Thursday's verdict is that November 11 was the day the broad claims got specific. Google began rolling out its November 2026 core update on November 10 at 10:06 Pacific, a global ranking change in all languages that may take up to two weeks, which puts its latest finish on November 24, three days before Black Friday. Google DeepMind announced AlphaGenome 2 the same Wednesday: a DNA model conditioned on individual cell types and states, trained on more than 107 million single-cell profiles across 68 tissues, finding 30% of causal variants on a 28-cell-type immune benchmark against 16% for its predecessor. IBM announced Nighthawk r3, designed for 10,000 quantum operations across 520 programmable qubits, with two chips linked into a 1,040-qubit system planned for 2027. Runway opened Praxis-1, its 27-billion-parameter robot world action model, under Apache 2.0 at 46.2% on RoboLab-120. Australia's regulator defended a Digital Duty of Care bill carrying fines up to A$109 million against the charge that it will chill speech. And Tether released BotPsy, a 0.8-billion-parameter open-weight model built to do one job on a phone: call tools well.
Google: the November core update rolls out three days ahead of Black Friday math
Google released the November 2026 core update to Search on Tuesday, November 10, logged on the Search Status Dashboard at 10:06 Pacific. Google calls it a regular update to surface relevant, satisfying content from all types of sites. It applies globally, in all languages, it is the third core update of 2026 after March and May, and the rollout may take up to two weeks. That arithmetic is the story. A full rollout ends November 24. Thanksgiving is November 26, Black Friday November 27, Cyber Monday November 30, and Google's own advice is to wait at least a full week after a rollout ends before judging in Search Console, which puts a clean read on December 1, after the biggest shopping weekend of the year is already over. In March, 79.5% of top-three URLs moved, per SE Ranking data cited in coverage of this update. Google named no target, no site type, and no new guidance.
For Chris' client work this is a calendar ruling, not a panic button. Rankings that move during a rollout are weather, not verdict, and any shop that rewrites its pages this week will never know whether the update or the rewrite did it. The discipline is dull and it is the whole job: note the start date, watch Search Console without touching the merchandise, and make the Black Friday plan on the rankings you have, not the rankings you hope the update grants by November 27. The board's practitioner chair adds the sharper point. Google chose this window. A company that tells publishers to wait a week past the end knows exactly when the end lands, and it landed three days before the sale. That may be coincidence and it may be that shopping queries are where ranking quality matters most in November. Either way, the sellers absorb the uncertainty, and the house keeps the calendar.
DeepMind: AlphaGenome 2 reads DNA cell by cell, not tissue by tissue
Google DeepMind announced AlphaGenome 2 on Wednesday, November 11. The first AlphaGenome predicted what a DNA variant does averaged across broad tissues. The new model is conditional: give it a sequence of up to a million letters plus a specific cell type or cell state, chosen from a catalog or encoded from a researcher's own single-cell experiment, and it predicts gene expression, chromatin accessibility, histone marks, and DNA methylation for that cell in that state. It was trained on human and mouse data including more than 107 million individual cell profiles across 68 tissues. On TenK10K, a benchmark of 28 primary immune cell types, DeepMind reports it recovered 30% of causal expression-linked variants at high precision against 16% for the earlier model fine-tuned on single-cell data, and got the direction of the effect right 43% of the time against 23%. In the demonstration case, it singled out the variant that overactivates the ETS2 gene in gut macrophages, reproducing a 2024 inflammatory bowel disease finding whose lead researcher told Nature the model reached in 30 minutes an answer that took his team a year and a half of experiments. AlphaGenome 2 is available now through the AlphaGenome API for non-commercial research, with commercial research in private preview. DeepMind states plainly that it is research only, not validated for clinical or diagnostic use, and the paper is a preprint. Even at high precision it recovers only a minority of nominated causal variants.
This is the frontier beat at its best, and the board should say why in plain words. Most disease variants do not break a gene. They turn its volume up or down in one kind of cell, at one moment, and a tissue average washes that signal out. A model that keeps the cell in the question turns a static atlas into a predictive engine, and the ETS2 case is the right kind of exhibit: a known answer, found again computationally, with the scientist who did the wet work on the record about the time it saved. The limits ride in the same paragraph as the praise, by house rule. Thirty percent is nearly double sixteen and it is still thirty. The claim today is a research instrument that points a lab at the right bench faster, and that is already a large claim. What it does to drug discovery over five years is the open question, and the paper leaves it open.
