public on GitHub and anyone can replay the verification in about eleven minutes. The paper has not been peer-reviewed yet, and the formal statement still needs expert confirmation that it matches the 2008 conjecture exactly. So the careful answer is: an AI research system produced a machine-checked proof that experts are now verifying — the strongest form of “AI proved it” seen so far, but not yet the final word.

What is the Aaronson-Ambainis conjecture in simple terms?

It says that if a yes-or-no question about a string of bits can be described by a low-degree polynomial, then at least one single bit must have a large influence on the answer. In quantum terms, it means quantum computers cannot be exponentially faster than ordinary computers on unstructured, random-looking problems — they need structure in the problem to earn their speedup.

Does this proof make quantum computers useless?

No. It rules out exponential speedups on unstructured inputs only. Structured problems — factoring with Shor’s algorithm, simulating quantum systems, and other pattern-rich tasks — are unaffected. The proof draws a boundary around where quantum advantage can exist; it does not erase the advantage inside that boundary.

What is Lean, and why does its approval matter?

Lean is a proof assistant: you write a proof as code, and a small trusted kernel checks every logical step. If Lean accepts, the theorem as written is proved, with no gaps a human referee might miss. It matters because it makes the proof independently checkable — but it only proves the statement as formalised, so experts must still confirm the formalisation matches the original conjecture.

Who is PSI, and what model did it use?

Physical Superintelligence is a Cambridge, Massachusetts startup launched in September 2026 that calls itself an AI-native physics laboratory. It says the Aaronson-Ambainis conjecture proof came from its internal research system running instances of a non-public Anthropic model. Anthropic has not named the model, and has not said it took part in or endorsed the work.

Where can I read the Aaronson-Ambainis conjecture proof?

PSI published the paper and the full Lean formalisation on GitHub, in the repository physical-superintelligence/aaronson-ambainis. The repository holds the 13-page paper, the 84 Lean files, and audit notes. You can also learn how Lean itself works at lean-lang.org.

Conclusion

The Aaronson-Ambainis conjecture stood open for eighteen years because it sits at a hard boundary: the line between problems where quantum computers can be dramatically faster and problems where they cannot. On 6 November 2026, an AI research system at a two-month-old startup crossed that boundary’s last gap in three days, and a machine checked every step of the result. That is a genuine milestone — for quantum complexity theory, and for AI research itself.

It is also a milestone with edges still rough: no peer review yet, an internal audit, a non-public model, one unformalised hypothesis, and a formal statement that experts are still confirming. The right response is neither breathless hype nor sceptical dismissal. It is the response the mathematics community is already giving: run the code, read the Aaronson-Ambainis conjecture proof, write the shorter exposition, and see whether it holds. If it does — and the early signs are that it will — then the Aaronson-Ambainis conjecture will be remembered as the first famous open problem closed by an AI research system, and the day the machines stopped only assisting mathematics and started advancing it.

For more on how AI systems are doing real scientific work — and how to tell a real result from a demo — see our explainers on Hydrogen-1.0 and the shelf life of AI research findings, and our look at Anthropic’s FlashProve verification system.

Leave a Comment

Your email address will not be published. Required fields are marked *