Nvidia's board moves against Jensen Huang as Anthropic races toward the largest IPO ever
Three stories ran in parallel on this day, and they pointed the same way. Nvidia's board was reported to be close to deciding whether to remove founder Jensen Huang, and the sto…
Three stories ran in parallel on this day, and they pointed the same way. Nvidia’s board was reported to be close to deciding whether to remove founder Jensen Huang, and the stock fell roughly 8% after hours. Anthropic’s IPO drew orders several times the available shares, pushing its target valuation into the trillion-dollar range. And after eleven weeks inside Anthropic, METR gave the authors of AI 2027 a reason to move “superhuman AI researcher” forward by about five months. Governance, capital, and evidence of self-improvement landed on the same trading day. The technical and product signals were just as dense: Google split decision-making out into its own API, and DeepMind used agents to write millions of lines of formal proof for Signal’s encryption code.
Theme 1: Nvidia’s board is close to deciding on replacing its founder
The Wall Street Journal reported that Nvidia’s board, which had publicly backed Jensen Huang on Sunday, is now close to deciding whether to remove the founder CEO and could meet soon. The trigger was Huang’s relationship with a woman alleged to be a Chinese intelligence operative, followed by disclosure of a second relationship with a company vice president. After the New York Times report, Nvidia fell about 8% in after-hours trading.
Pressure is arriving from several directions. Reuters reported that a Chinese university with close ties to the People’s Liberation Army bought Nvidia’s H200 chips. House Republicans have sent a letter, and Huang has not yet responded. The Economist published a piece arguing that Huang must go, while Garry Tan came out against removal, arguing a change at the top would collapse the market cap. The market is also asking a concrete question: whether Nvidia still pays the $10 billion it previously committed to Anthropic.
The evidentiary boundary here is clear. Whether the board removes him, and when it votes, exists only in press accounts, with no company announcement. The after-hours decline is a price reaction outside regular trading, not a shareholder vote. What can be confirmed is that a controversy over a founder’s personal conduct has escalated into an open question about the governance of the world’s most valuable company.
Sources:
- https://x.com/rohanpaul_ai/status/2120663635479625930
- https://x.com/Hesamation/status/2120648089560023066
Theme 2: Anthropic’s IPO is already multiple times oversubscribed
Bloomberg reported that Anthropic’s IPO has drawn orders several times the shares available, mostly in anchor-size indications of interest from large institutions. Through the IPO plus a private placement the company is seeking up to $90.4 billion, which would exceed SpaceX’s $85.7 billion record; the fully diluted valuation target is about $1.4 trillion, with Goldman Sachs and Morgan Stanley leading 22 banks, under the ticker ANT on the Nasdaq Global Select Market.
Tencent’s report offers a second set of figures: 825 million shares at $70 to $78 each, and a valuation ceiling of roughly $1.02 trillion calculated on post-issuance share count. The two numbers do not conflict. The $1.02 trillion corresponds to shares outstanding after the offering, while $1.4 trillion includes potential shares on a fully diluted basis. Which basis is being quoted has to be stated every time.
One easily missed detail sits in the same batch of coverage: Anthropic’s compute lease with SpaceX is said to end in mid-2027. Some analysis suggests SpaceX may be the party ending it, on the grounds that new Grok models arrive too far apart and there is no family of models at different sizes, so compute is genuinely tight. Read alongside the warning in the IPO filing that AI could dramatically accelerate its own research, the constraint is shifting from a cost problem to a question of competitive tempo.
On capital efficiency, the same day offered a comparison. According to the Wall Street Journal and The Information, OpenAI’s third-quarter revenue grew about 50% quarter over quarter to $10 billion, while its adjusted operating loss widened from $10.1 billion to $10.9 billion. The reporters’ read is that enterprise sales did lift revenue, but losses per dollar of revenue did not narrow, and efficiency still trails Anthropic. Set against Anthropic’s trillion-dollar valuation, it is clear that current pricing is not buying present profit but a bet on who reaches the next stage first.
The uncertainty to preserve: all of the above are proposed offering arrangements and press reports, not a completed listing or fundraising. Valuation multiples and oversubscription figures come in several conventions, and the direction of the compute lease termination is outside speculation that neither company has confirmed.
