Wednesday’s shape is verification. OpenAI says its models produced three proofs across mathematics, computer science and physics, including a disproof of a hypothesis that complexity theory has used for decades to set limits on algorithms. Anthropic and Theorem say Claude generated machine-checkable security proofs across 123 open-source libraries and found bugs that fuzz testing had missed. IBM linked two 120-qubit quantum processors with cables most of a metre long and moved quantum states between them at 99% fidelity. Google DeepMind taught its world model to speak. OpenAI, the loudest voice in the race for Nvidia silicon, reportedly plans to buy Google’s chips instead. And Anthropic discounted Claude by 40% for every level of American government. The models are producing claims faster than the institutions around them can check those claims. That gap—between generation and verification—is where Wednesday’s news lives, and it is a healthy gap to watch, because checking is the part of science that keeps wonder honest.
OpenAI claims three proofs, and hands the checking to outsiders
OpenAI announced on October 26 that its internal models produced three results: a disproof of the Exponential Time Hypothesis in computer science, a proof in mathematics covering the local smoothing conjecture with implications for the Kakeya and restriction problems, and an existence result in quantum field theory that includes quantum chromodynamics. A separate briefing reports that Anthropic published its own refutation of the same hypothesis, reaching the same conclusion by a different route. Every one of these claims awaits external peer review, and OpenAI labels the manuscripts research preprints. The computer science result is under review by the hypothesis’ original authors and by Professor Ryan Williams. The mathematics proof and its Lean formalization are public. The physics materials are withheld while the collaboration, led by David Gross, prepares its own explanation first—a choice the Advisory Group on Mathematics and Artificial Intelligence publicly challenged on October 26, calling for immediate release.
"OpenAI’s proofs remain mere claims until humans verify them."—Kim KJ briefing commentary, October 28
Hold the claim and the checking apart, because the checking is the story. Lean can confirm that a formal argument follows its rules; it cannot confirm that the formal statement captures the problem scientists intended to solve. A pure Yang–Mills existence result, for example, leaves the Clay Millennium Prize untouched, because the prize also requires a positive mass gap. The Simons Foundation has put $15 million behind verification work, with a $5 million commitment from OpenAI announced in early October, which tells you the labs themselves now budget for doubt. This is Chris’ standing interest in machine intelligence at its best: a serious claim, exhibits on the table where they exist, the question left open where it is open. If the Exponential Time Hypothesis falls, conditional limits across algorithm theory get reassessed. The P versus NP problem stays exactly where it was.
Claude writes the proofs for 123 software libraries
Anthropic reported on October 22 that, working with verification company Theorem, Claude generated roughly 180 million lines of machine-checkable mathematical proofs covering 239 test entry points across 123 open-source libraries. The method converted existing OSS-Fuzz tests into specifications, compiled the software for RISC-V, and used the Rocq proof checker to test Claude’s arguments. The work uncovered hundreds of bugs, including a decompression error that needed a roughly 20 MB input in a particular arrangement and an image-decoding defect that needed 64 MB and more than 2 GB of memory—inputs conventional fuzzing rarely assembles. A public repository holds proofs for 60 entry points across 25 libraries. Anthropic reports that Opus 5.5 finished half of a 36-task verification benchmark with about 40% fewer tokens than Opus 5.
The limits deserve the same print size as the headline. Each proof guarantees a specified property under specified conditions, for a single-threaded RISC-V build with modelled Linux behaviour and inputs capped near 64 MB. A proof says nothing about whether the program finishes, and nothing about defects outside its specification. Rocq checks the argument against the spec; humans still decide whether the spec describes the security property that matters. The reference point is seL4, the verified kernel that took about two person-years to build and twenty to verify. If Claude cuts that second number by an order of magnitude, verified libraries become ordinary infrastructure, and every application standing on those libraries inherits the evidence. Theorem’s OSS Verify programme offers the service free to qualifying open-source projects, with proofs taking several days.
IBM wires two quantum processors into one machine
IBM Research connected two 120-qubit Nighthawk R1 processors with four superconducting cables, each 80 centimetres long, inside one dilution refrigerator, and transferred quantum states between the chips at an average fidelity of 99.00%. The team also demonstrated remote gates between qubits on different chips and entanglement across the link, and the cables survived three warm-up and cool-down cycles. The paper, posted to arXiv on October 21 with 42 authors led by Neereja Sundaresan, is a preprint. Precision matters here: 99% is the fidelity of state transfer, the headline operation. The work does not demonstrate a computation coordinating all 240 qubits, and only a subset of qubits near the connections was wired for measurement.
"Superconducting processors have been connected for the first time using multiple quantum channels with an average transfer fidelity of 99%."—Network World, October 27
The significance sits in the scaling path. Quantum machines grow by adding qubits to a chip or by linking chips into a system, and the link has been the weak joint. IBM says the approach fits Nighthawk today and the fault-tolerant Starling machine planned for 2029. The same day, IBM’s chief executive told investors that three independent demonstrations achieved quantum advantage on IBM hardware this past summer and that the company has signed 16 enterprise engagements this year, including AT&T and General Motors. Treat the advantage claim as a company claim from an earnings call. Treat the cables as an exhibit: photographed, measured, and described in a paper anyone can read.
