Latest / Quickly Quantum / August 2, 2026 — The Werner-State Race: AI, Humans, and One Open Problem
Transcript
- Twenty years unsolved, then five days flat — with one of the winning teams crediting a chatbot for finding the actual proof. That's the tension driving this whole episode. In the back half of July, at least three independent teams cracked a famous open problem in quantum information theory, one that had resisted every attack since the year two thousand. One team did it the old-fashioned way — pen and paper, a self-contained human derivation. Another team says the core proof came from a large language model, and the humans' job was just to check its work. Since it's Sunday, we're spending the whole show on this one question, from every side: what does it actually prove when AI and human mathematicians reach the same finish line, days apart, on a problem people thought was genuinely hard? Welcome back to Quickly Quantum, your daily brief on the quantum frontier. It's Sunday, August 2, 2026. Let's get into it. So what actually is the problem, and why did physicists bother naming it? Back in 1989, physicist Reinhard Werner introduced a family of quantum states — now called Werner states — built around a single tunable parameter. They matter because of entanglement, the quantum correlation that lets two particles behave like a linked pair no matter the distance, and because of distillation: taking a pile of weak, noisy entangled pairs and using only local operations plus a bit of classical communication to squeeze out fewer, cleaner, maximally entangled pairs — the kind you actually need for quantum teleportation or a future quantum internet. Some entangled states carry a mathematical fingerprint called a non-positive partial transpose, and for more than two decades nobody knew whether every state with that fingerprint could be distilled at all. Here's the elegant part: through a technique called twirling, that huge general question collapses down to one narrow question about Werner states specifically — solve it there, and you've solved it everywhere. That's why Werner states are the field's universal test case. In 2022, three prominent theorists — Paweł Horodecki, Łukasz Rudnicki and Karol Życzkowski — published a paper in PRX Quantum called Five Open Problems in Quantum Information Theory, naming the questions they considered ripe but stubborn. Problem five asked something specific: does a particular two-ququart Werner state — ququart meaning a four-level quantum unit, not the usual two-level qubit — become distillable if you're allowed two copies instead of one? A registry that tracks major open math problems, posting as @openconjectures on X, had logged it as, quote, one of five KCIK prize problems in quantum information theory, distillability of Werner states, end quote. Then, in the back half of July 2026, it broke open almost all at once. On July twenty-third, a team from Wuhan University — Fu, Gao and Park — posted a fully human proof. Four days later, on July twenty-seventh, a separate paper from Thomas Fraser, Felix Huber and coauthors answered Problem five directly — and stated, right in the paper, that the results were found and written up with AI tools. Let's start with the team making the loudest claim: Thomas Fraser, Felix Huber, and their coauthors. Their position, stated plainly in the paper, is that a large language model generated the actual mathematical content — not a hint, the proof itself — and the humans' role afterward was refinement and verification. One version of their preprint puts it this way, quote, the proof presented in this work was found with ChatGPT Sol-5.6 and refined with Claude Opus and Fable; all results have been carefully checked by the authors, end quote. That's not AI replacing mathematicians — it's AI doing the generative step, humans doing the auditing. Coauthor Balázs Pozsgay laid out the timeline himself on X, and it's worth hearing directly because it shows how fast this moved. Quote, crazy times, end quote, he wrote — his team found a solution on July twenty-second, and he added, quote, we checked the computations and they are correct, end quote. He described watching two other groups converge on the same result within about forty-eight hours, one already on arXiv, one circulating as a public document, and noted, quote, interestingly, our proof is different from the one on arxiv, end quote. That detail matters: if the AI-assisted route and the human route found genuinely different proof strategies for the same answer, that's not one team copying another — it's two separate paths landing on the same mathematical truth. The team's broader framing across their preprints points toward, quote, a structural change affecting the field of quantum information and computation, end quote. And they're upfront about the collision itself — one version of their paper explicitly notes becoming aware, mid-writeup, of the Wuhan team's paper and of a third group working an entirely different angle. Their