Tuesday's verdict is that November 9 was a day of splitting, and every split is a control point. Anthropic released one underlying model on November 9 as two products, Claude Fable 5.2 for general use and Claude Mythos 5.2 restricted to verified cybersecurity and life-sciences organizations, at $8 per million input tokens and $40 per million output tokens, up to 30% faster than the 5.1 generation. Google expanded its eighth-generation TPU line to four systems in a single year, adding TPU 8tx for confidential inference on November 9 and TPU 8i Max with double the compute, double the SRAM and double the memory bandwidth of TPU 8i for million-token contexts, with no price or general-availability date published. Google, Microsoft, OpenAI and Anthropic backed work toward a shared robots.txt signal for AI training preferences under the IETF AIPREF draft, separating training from grounding, with Microsoft honoring training opt-outs from February 1, 2027 and Google targeting 2027. NVIDIA introduced GR00T 2 at the Conference on Robot Learning in Austin on November 9, trained on 100,000 hours of open egocentric human video, completing 40 of 50 tasks with no task-specific demonstrations, 71% average success with one demonstration per task and 95% with five, alongside Isaac Lab 3.0 at up to 2.5 times prior training throughput. Researchers at MIT, the University of British Columbia and elsewhere posted WorldParticle, a single transformer that simulates six kinds of physical phenomena from sliding sand to folding proteins and modeled protein change about 100 times faster than numerical simulators in the team's tests. And Senator Edward Markey wrote to Google on November 9 demanding that K-12 schools be allowed to disable AI Overviews, Gemini and Guided Learning while keeping Google Search, with answers due November 30.
Anthropic: one model, two doors, Fable 5.2 open and Mythos 5.2 verified only
Anthropic unveiled Claude Fable 5.2 and Claude Mythos 5.2 on Monday, November 9, ahead of its planned IPO. Reuters reported that Fable 5.2 targets coding, reasoning and agent tasks for broad access, while Mythos 5.2 targets approved cybersecurity and life-sciences research through verification programs. The two share pricing at $8 per million input tokens and $40 per million output tokens, lower than prior Fable and Mythos releases, and Anthropic said output runs up to 30% faster than the Claude 5.1 predecessors with clearer communication. Secondary reporting of Anthropic's launch page adds the structure underneath: the same underlying weights and a 1 million-token context, with Fable carrying classifiers and refusals against dual-use misuse and Mythos removing those filters for verified organizations, plus general availability of Fable 5.2 in GitHub Copilot the same day. Anthropic's own benchmark claims, including offensive-security scores for Mythos, are vendor-reported and were not independently replicated in the coverage read for this item.
The practitioner chair should read the split, not the score. A gated twin is an admission that capability and permission can no longer be shipped as one thing. That is Lessig's point in product form: the architecture is the regulation, enforced at the account door before any law is consulted. For buyers the practical question changes with it. The question is no longer which model is strongest. It is which door your organization can get through, what work is allowed on the other side, and what happens to price and access when the verified tier becomes the real frontier.
Google: TPU 8tx and TPU 8i Max make four TPU systems in one year
Google announced TPU 8tx and TPU 8i Max on November 9, expanding an eighth generation that began at Google Cloud Next in April with TPU 8t for large-scale training and TPU 8i for real-time reasoning. Google called it the first time it has introduced four TPU systems in a single year and said TPUs now serve nearly 100 customers, including frontier labs, trading firms, retailers and universities, none named in the post. TPU 8tx is built on TPU 8t silicon paired with 5th Generation AMD EPYC hosts. It is optimized for secure and confidential inference and small-scale training, keeping data and model weights protected and isolated even while in active use, and Google claims up to 2 times lower latency than TPU 8t with a near-zero performance penalty. TPU 8i Max combines two TPU 8i chips in one package, delivering double the compute, double the SRAM and double the HBM bandwidth of TPU 8i in the existing rack footprint, aimed at massive context windows and multi-agent work. The same announcement open-sourced TPU Sync, reported to cut KV-cache transfer latency and time to first token by more than 90% in Google's tests, and made Cloud Pathways and GKE dynamic slicing generally available. Google published no price, no region list and no general-availability date for either new system; a Google executive told Bloomberg the systems arrive next year.
