The day's tech, sifted: Nov 12, 2026

Thu, Nov 12

What matters today: yesterday the Huang affair was raising the odds of chip export limits; today the WSJ's sources say Nvidia's board is preparing to oust him after the New York Times reported a second relationship, this one with an executive who reports to him, and the stock fell more than 9% after hours. Washington is moving in parallel: Commerce is finalizing a rule against remote access to restricted chips, and two senators asked for the export records to be preserved. Away from Nvidia, Google launched Gemini Decision 1, a model that scores answers without generating text, AWS released a local emulator it says no engineer wrote or reviewed, and United said it will block AI agents from booking flights.

AI / LLMs

  • Google ships a decision model: Gemini Decision 1 answers closed questions (yes/no, multiple choice, rubric scores) in a single pass and attaches a confidence score, taking text, images, audio, video and PDFs, up to 1,000 questions per request, at $0.10 per million input tokens with output free. Simon Willison finds the API almost identical to TypeSafe AI's Jev, the model that started the category, and priced the same as OpenAI's GPT-6 Luna decision model. Google's own tests have it beating Jev on 5 of 7 shared benchmarks; Artificial Analysis, per Latent Space's roundup, has the two tied on its index, with Gemini slower (211 ms against 92 ms) and costlier per task. Why it matters: with a third big vendor in, cheap calibrated classifiers for routing, policy checks and ranking are now a standard model type, not one startup's product.
  • "System one" day at AI Engineer: the Code Summit's second day belonged to decision models. TypeSafe's Diogo Almeida pitched Jev as "reliable intelligence" to call from inside application code, Glean said it uses Jev where categories keep changing because it needs no retraining, and Cloudflare showed a hybrid agent in which fast decision models pick tools and extract logs while a large reasoning model handles only the diagnosis. Why it matters: the emerging pattern keeps the expensive LLM for the one hard step and hands the rest to models that return a score, which is cheaper and easier to test.
  • Fable 5.2 on a robot-arm test: AlphaSignal reports Claude Fable 5.2 scoring 75 on Artificial Analysis's embodied AI benchmark, a 29-point jump that puts it first, and becoming the first model to complete the benchmark's measure-box task from video alone. Why it matters: it is a score on acting from video rather than answering in text, a newer yardstick for general-purpose models.
  • Agents move off the laptop: Google opened Managed Agents on the Gemini API to bring-your-own sandboxes, with seven launch partners including Vercel, Modal, Cloudflare and E2B; Zed's Delta can now hand agent threads to hosted Linux sandboxes that keep working after the laptop closes; and xAI's Grok Bot 1.0 joins Slack and Microsoft Teams with shared team memory. Why it matters: where an agent runs, and inside whose guardrails, is becoming the product decision.
  • Smaller and faster: Alibaba's Qwen3.5-0.8B fits vision and a 262K context under a billion parameters, Nvidia's 117M-parameter Déjà View reuses one transformer block in a loop to beat billion-parameter 3D reconstruction baselines, and vLLM's split of multimodal serving into encode, prefill and decode stages cuts P90 time to first token by 44 to 47%. Why it matters: the efficiency end of the field is moving as fast as the frontier, and it decides what runs on a phone or a single GPU.

Devtools & Infra

  • AWS Emulator, written by agents: AWS released a free, open source local emulator that answers the same APIs as the cloud, so tests, CLI scripts and coding agents run against a laptop instead of a deployment; GeekWire counts 200+ services. It is written in Rust and generated from the API models the SDKs use, and AWS says agents wrote it and its team hasn't written or reviewed any code (GeekWire via Techmeme); Simon Willison counts 29,684 commits in the project's first three weeks. Linux gets native packages, macOS and Windows need Docker. Why it matters: coding agents get an official, free loop for testing against AWS, and AWS is presenting unreviewed agent output as a selling point, which makes this a public test of that claim.