OpenAI reportedly plans to buy Google’s chips
OpenAI plans to buy Google’s Tensor Processing Units, The Information reported on Tuesday, according to Reuters, citing a person familiar with the matter. Terms and volumes are undisclosed. Google did not respond to Reuters outside regular hours and OpenAI declined to comment, so this item runs as a sourced report, with that status stated plainly. OpenAI has rented Google TPUs since June 2025, made its frontier models available through Google Cloud last week, and unveiled its first custom chip with Broadcom in June, planned for deployment by the end of this year. Alphabet reports quarterly results later on Wednesday. Google began selling TPUs directly this year, expects to recognise a small share of TPU sales revenue by the end of 2026, and expects the vast majority in 2027.
The board reads this as a supply-chain story wearing a rivalry costume. Search competitors can still do serious business when the scarce input is compute: Amazon, Microsoft and Alphabet plan as much as $600 billion of combined AI infrastructure spending this year, and Anthropic’s own S-1 describes a four-family chip strategy spanning Amazon Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. Nvidia keeps the centre of the market; the edges now route around it wherever price or availability demands. For a business owner, the downstream effect is the one that counts—more silicon from more vendors means steadier capacity behind the APIs and answer engines that client work already depends on.
Genie 4 gives the world model a voice
Google DeepMind announced Genie 4 on October 27: a real-time interactive world model that generates native audio—music, ambient sound and interactive dialogue—alongside the scene it simulates. Users steer with keyboard controls and rewrite the world in natural language as it runs, changing weather, camera or cast mid-simulation, while Gemini models watch the frames, break goals into steps and keep the sequence coherent. DeepMind reports stronger physics, more natural human motion, and experiments pairing Genie 4 with Gemini Robotics on a simulated humanoid checkers game. Access opens as a research preview for trusted testers, then reaches Google AI Ultra subscribers in Project Genie and early-access partners through the Gemini API and Google AI Studio. Every generated video and audio track carries a SynthID watermark.
"Understanding and simulating the world is a crucial aspect of building AGI."—Google DeepMind, October 27
DeepMind states the limits in its own post: coherence holds over short sessions, object permanence drifts over time and after repeated edits, and legible text, spatial accuracy and long-term consistency remain open research problems. Negroponte’s chair takes the demo seriously anyway. A world you can enter, redirect and hear is a training ground for robots and agents, a pre-visualisation stage for filmmakers, and a way to rehearse a flood or a factory floor before either exists. The honest label is preview. The direction is unmistakable: simulation is becoming a medium, with a reasoning model directing and a world model rendering, frame by frame.
Anthropic takes Claude to every level of government, at 40% off
Anthropic expanded its General Services Administration agreement on October 27: a 40% discount on Claude Enterprise, Claude for Government and the Claude API, extended for the first time to state, local, tribal and territorial governments, with no minimum spend. Anthropic estimates the agreement could save government organisations $1 billion. About 500 US government organisations now use Claude; weekly Claude for Government active users are up about eightfold since February, and federal civilian usage has grown about twentyfold in a year. The company cites Utah giving Claude Code to 34 developers across 10 state agencies—80% had a working environment within 30 minutes and 71% reported saving more than 20 hours in a four-week pilot—and a state team that moved nearly 900 API methods to a new framework in two days against a month-long scope. Claude for Government reached general availability on September 30 at FedRAMP High authorisation.
"Those organizations run some of the most important systems in the country, usually on tight budgets and with long procurement cycles."—Anthropic, October 27
Gigi Sohn’s chair and Lessig’s chair read this item together. A discounted, authorised model inside records, forms, backlogs and legacy code becomes civic architecture: the defaults inside that deployment shape how the state writes, decides and remembers. The reported counterpoint from Europe sharpens the question—Semafor reports the Netherlands plans to eliminate 25,000 government jobs, more than 15% of its civil service, as part of savings associated with AI adoption. One government buys the tool and counts saved hours; another counts saved positions. Both numbers deserve verification as they land, and the difference between them is a policy choice, made in the open, about what automation is for.
What the day adds up to
Wednesday’s through-line is a division of labour coming into focus. Models generate—proofs, worlds, code, dialogue—at a pace no institution matched a year ago. Humans, checkers and cables do the slower work: Lean and Rocq test the logic, outside professors test the meaning, a refrigerator full of superconducting cable tests whether two quantum chips can act as one machine, and a procurement agreement tests whether a republic can adopt the tool without surrendering judgment to it. For a business reading at 7 AM, the practical residue is concrete. The answer engines and office tools on your desk now run on a diversifying chip supply, which steadies them. The libraries underneath your software may soon carry machine-checked evidence, which hardens them. And every impressive claim arriving this season deserves the question Wednesday’s scientists asked first: show me the proof, and show me who checked it.
Sources: OpenAI three proofs via The Neuron, October 27, 2026, and Kim KJ briefing, October 28, 2026; Anthropic and Theorem verification via The Neuron, October 26, 2026; IBM quantum interconnects, Network World, October 27, 2026, and XenoSpectrum preprint summary, October 27, 2026; OpenAI TPU plan, Reuters via BusinessWorld, October 28, 2026; Genie 4, Google DeepMind, October 27, 2026; Anthropic GSA expansion, Anthropic, October 27, 2026; Netherlands civil-service plan, Semafor via The Neuron, October 28, 2026. Doodle by Eunice.