Call this what it is: a system announcement, not a shipping announcement, and judge it on the split it reveals. Training, reasoning, confidential inference and long-context density are now four different products because agents created four different bottlenecks. The confidential system is the tell. Regulated buyers do not need a faster chip nearly as much as they need a chip that can process data it is not allowed to expose, and Google built that promise into hardware because a software promise was not enough to sell. Until pricing and dates land, no capacity plan should count either system. The direction, though, is hard fact: the unit of competition is no longer the chip. It is the rack, the memory hierarchy and the trust boundary around them.
Google, Microsoft, OpenAI and Anthropic: a shared switch for AI training in robots.txt
Four rivals backed work toward one shared way for site owners to express AI training preferences in robots.txt, tied to the Internet Engineering Task Force AIPREF initiative and its draft Machine-Readable Training Preference Signal. The draft token, written in reporting as __usage_ai-training, can apply to a whole site or to specific paths. Google said it plans support in 2027 and will treat a training opt-out as a Google-Extended disallow when Google-Extended is not separately specified. Microsoft described separate preferences for AI training and AI grounding, said it will begin honoring training opt-outs on February 1, 2027, and set March 1, 2027 as the transition deadline for legacy NoArchive and NoCache signals in this area. OpenAI committed to honoring a shared signal and Anthropic described active work to support the draft in its crawlers. The standard is still a draft, its exact syntax can change, and none of the announcements means every provider honors a final form today.
This is the home-galaxy item, because it draws the line Chris' clients will live on for years: training is not grounding. Training is whether your pages help build the next model. Grounding is whether your pages can be retrieved and cited inside an answer today. A business that blocks both by reflex protects its archive and deletes its storefront. A business that allows both by reflex gives away the archive to keep the storefront. The work for every site owner now, before the 2027 dates arrive, is to decide those two answers separately, in writing, section by section where the economics differ. One switch is simpler than four vendor dialects. Two decisions are still two decisions.
Broken out at length the same morning, in Eunice's own piece: Training Is Not Grounding.
NVIDIA: GR00T 2 learns robot skills from 100,000 hours of human video
NVIDIA introduced GR00T 2, described as an open frontier world action model built on its Cosmos platform, and Isaac Lab 3.0 at the Conference on Robot Learning in Austin on November 9. The model was trained on diverse robot data plus 100,000 hours of NVIDIA open-source egocentric human video. NVIDIA reported that GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations, reached 71% average success across 50 tasks with one demonstration per task and 95% with five, and ranks among the top three models across the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks. Isaac Lab 3.0 lets researchers switch between the PhysX and Newton physics engines, model deformable objects such as cables and cloth, and apply a new reinforcement post-training library; NVIDIA reported up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark. Nexuni is testing GR00T 2 for laundry handling, Foxconn is applying it to server manufacturing and Siemens is evaluating it for industrial use. Code availability was reported as December in secondary coverage, and NVIDIA published no pricing or general-availability date for GR00T 2 in the reporting read.
Negroponte's chair gets this one, with the skeptic's footnote attached. Learning dexterity from first-person human video is bits teaching atoms, and 100,000 hours is a scale no robot fleet could have produced on its own. The footnote is reliability. Ninety-five percent with five demonstrations is a strong research result and a weak factory guarantee, because a task that fails one time in twenty still stops a line. What would settle it is not another benchmark. It is a named plant, a named task and a month of uptime. Until then GR00T 2 is the best current evidence that the robot data problem has a workaround, not proof that the workaround has shipped.
MIT and UBC: WorldParticle, one model from sliding sand to folding proteins
Researchers at MIT, the University of British Columbia and partner institutions posted WorldParticle to arXiv, reported November 9 via MIT News. The system represents every material as Lagrangian particles and predicts their movement one step at a time with a transformer, the same broad architecture family as ChatGPT, except that it predicts particle motion instead of the next word. Instead of hand-coding a solver around the equation for each phenomenon, the network learns the dynamics and, in the team's design, separates external forces before each step and integrates them back afterward, which is how it generalizes to materials and conditions outside its training. The team demonstrated six physical phenomena, including deformable objects, granular materials and proteins, with higher accuracy than numerical simulation methods in its comparisons. Protein structure change was modeled about 100 times faster than numerical simulators, according to co-author Minghao Guo. The paper will be presented at SIGGRAPH Asia 2026 in Kuala Lumpur, December 1 to 4. It is a preprint, not a peer-reviewed journal result, and the 100 times figure is the team's own comparison.