claim, distilled: AI can now do real generative mathematics on a named, unsolved problem, provided a human is standing at the end of the line checking every step before it goes out the door. Flip that around, though, and you get the team that reached the same answer without any AI at all: Jinshi Fu, Li Gao and Sang-Jun Park at Wuhan University. Their July twenty-third preprint proves the same core equivalence — that Werner states in arbitrary dimension are two-copy distillable if and only if they're one-copy distillable — through a self-contained human derivation, a block-operator inequality and variational analysis of matrix blocks, classic mathematical-physics technique, zero language models involved. And they posted first, before the AI-assisted paper even appeared. Meanwhile, a further group, Kishor Bharti, Rishikesh Gajjala and Tobias Haug, was reportedly closing in on the same result through a third method entirely — according to the AI-assisted team's own paper, which noted becoming aware of that work mid-preparation. Put those three data points together and you get an argument that cuts against the flashier headline: if a fully human team reaches the same finish line in the same week as an AI-assisted team, and a third group is independently closing in via a different route altogether, then maybe what changed wasn't that AI unlocked something previously impossible. Maybe the surrounding math had simply matured to the point where the answer was reachable by anyone with the right specialized tools, human or machine. Researcher Dmitry Grinko, posting on X, laid out exactly this convergence — listing all three teams and their preprints side by side without editorializing, which is itself a useful, sober counterweight to the more triumphant framing floating around elsewhere. The human team's implicit position isn't that AI is useless. It's more modest, and frankly more testable: extraordinary claims about AI cracking open problems need a control group, and here we got one, almost by accident, inside the same seven days. Zoom out further, and you get the view from Terence Tao, the Fields Medalist, giving the marquee lecture at the International Congress of Mathematicians in Philadelphia this summer. His framing is what makes this Werner-state story feel like more than a niche physics curiosity. Tao argues mathematics is shifting from what he calls proof scarcity to proof abundance — for most of the field's history the bottleneck was generating proofs at all; now, he says, AI is strong enough at generating them, and improving fast at verifying them, that the constraint is flipping. He broke the mathematical workflow into five stages: proof generation, verification, exposition, publication, and what he calls canonicalization, the slow process of a result actually getting absorbed into the field's shared body of knowledge, taught, cited, built on. Which, if you think about it, sounds a lot like a corporate workflow chart applied to centuries of human thought. His point, though, is serious: AI is already good at the first two stages. It's the back three — writing a result up so humans trust it, publishing and attributing it properly, folding it into what the field considers settled — where the human bottleneck remains, and where he thinks automation will be hardest. Notice how neatly that maps onto the Werner-state race: three separate proof-generation events within days, and now the field has to do the slower work of sorting out which proof is cleanest, who gets credited, and whether the AI-assisted one holds up to the same scrutiny as the human one. Tao isn't an AI skeptic here — he's on record saying he was, quote, very impressed by the Leiden Declaration, end quote, and endorsed it wholeheartedly, including its call for mathematicians to publicly debate what mathematics is even for in an era where proofs get cheap. That's a measured position, not doom, not hype — a structural warning that the field's human infrastructure, the referees, the journals, the textbooks, was built for a world with far fewer proofs to check than the one Tao says is arriving now. The institutional pushback comes next, and it's the reason this isn't just an academic curiosity. In June 2026, a document called the Leiden Declaration on AI and Mathematics gathered over a thousand signatures within twenty-four hours and went on to be formally endorsed by the International Mathematical Union — the field organizing a collective public statement of values, which practically never happens. Its warning is specific: unchecked reliance on AI-generated proofs risks flooding the field with arguments that look plausible but are subtly wrong, it risks overwhelming peer review with more candidate proofs than qualified humans can check, it muddies who deserves credit when a model generates the core idea, and it worries research priorities could tilt toward whatever problems AI companies find useful to showcase, rather than what's mathematically important. Notice that last point