IBM: Nighthawk r3 aims at 520 qubits, and two of them at 1,040 in 2027
IBM announced Quantum Nighthawk r3 on Wednesday, November 11, at its Quantum Developer Conference in Yorktown Heights. The design target is 10,000 quantum operations across 520 programmable qubits, with 2,034 total physical qubits on the chip. In 2027 IBM plans to link two r3 processors with chip-to-chip connectors it calls l-couplers into a single 1,040-qubit system. The current machine, Nighthawk r2, in service since September with 120 programmable qubits, has run 34 billion circuits for partners including E.ON and Q-CTRL, with reported speedups of up to 19 times over earlier IBM processors. IBM says it has linked r1 processors with four l-couplers at 99.3% state-transfer fidelity, against a 99.9% target, and connected two cryogenic modules into one system for the first time. The r3 figures are design targets, not a built processor. No error rate for r3 and no date for outside access were given. The fault-tolerant Starling system remains the 2029 target.
Negroponte's chair and the Slashdot chair split this one cleanly, and both halves print. The long view is that the unit of progress just changed. Quantum roadmaps used to be one bigger chip. This one is a linked system: prove the coupler, then scale by joining processors the way classical computing learned to scale by joining machines. IBM has shown the links work at 99.3%, which makes 1,040 qubits a build schedule rather than a wish. The skeptic's half is just as plain. A roadmap announcement is not a computer, 520 is a target, and a field that has promised fault tolerance by the end of the decade for several decades owes its readers the word target in every sentence. Chris' readers should track the coupler fidelity number, 99.3 toward 99.9. That gap, not the qubit count, is where this plan lives or dies.
Runway: Praxis-1 robot model ships open, weights, code, and a commercial license
Runway made Praxis-1 generally available on Wednesday, November 11, with open weights under the Apache 2.0 license, including for commercial use. Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, built on the large-scale video pretraining behind Runway's world models: it imagines the next seconds of a scene and derives the movement that produces them, 24 action steps at 15 hertz per call. Runway reports 46.2% success across 1,200 simulated episodes on NVIDIA's RoboLab-120, 555 of 1,200 tabletop tasks, ahead of FLUX 3 Action at 42.9%, and 80.0% on SimplerEnv after fine-tuning on a robot it never saw in training, against 52.4% for the original policy. The release is unusually complete: three checkpoints, inference code, full and LoRA fine-tuning code, on GitHub and Hugging Face with native LeRobot support, serving through vLLM-Omni, running from H100 and B200 data-center cards down to an RTX 5090 workstation and the Jetson Thor computer inside the robot. Three launch partners have run it on real machines, a legged humanoid, a six-axis industrial arm, and a dual-arm warehouse robot. Every benchmark figure is Runway's own, not independently reproduced, and 46.2% still means failing more than half the time, in simulation, on tabletop tasks.
Lessig's chair claims this one, because the license is the architecture. A robot brain you may download, fine-tune on your own machine, and sell work built on is a different industry from a robot brain rented by the call from three labs. NVIDIA opened GR00T 2 days earlier and Google's Gemini Robotics 2.5 is already working a Mercedes-Benz line, but Runway's entry is the most permissive of the three, and permissive compounds: every lab that fine-tunes Praxis-1 on a new embodiment teaches the commons something. The caution belongs to the practitioner chair and it is physical, not legal. Open weights do not make motion safe. A model that predicts plausible actions can still drop, crush, and misjudge, and it now ships to anyone with a workstation card. The demo-or-die test for 2027 is whether those partner machines are still running it when the cameras leave.
Australia: the regulator says a duty of care will not chill speech, and fines up to A$109 million say otherwise
Australia's eSafety Commissioner told a Senate inquiry on Tuesday, November 10, that warnings about the government's Digital Duty of Care legislation were wrong, Reuters reported. The bill would require internet companies to proactively check for and stop potential user harms or face fines of up to A$109 million, about $75 million. It has passed the House of Representatives and needs crossbench support in the Senate. The government wants it passed by the end of the year. Tech companies and free-speech advocates argue the duty gives regulators too much say over what counts as harmful and will push platforms to remove lawful content rather than risk the fine. The regulator's answer was that platform business models reward engagement, not caution. The bill would also give users a right to opt out of algorithmic recommendation and see only accounts they chose to follow. A section requiring age checks where a child might see pornography drew specific questions from Google's Australia executive at the inquiry.
