The Signal — November 7, 2026

A research system running an unreleased Anthropic model produced a machine-checked proof of a quantum computing conjecture open since 2008. The White House set rules for who can sell synthetic DNA to federally funded labs, and OpenAI tested how well AI reads raw lab data.

A research system running an unreleased Anthropic model produced a machine-checked proof of a quantum computing conjecture that had been open since 2008, though the experts now reading it haven't published their verdict. The White House also set rules for who can sell synthetic DNA to federally funded labs, and OpenAI published a test of how well AI models read raw lab data.

An unreleased Anthropic model helped prove a quantum computing conjecture open since 2008, and experts are still reading the proof

Physical Superintelligence (PSI), a physics research lab, said on November 6 that a research system it built in-house, running instances of a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture. The system had a complete proof on the third day of the run, and Lean, a proof assistant in which software checks every step, accepted it, according to PSI's announcement. PSI then set the system to work on related open problems, and along the way it improved the result twice; Lean accepted the final version as well.

Most dramatic quantum speedups, such as Shor's factoring algorithm, depend on structure: a promise that the input hides a rare pattern, like a repeating sequence. Scott Aaronson and Andris Ambainis asked whether a quantum computer could still be exponentially faster on random inputs, which almost never have such patterns, and showed that the answer would follow from a statement about polynomials. With that statement proved, an ordinary computer can estimate a quantum algorithm's chance of answering yes to a yes-or-no question, up to a small error, on all but a small fraction of inputs. The number of times the ordinary computer has to read the input grows at most like a fixed power of the quantum algorithm's reads, not exponentially. The Quantum Insider points out that the theorem limits reads of the input rather than total computing time, so it doesn't settle which applications will get useful quantum speedups. The outlet expects it to steer quantum algorithm research toward problems whose mathematical or physical structure offers an advantage.

"The mathematical statements in this paper were produced by AI models," the paper's AI disclosure says. David Turturean of PSI scoped the problem, built the research harness and ran it, then prompted the model to write a first draft of the paper. He reviewed and edited the draft and is responsible for its correctness. Inside the harness, model instances searched for existing ideas, proposed and tested new ones, and supervised the other instances. The proof follows a plan by Shachar Lovett and Jiapeng Zhang, replacing an inequality they had left as a conjecture with a new one proved in the paper. PSI advisor Jim Halverson audited the proof.

The paper is posted on GitHub and has not been peer reviewed. Four Stanford researchers, Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan, proved a special case in August and are now reading the full proof and writing a shorter version for experts. PSI quotes them calling it "a huge result." The paper thanks Aaronson for referring PSI to that group, and it notes that an earlier attempt at a proof was announced and later withdrawn. Anthropic hasn't said which model was used and didn't respond to Business Insider's request for comment. As of this morning, Aaronson hadn't written about the proof on his blog.

Sources: PSI · PSI paper · The Quantum Insider · Business Insider


Federally funded labs will have to buy synthetic DNA from suppliers with verified screening and customer checks

A White House memo dated November 5, from National Security Advisor Marco Rubio and White House science office director Michael Kratsios, set a national policy for screening synthetic nucleic acids, which The Daily Caller describes as the raw material scientists use to build viruses in the lab. Every recipient of federal life sciences funding will have to buy synthetic nucleic acids and benchtop synthesis machines only from certified suppliers. The policy replaces a 2024 framework that, according to the memo, relied on self-attestation rather than independent verification.

To be certified, a supplier has to run every order of 50 or more nucleotides through a screener verified by the National Institute of Standards and Technology (NIST), check who its customers are, report flagged orders to authorities such as the FBI, and keep screening records for at least eight years. The supplier still makes the final call on each order. Benchtop synthesizers, which work outside the controls of a central synthesis facility, must refuse to start without a valid authorization from a verified screener, apart from limited offline use that NIST standards will define.

The schedule is phased. NIST has 90 days to issue screening standards, the Department of Health and Human Services is to publish the first list of approved suppliers within 160 days, and funded labs must buy nucleic acids only from those suppliers by day 250 and benchtop equipment by day 340. The 2025 executive order behind the policy asked for it within 90 days, and The Daily Caller counts it as roughly 460 days late. The memo cites risks "including those from SI uplift," using the administration's term "super intelligence" for AI, and the policy tells NIST to develop "super intelligence-assisted" screening methods, including ways to catch sequences designed to evade comparison with known ones.

The requirement binds only labs that take federal money. The policy says it should be implemented in a way that "encourages its applicability to all synthetic nucleic acid orders," but it doesn't require that. The Daily Caller notes that Congress could take up the Biosecurity Modernization and Innovation Act, which would extend screening to privately funded labs and regulate DNA printers. Gerry Parker, who drafted the executive order while at the National Security Council, said the policy "replaces an honor system with a real, verifiable system for accountability," and Nathan Calvin of the AI safety group Encode AI called screening "even more urgent to enact in the era of rising AI capabilities." We haven't yet seen a critical response from suppliers or researchers.

Sources: The White House · The Daily Caller


OpenAI's wet-lab benchmark finds GPT-6 Astra rejecting an unsuitable control plate, then picking the wrong colonies on another task

OpenAI and Ligo, a biotech company focused on enzyme design, published BioBasicsBench on November 6. Its 100 tasks come from scientists' own experiments and use raw evidence such as gel and blot images, photos of colony plates, microscopy stacks, sequencing reads, plate-reader exports and videos of bench work. Agents have to interpret the data, catch errors and recognize when the evidence can't support a conclusion. According to OpenAI's post, each task was checked by another Ligo scientist who hadn't written it and an outside PhD-level scientist working in a wet lab, then underwent a final internal review. Answers earn partial credit against expert-written rubrics.

At the "xhigh" reasoning setting, GPT-6 Astra earned 66.7% of the available credit, against 43.8% for GPT-5.6 Sol, while using about half as many output tokens per task. OpenAI suggests stronger native multimodal capabilities "may be key" to the gain. In one task, GPT-5.6 Sol saw no colonies on a negative control plate and built its conclusions on that, while Astra noticed the plate was labeled for kanamycin, a different antibiotic from the ampicillin used in the experiment, and rejected it as a control. In another, Astra recognized that a microscope recording showed a z-stack, a series of images taken at different depths. That explained why cells came in and out of focus; Sol had looked for a biological explanation.

OpenAI also shows where Astra fails. On a colony PCR gel, it sorted the bands into the right groups but assumed the larger product marked a successful clone, reversing what the controls showed and selecting colonies that would have led to purifying the wrong plasmid. R&D World notes that a wrong colony pick can cost a researcher a day of culturing cells and purifying a plasmid the lab already had. Astra also discounted microscope observations that contradicted its preferred reading and overcounted one plate's colonies by about 12%. "Higher benchmark scores do not yet mean models can make reliable judgments in the lab," OpenAI writes.

OpenAI built the benchmark with Ligo and tested only seven of its own models, as R&D World points out. We haven't seen results from anyone outside OpenAI. The score is partial credit, not the share of tasks fully solved, and the tasks involve reading data rather than running experiments. OpenAI ends its post by saying that tracking these capabilities will help it understand their implications for biosafety and biosecurity.

Sources: OpenAI · R&D World


On the Editor's Desk

We held the Justice Department's new Super Intelligence Task Force, which absorbs the department's January AI litigation task force and hasn't yet named cases or issued the prosecution guidance it is supposed to write. Crusoe's confirmation that Google bought its Wyoming data-center site adds little to a deal Bloomberg first reported in June, according to TechCrunch. Ai2's ToyBlocks robot-learning simulator can wait for the technical report its developers say is coming.