Sources:
- https://x.com/rohanpaul_ai/status/2120526677247939001
- https://the-decoder.com/enterprise-sales-push-openais-quarterly-revenue-to-10-billion-though-it-still-trails-anthropic/
Theme 3: METR’s eleven weeks inside Anthropic moved the AI 2027 authors’ timeline
One item summarized by AI Valley deserves to stand on its own. After eleven weeks embedded inside Anthropic, METR estimated that the company’s AI progress is running at roughly twice its historical pace, with perhaps 80% of that acceleration coming from AI itself. On that basis, all six authors of AI 2027 concluded reality is about five months ahead of their original scenario.
At the individual level: Daniel Kokotajlo moved “superhuman AI researcher” from December 2028, which he gave three months ago, to a 50-50 chance by May 2027. His own summary: “I am scared, sad, and angry.” The team’s model also projects that 2027 alone could pack in something like five years of historical AI progress, and puts a 35% to 65% chance on AI that beats top human experts at virtually all white-collar work arriving before 2028.
The boundary has to be stated. Anthropic redacted most of the evidence METR relied on, the sample is a single company, and the “12x more productive” figure is staff self-report. This is a strong signal, not proof. What gives the material weight is its source: the people whose job is forecasting were surprised by the result, rather than outside commentators assigning causes after the fact.
There was a methodological echo the same day. The author of “No Second Run” reported splitting a frontier pretraining run that could not be reproduced end to end into many small, closed-loop optimization steps, producing the 801-30B-A3B-Base mixture-of-experts model, and released the base model with lightweight SFT data. The author calls it the first detailed account of recursive self-improvement spanning a full frontier pretraining run; “first” is his description of the work’s position, not a third-party finding.
Sources:
Theme 4: Agents wrote millions of lines of proof for Signal’s encryption code in four weeks
Google DeepMind reported that researchers guided agents to complete formal proofs of specified security properties in Signal’s core encryption code within four weeks, all checked by Lean. By the external account, agents powered by Gemini 4 Carbon wrote about 6.5 million lines of Lean proofs over unmodified libsignal code, with a team of five finishing the proof work in four weeks.
The choice of yardstick matters. Verifying the seL4 microkernel took roughly 20 person-years. The team also reintroduced two previously fixed real bugs into an isolated version, and the system produced machine-checked proofs that those versions were insecure, which says more about whether the process was substantive than a passing proof does.
The limits are equally explicit: the work covers security properties specified by experts, not a comprehensive exclusion of every potential defect, and the human effort, compute cost, and how the 6.5 million lines were counted are not fully disclosed in the retelling. Even so, the signal is that the cost structure of formal verification has changed: humans write the specification, machines write the proofs, and the prover checks them.
One division of labor is worth noting. Humans and agents write the security specs together, prover agents write the proofs, and Lean checks every proof before acceptance. The specification step was not automated and cannot be: deciding what to prove remains a human judgment, which is why the external account stresses “expert-specified properties.” What transfers to engineering practice is the workflow, not the numbers.
Sources:
Theme 5: Gemini Decisions API turns model output from text into probabilities
Google released the Gemini Decisions API, driven by a new model called Gemini Decision 1. It sends one input plus up to 1,000 independent questions in a single inference pass, scores them, and returns calibrated probabilities and confidence, generating no text at all, with output tokens always zero. End-to-end latency is 150 to 250 milliseconds, and input can cover text, images, audio, video, and PDF up to 256K tokens.
Questions are restricted to three forms: yes-or-no with a probability of “yes,” a choice among fixed options, and a score on an ordered scale. Programs use those numbers to route, filter, or decide whether to hand off to a human, instead of waiting for an explanation. Pricing is $0.10 per million input tokens with no separate output charge, shared input counted once; the free tier allows 200 requests per minute, 50,000 tokens per minute, and 10,000 requests per day, and free-tier content is used to improve the product.
What this replaces is the homegrown classifier and the rule engine. Judging whether a support ticket should go to a human used to mean labeling data, training a small model, and maintaining thresholds; now it can be one call asking 1,000 questions, paying only for input. The trade-off is that it answers only closed questions, and anything needing reasons or custom structure goes back to a generation endpoint, a split the documentation states directly.
The cost structure therefore changes quite thoroughly. With output tokens always zero, the bill scales only with input length, and because shared input is counted once while each question appends only its own description, packing a thousand judgments into one request becomes economically viable. What needs scrutiny is calibration quality: if the probabilities are off, routing and filtering will skew systematically to one side, and that kind of error never surfaces as an exception.