Godwin's chair was built for exactly this hearing, and his test is one sentence long. Judge a speech rule by what a careful company does the day before anyone is fined. With A$109 million behind a duty to stop potential harms, potential doing heavy work, the careful company removes first and reads the appeal later, and it removes the borderline political speech, the satire, and the awkward lawful post long before it removes anything a regulator named. The opt-out of the algorithmic feed is the one genuinely pro-user clause in the bill, and Gigi Sohn's chair would keep it and strike the rest. The engagement argument cuts the other way than the regulator thinks. A business model that maximizes engagement under a vague harm duty does not become careful about speech. It becomes careful about liability, and those are opposite instincts wearing the same suit.
Tether: BotPsy puts a 0.8-billion-parameter tool caller on the phone itself
Tether AI Research released QVAC BotPsy-0.8B on Wednesday, November 11, an open-weight model of 0.8 billion parameters built on Alibaba's Qwen3.5-0.8B and specialized for one layer of work: function calling and multi-step tool use on personal devices, across calendars, messaging, files, settings, web research, travel, and commerce. In Tether's own model-card evaluation it roughly doubled its base model on agentic benchmarks, BFCL v4 from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and ToolSandbox from 30.5 to 64.6, beating the larger Qwen3.5-2B-Think on five of six. The GGUF release for llama.cpp runs to about 0.53 gigabytes in its smallest quantization. The limits are printed on the card: English-first training, unsafe responses in 21% of automated red-team cases against 36% for the base, and an explicit warning against safety-critical, financial, health, or other sensitive automation. Training code and data were not released.
The practitioner chair loves the shape of this even while discounting the scoreboard. The frontier race assumes the agent lives in a data center and rents its brain by the token. BotPsy argues the opposite for the routine layer: a small, specialized model on the device, free to run, private by construction, good enough to book, file, message, and fetch. Specialization beating scale at one named job is the oldest story in computing, and 0.8 billion parameters doing it at half a gigabyte is a real exhibit, vendor benchmarks and all. Tether's longer design hands hard tasks up to a home PC or trusted peers and routes web queries through encrypted peer-to-peer connections. Whether that network arrives is an open question. The model is here today, and anyone can test the claim on their own phone this morning, which is more verification than most launches offer.
What the day adds up to
Thursday adds up to specificity beating scale as the day's method. Google's update is not a new philosophy of search; it is a dated change with a two-week window that collides with a dated sale. AlphaGenome 2 did not get bigger so much as it learned to ask which cell. IBM's road to a thousand qubits is not a bigger chip but a coupler with a fidelity number attached. Runway's contribution is a license and a download, precise enough to build on. Australia's argument turns on one word, potential, inside a duty backed by a specific fine. And the smallest model of the day won its benchmarks by refusing every job but one. The broad claims will keep arriving by the truckload. The reader's edge is the habit this paper keeps: ask what exactly, in which cell, by what date, under which license, at what number. November 11 rewarded every version of that question.
Yesterday's paper watched AI move into the institutions. Today's paper watches the claims get specific enough to check.
Sources: PPC Land and Search Engine Journal on Google's November 2026 core update, dashboard timing, and Black Friday arithmetic, November 10, 2026, with SE Ranking movement data from the March update as cited; Google DeepMind, AlphaGenome 2 announcement and preprint, November 11, 2026, with Nature reporting on the ETS2 case and its research-only status; IBM Newsroom, Nighthawk r3 announcement, November 11, 2026, with Quantum Computing Report and Ars Technica coverage of the r2 record and the 2027 linked system; Runway, Praxis-1 availability post and research report, November 11, 2026, benchmark figures company-reported; Reuters on the Australian eSafety Commissioner's Senate testimony and the Digital Duty of Care bill, November 10, 2026; Unite.AI on Tether's QVAC BotPsy-0.8B release and model card, November 11, 2026, evaluation figures company-reported.