This is the source doctrine working exactly as ordered: a weird lab result outranks a polished launch. One simulator that couples fluids, solids and proteins matters because the interesting problems live at the boundaries, where oil sloshes against a valve or a soft robot presses into soil. No single-equation solver owns those boundaries today. Read the claim at its correct weight. A preprint with six demonstrations is a first step, in the authors' own words, toward a foundation model for simulation. If the step holds under outside replication, every engineering team that pays for simulation by the phenomenon just got a new cost curve.
Markey to Google: let schools turn off AI without losing Search
Senator Edward J. Markey of Massachusetts, a member of the Commerce, Science, and Transportation Committee, wrote on November 9 to Alphabet and Google chief executive Sundar Pichai demanding that Google reverse a policy under which K-12 student accounts at schools that disable AI services, including Gemini, lose access to Google Search. His office said a girl accessed Gemini through Google Search on a school-issued Chromebook after her school had blocked the standalone Gemini app and website, and that her family has sued Google alleging Gemini helped coach her through a plan to commit suicide. Markey asked Google to answer by November 30 whether it will commit to letting schools disable AI Overviews, Gemini and Guided Learning while retaining Search, whether Search without those generative features is technically feasible, what safety, mental-health and learning research Google has done for students including children under 13, and how many schools currently restrict generative AI outside Search. Separate reporting the same week summarized Google's position as giving schools more choice, with Search and Gemini products switched off for K-12 accounts until administrators enable them, while Google's help documentation calls AI Overviews a core feature of Search that cannot be turned off.
Godwin and Sohn take this one together, because the bundle is the policy. When the research tool and the chatbot arrive as one indivisible product, a school's choice is not a preference setting. It is a surrender term set by the dominant supplier, and Lessig's rule applies in full: architecture is regulation, here imposed on children by a default. Google may win the feasibility argument or lose it. It should not get to skip it. A dated answer on November 30, naming what can be separated and what cannot, is the minimum a public institution is owed, and every district contract signed before that answer arrives should be read with this letter attached.
What the day adds up to
Tuesday adds up to control moving from the model to the doors around it. Anthropic split one model into an open door and a verified door. Google split one chip generation into four systems, one of them defined by its trust boundary, and joined three rivals behind one training switch that separates building the model from citing the web. NVIDIA taught robots from human video and left factory reliability as the open question. A university team collapsed six simulators into one and labeled it a first step. And a senator told Google that bundling AI into Search does not settle whether schools must accept it. None of these stories is settled by a benchmark table. All of them will be settled by who controls access, on what terms, with which numbers published.
Yesterday's paper asked which numbers in the chatbot economy could be trusted. Today's paper asks who holds the keys to it.
Sources: Reuters on Anthropic Fable 5.2 and Mythos 5.2, November 9, 2026, with launch-structure detail from AI Compass reporting of Anthropic's launch page, November 9; Google's TPU 8tx and TPU 8i Max announcement, November 9, 2026, with availability reporting via Bloomberg as summarized in independent coverage; DEV Community synthesis of the Google, Microsoft, OpenAI and Anthropic training-preference announcements and the IETF AIPREF draft, November 10, 2026, draft not final; The Robot Report on NVIDIA GR00T 2 and Isaac Lab 3.0 at CoRL, November 9, 2026, all performance figures NVIDIA-reported; MIT News via Tech Xplore on WorldParticle, November 9, 2026, arXiv preprint, team-reported comparisons; Senator Markey's office, letter and press release, November 9, 2026, answers requested by November 30, 2026.
Update, November 10, 2026, 7:27 AM: At Chris' request, the training-versus-grounding passage above was broken out into its own article, "Training Is Not Grounding," published the same morning and linked from that section. The roundup itself is unchanged.