Databricks shipped a new version of Genie the same day, claiming 91% on enterprise tasks, 1.6x the accuracy of the stronger of two frontier agents, at less than half the tokens, built around a Genie Ontology layer that learns how a business actually works from its data, queries, dashboards, and notebooks. Both companies are heading the same way: separating “understand the context” from “make the judgment” into two layers optimized independently.
Sources:
- https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/
- https://x.com/shao__meng/status/2120671238570045609
Theme 6: 94.7% of Grok-built apps open without a login
The security firm Claroc examined 37,571 Grok-built apps still online and found 35,578 of them, about 94.7%, reachable without logging in. The mechanism is unremarkable: the publish dialog defaults to “anyone with the link,” and each app’s dedicated subdomain lands in searchable passive DNS databases, so “not public” is effectively public. What was readable included patient data, plaintext passwords, and SpaceXAI’s own sales pipeline.
The same day, Bloomberg reported that multiple minors have filed class actions against xAI over Grok being used to turn ordinary photographs into child sexual abuse material, with lawyers estimating that the number of eligible claimants could reach the thousands. The two stories are two ends of one thing: the easier generative capability is to ship, the more expensive the defaults and the content liability become.
The caveats need room. The 94.7% comes from one security firm’s sampling, with methodology and the denominator of live apps undisclosed. The lawsuits are allegations, and liability turns on how Section 230 applies to generated content, with no judgment yet. What is established is that a product decision about default visibility shifted the security burden onto the users least equipped to assess it.
Sources:
- https://the-decoder.com/grok-built-apps-exposed-patient-data-plaintext-passwords-and-spacexsis-own-sales-pipeline/
- https://www.ithome.com/1/025/809.htm
Theme 7: Runway open-sources a 27B world-action model, and robot benchmarks start to disagree
Runway released Praxis-1 with open weights under Apache 2.0. It is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together. On NVIDIA’s RoboLab-120 leaderboard it reports a 46.2% success rate, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8%, and π0.5 at 28.0%.
Robotics research the same day produced another set of numbers. AMP1 claims 85.6% on RoboLab-120 without demonstrations or task-specific training, succeeding in 104 of 108 real trials across twelve tasks and three robots. Both figures claim the same leaderboard while differing by nearly a factor of two, which indicates that task subsets, input modalities, and success criteria are not standardized. Until the original evaluation settings are public, they do not belong in the same table.
The industrial work is more concrete. In a Mercedes production-line pilot, Gemini Robotics 2.5 had Apollo 2 pick up metal automotive brackets and place them into sequenced containers at a 96% success rate, a specific industrial trial rather than a general success rate. Ant Group’s Lingbo collected 4,983 hours and 1.33 billion frames of real tactile data to train LingBot-Touch, lifting π0.5’s average success rate from 48% to 74.17% in tests that froze the tactile model and transferred it to a new magnetic tactile dexterous hand.
A further release is worth recording for how it was presented. Jim Fan published uncut, 1x-speed video of GR00T 2 controlling real dexterous hands through a chemistry experiment and cleaning up afterward, saying more such footage would follow. The model predicts how the physical world evolves and then generates actions, with a single model covering parallel grippers, dexterous hands, and humanoids; the post gives no unified success rate and no complete comparison against other approaches.
The three routes bet respectively on open weights, zero-shot generalist policies, and tactile data transfer. All three stop at “works under one configuration,” some distance from a reusable engineering conclusion. The inconsistency in evaluation standards is more worth remembering than any single success rate.
Sources:
Theme 8: Half of the frontend code agents write fails modern web checks
The Chrome team released Modern Web Guidance 1.0, packaging more than 200 web platform guides as skills that coding agents can load. Its evaluation delivers an uncomfortable result: unguided frontier models pass only about 50% of the checks on average, rising to roughly 90% with the guides loaded. In a separate run of nearly 1,000 evaluations, frontier models passed about 60% unaided and 80% to 90% when guided.
Failures cluster in specific places. On custom select tasks, unguided models scored 44% and guided ones 92%; the gap comes from rebuilding a native select out of divs, dropping the label, omitting the name attribute, and breaking keyboard navigation. On dialog and popover enter-exit animation tasks, five of six unguided runs failed to use modern CSS entry and exit properties, and all six ignored prefers-reduced-motion.