lands directly on our story: is Problem five getting solved this month because it was mathematically ripe, or partly because it made a clean, marketable demonstration of what a named AI model can do? The Declaration doesn't answer that for this specific case, but it's exactly the kind of question it exists to raise. And there's a real-world preview of the failure mode, from theoretical physicist Steve Hsu at Michigan State. Hsu published a Physics Letters B paper in December where he credits GPT-5 with originating the core idea of a real physics result, using what he calls a Generate-Verify protocol — one model proposes, a separate model checks. His account of that process is a genuinely useful data point: the AI performed like a brilliant but unreliable collaborator, capable of the central creative leap but also capable of confidently wrong reasoning that could mislead even expert humans without rigorous cross-checking. That's the same shape as the caveat built into the Werner-state AI paper itself, where the authors explicitly say every AI-generated result was carefully checked before publication. Put Hsu and the Leiden Declaration together and the shared worry isn't that AI can't do good math — it clearly sometimes can. It's whether the humans checking that math can keep pace once this becomes routine instead of a headline. Here's where I land, honestly: the most interesting fact in this whole story isn't that AI helped solve an open problem. It's that a human team, with zero AI assistance, solved the same problem in the same week, using a completely different method. That's the control group nobody planned for, and it should make everyone slow down before declaring this an AI moment. If proof-generation had truly been unlocked exclusively by a language model, you wouldn't expect a pen-and-paper team to arrive independently, days apart, via a different route — and you certainly wouldn't expect a third group closing in from a third direction at the same time. What that convergence suggests to me is that the surrounding mathematical toolkit had matured to the point where this particular door was ready to open for anyone strong enough to push on it. AI happened to be one of the hands on that door. That doesn't make the AI-assisted result meaningless — it's a real, checked, published proof of a genuinely open problem, and that's not nothing. But it undercuts the cleanest version of the headline, the one implied by some of the more excited reactions online. One futurist account, @Dr_Singularity, reacted on X with, quote, 1000x science and tech progress acceleration is near, end quote. I don't think one problem, solved in parallel by humans and machines in the same week, tells you that. What it does tell you, and this is where I think Tao's framing is exactly right, is that the bottleneck is moving. It's not can AI generate a candidate proof anymore — clearly, sometimes, yes. It's whether the field's referees, journals, and citation norms can absorb candidate proofs arriving this fast, from this many directions, with this much ambiguity about who gets credit when three teams and at least two AI models converge on one answer inside a week. Nobody in this story has a clean answer to that yet — not Tao, not the Leiden signatories, not the teams themselves, who are still sorting out priority and method differences even now. And that's honestly the most useful thing this episode can leave you with — not a verdict, but a genuinely open question the field hasn't resolved. Time for the Hype Check. I'd put this one at a six out of ten on substance. The underlying math is real — a genuine twenty-plus-year-old problem, three independent teams, a checked and published proof. But the framing running around online that AI simply solved this problem oversells it relative to what actually happened, which is closer to: the problem had become reachable, and multiple approaches — some AI-assisted, some not — happened to reach it in the same week. If there's one belief this week's news should shake loose, it's the assumption that whether AI wrote the proof is the meaningful dividing line in a story like this. The more useful dividing line is whether anyone, human or machine, checked the work rigorously before it went out the door. That's the standard both teams here actually met, and it's the one that matters most as this starts happening faster, with far less time for anyone to check the work. If this is the kind of story that gets you thinking — a real math problem, cracked in parallel by humans and machines, days apart — that's exactly the show I'm trying to make every week, so follow Quickly Quantum wherever you listen, and if today's episode was worth your time, send it to one person who'd enjoy arguing about it with you. This has been Quickly Quantum, an AI-voiced podcast, created and built by a real human using today's cutting-edge technology. Nothing you heard on this show is financial advice. I'm Brian Lampert, and I'll catch you all tomorrow — take care!