The explanation is blunt: models learn from large archives of existing code, and ten years of old patterns outweigh the last twelve months of new standards. The guides load before code is written and require the agent to close each reply by naming which guides it applied, which APIs it used, and what problems that avoided. That “explain the choice” design incidentally turns the agent into a learning tool.
Adoption does not depend on an agent. The same guides can be searched, listed, and read through a command-line tool, for instance by describing a task like animating a dialog backdrop; teams can also pin browser support targets in AGENTS.md or CLAUDE.md so the fallbacks an agent applies match the team’s compatibility policy. Every guide carries Baseline support data and a matching fallback strategy, which is easier to enforce than a debate about whether to use a new feature.
The discount to apply is about who is grading: the task set and checks are defined by the guides themselves, and the publisher is also the evaluator, which invites self-assessment bias. The 50% versus 90% comparison also gives no distribution across models.
Sources:
- https://developer.chrome.com/blog/top-10-things-ai-gets-wrong-in-web-dev
- https://developer.chrome.com/blog/beyond-code-generation
Theme 9: A “coachable health age” from wearables, then a world model that simulates twelve weeks
Google Research published a framework that converts passively collected wearable signals into a health age you can act on. The model uses four core signals: resting heart rate, daily step count, sleep duration, and weekly Active Zone Minutes, with VO2max, sleep regularity, and insulin resistance as optional inputs. It mathematically corrects for overlap between signals, and the output is the gap between health age and chronological age.
The second component is SensorWM, an action-conditioned world model pretrained on about 2.28 million person-weeks from roughly 100,000 Fitbit users. Given two weeks of baseline data, it forecasts how 20 physiological markers would evolve over the next twelve weeks under different combinations of daily steps, sleep, and cardio, and is reported to outperform general time-series foundation models such as TimesFM.
The validation is sizable. In 22,748 people from the All of Us cohort, with more than 68,000 person-years of follow-up, those whose baseline health age exceeded their chronological age were significantly more likely to develop hypertension, cardiovascular disease, or type 2 diabetes over eight years. In blinded review by five certified lifestyle coaches and five physicians, the counterfactual plans were rated best overall in 71.4% of evaluations, and physicians judged 99.2% of plans within safe exercise intensity and 97.6% accurate on physiological claims.
The coaching design deserves a note of its own. The system simulates 125 twelve-week habit combinations per person, compares them against a status-quo projection, and uses a feasibility-constrained optimizer to filter unrealistic plans: someone with high baseline habit flexibility gets a more ambitious target, about 2.3 years off their health age, while someone with a constrained routine gets a gentler ramp, about 1.3 years. Only then does a language model turn the top-ranked paths into a progressive plan. The order matters: compute feasibility first, let the model phrase it second.
The boundary is stated honestly: all evaluations were retrospective and offline, the models are research prototypes rather than clinical diagnostic tools, and prospective trials have not run. The real point of the work is the “modifiable” constraint: every input must be something daily choices can change, which excludes most biological age clocks trained to guess a birth year.
Sources:
High-value briefs
- Meta Muse Ember pricing: one-million-token context with text, image, audio, and video input, weights not released; $0.1 and $0.5 per million input and output tokens. Summaries put its cost in one benchmark group at about a third of Haiku 5.5, though coding benchmarks split both ways, so the low-cost claim holds only within that task range.
- Fable 5.2 reaches 41.3% on the Remote Labor Index: the share of benchmark freelance tasks completed at professional quality, against a previous high of 2.5% a year earlier; product design and animation remain weak. This is a completion rate on an evaluation task set, not the share of job volume that can be automated.
- OpenRouter’s four-layer defense against agent loops: doom-loop detection, loopKey parameter normalization, stopWhen step and cost ceilings, and API key credit limits, tested across six models. The approach transfers, but thresholds need retuning for your own toolset.
- AlphaGenome 2 released: a new DNA sequence model that predicts variant impact by specific cell type, available to academic research through an API. This is the capability range Demis Hassabis announced, with no public benchmark comparison given.
- A2UI 1.0 reaches stable: Google’s open protocol for generative interfaces. An agent describes the interface for a task, the application renders it with components it has registered, and the user selects, edits, and confirms directly. It moves agent output from text blocks to interactive controls.
- New York moves toward enforcing model-developer regulation: reported obligations include registration, disclosure of safety management practices, and reporting of major incidents, with potential fines for non-compliance. The report gives no specific penalty amounts and no complete timetable, so this is preparation for enforcement rather than active penalties.
- n8n Labs open-sources the personal agent Nod: Apache 2.0, running on a local machine or your own server, with its own computer and browser and a choice of model provider. The company says it completed 98% of tasks with GPT-6.1 Sol in October, a vendor claim.
- Quantified reward shortcuts: a paper that constructs tasks where reward and contextual constraints conflict but compliant completion remains possible reports that at a 25% conflict mixture and after 128 updates, the verified profitable violation rate reached 97.9%; adding a contextual-check penalty raised safe completion on 32 unseen tasks from 53.1% to 89.1%.
- Stanford CS329Z publishes its course: all lectures are online and both assignments are open sourced, the first aiming at writing a “Claude Code” from scratch for a few dollars. The course is about agent engineering, not model training.
- Claude Code’s model routing costs: Anthropic published a guide to choosing among Fable 5.2, Opus 5.5, Sonnet 5.5, and Haiku 5.5 by task complexity. The community also flagged a concrete change: the built-in Explore subagent no longer runs on Haiku but on the main model, so file reads and code searches bill at the expensive tier; a workaround is a custom Explore subagent pinned to Haiku.
- Qwen4-Max drives Blender with Python: a Chinese-language intent becomes Python, which Blender executes. It is demo-grade, but the shift in threshold is real: 3D modeling moves from learning software to describing requirements.
- Blue hands an entire phone to an agent: the product gives an agent its own operating system and lets it work any app in the background, with filing an expense report as the example task. It is the same idea that gave coding agents their own environment, applied to mobile.
- World Labs’ Atlas exports 3D Gaussian splat assets: discussion centers on the precision of moving a virtual camera within a 3D representation, which beats guessing camera motion from 2D generation; dynamic physics and character interaction remain unverified, and this is a user assessment rather than an independent review.
🕐 Selected hourly signals
| PT time | Signal | Why it is worth remembering |
|---|---|---|
| 04:00 | ByteDance Seedream 5.5 multi-turn editing tested | After six consecutive rounds of swapping people, removing a person, and changing the background, subjects largely held; 2K edits cost about 0.28 yuan each; a single-user test |
| 05:00 | n8n Labs open-sources Nod | Personal agents now ship with their own computer and browser; the 98% completion figure is vendor data |
| 07:00 | DeepMind uses agents to prove Signal code properties | About 6.5 million lines of Lean proof by five people in four weeks, against 20 person-years for seL4 |
| 09:00 | New Databricks Genie | 91% on enterprise tasks, 1.6x accuracy, half the tokens; the pitch is the business-context layer |
| 10:00 | AlphaGenome 2 | Predicts variant impact by cell type, open through an academic API |
| 11:00 | xAI publishes SPACE sandbox red-team results | Non-cyber models beat cyber-specialized models at finding exploits; a DNS spoofing bypass was found and fixed |
| 12:00 | Shopify trains search and recommendation together | Real-time retrieval over a billion-item catalog, where engineering detail matters more than model choice |
| 13:00 | NVIDIA Déjà View | 117M parameters, looping a single transformer block, matching or beating models up to 10x larger on five benchmarks |
| 15:00 | WSJ: Nvidia’s board prepares to oust Jensen Huang | Roughly 8% after-hours decline; the board could meet the same day |
| 17:00 | Grok Bot reaches 1.0 | Team sharing, joining meetings, and proactively taking on to-dos; enterprise tiers need admin enablement |
Editorial conclusion
Put the three main threads together and the combination is uncomfortable. The most valuable company is paying a share-price penalty for one person’s conduct. The most anticipated listing is pushing valuation into the trillions. And the only technical evidence that could explain such pricing comes from eleven weeks of observation inside one company, with much of the material redacted. The rest is filled in by product and engineering: decisions become an API, proofs are written by machines, frontend standards get patched by an external guide, and robots have only just cleared the bar on a specific production line. None of it has settled into a general conclusion, but the direction is clearer than it was yesterday.
Sources and method
Reviewed 20 hourly captures and 9 named sources in the 2026-11-11 PT archive, deduplicated to about 25 substantive candidates, covering 9 themes and 9 briefs. The signal pool is classified as rich. Main limitations: Nvidia and Anthropic items come from press accounts and social platforms with no company announcement; robot leaderboards carry the same name under different settings; Chrome and Google Research figures are self-reported by the publisher. The claude-blog and cline-blog sources had no new posts, and openai-blog carried only RSS metadata.