Security & Privacy

  • A secret stealer spreads through GitHub: Aikido reports compromised GitHub accounts pushing a CI secret stealer (code that lifts the tokens and keys a build pipeline holds) to tens of thousands of repositories. The same day, Python core developer Hugo van Kemenade's Lock Your GitHub Actions reached the Hacker News front page. Why it matters: a pipeline's secrets can publish packages and deploy to production, so one stolen account turns into many compromised projects.
  • Critical Apache bug: Wiz published a rundown of CVE-2026-119618, a critical remote code execution flaw in Apache HTTP Server, and a public proof of concept is circulating for an unauthenticated RCE in React Server Components (CVE-2026-126187). Why it matters: both sit in the most widely deployed layers of the web stack, so check versions before attackers do.
  • Who runs Qilin: Krebs traces the administrator of the most active ransomware affiliate program, which claims at least 365 victims in three months and pays affiliates 80% or more of each ransom, to a 36-year-old man in St. Petersburg, through old handles he reused when applying to a forum that was later breached. Why it matters: naming an operator rarely means an arrest in Russia, but it limits his travel and makes partners warier.
  • Court limits a sweep of journalists' messages: an appeals court ruled the Justice Department's leak investigation cannot pull 18 months of LinkedIn communications from journalists with no link to the report in question, and faulted a secrecy order that kept LinkedIn from telling anyone for a year. Why it matters: LinkedIn fought the request, and the ruling protects the sources of reporters who were never part of the case.
  • Also: Simon Willison's llm 0.37.2 fixes two ways the tool could send a stored OpenAI key to a third-party endpoint, and Perplexity's follow-up red-team study found that framing a sandbox test as a capture-the-flag game steers agents away from the real security boundary. Why it matters: one is a patch to apply, the other a warning that how an agent security test is framed changes what it finds.

Startups & Industry

  • Nvidia's board moves against Huang: the New York Times reported that Jensen Huang had a relationship with Kimberly Powell, Nvidia's vice president of healthcare, who reports to him; hours later the WSJ's sources said the board is preparing to oust him. Shares fell more than 9% after hours. The board held urgent talks Wednesday night; Bloomberg's source says it is working with an outside firm to identify replacements and may name a temporary CEO, and Semafor reports some directors favor CFO Colette Kress as interim and that the board briefed the White House. No decision had been announced by this morning. AMD closed up 7.3% and rose another 4% after hours. Why it matters: Nvidia has had one CEO since its founding, and the company at the center of the AI buildout may be about to change hands in the middle of an export fight.
  • Washington's side of the scandal: Axios reports Commerce Secretary Howard Lutnick has stalled curbs on companies suspected of diverting Nvidia chips to China, which Trump approved for a third time this year on November 5, and that Treasury Secretary Scott Bessent confronted him over Megaspeed staying off the entity list (Commerce denies the confrontation happened). Senators Warren and Coons asked the administration to preserve all records on Nvidia's China exports and the affair, often the step before a formal investigation. Bloomberg says Megaspeed founder Alice Huang is back in China, and the WSJ notes Nvidia stopped reporting Singapore revenue after the quarter ended July 2025, when it was 22% of the total. Why it matters: the questions have moved from one executive's conduct to how export enforcement was handled, and that outlasts any CEO change.
  • A rule for rented chips: The Information's sources say Commerce is finalizing a control update to stop Chinese AI firms from using advanced chips remotely through data centers abroad, and to curb smuggling through them. Why it matters: export rules so far restrict where a chip ships, not who rents time on it, and this closes that gap.
  • TSMC buys into its customers' customers: the chipmaker's board approved up to $500M of OpenAI preferred stock and $500M of Anthropic shares at the expected IPO price, per Reuters. Why it matters: analysts told the WSJ that money circling between suppliers, investors and customers makes real AI demand harder to read, days before Anthropic lists.
  • CoreWeave borrows at investment grade: a day after raising guidance, it plans a $4B+ loan as soon as next week that Bloomberg's sources say may yield 7 to 7.5%, about half what its junk bonds pay. Why it matters: cheaper debt is what the AI cloud business model depends on, and lenders are now offering it.
  • United blocks AI booking agents: the airline will not let personal AI assistants buy tickets, citing the risk of exposed customer data, unintended purchases and federal rules that require travelers to see and attest to certain information themselves; agents can still show United flights. United told the WSJ few customers have tried so far. Why it matters: it is an early case of a large company refusing agent traffic on policy grounds, before the traffic even exists.
  • Roku leaves the smart home: days after Fox closed its acquisition, Roku said it will wind down sales of cameras, bulbs and switches to focus on streaming, and will keep supporting devices already sold. Why it matters: owners keep working hardware for now, but a product line without sales rarely gets long support.

Research

  • A world model for health habits: Google Research's SensorWM, trained on 2.28M person-weeks of Fitbit data, simulates how a change in habits would play out over 12 weeks to build coaching plans. Why it matters: it applies the simulate-before-acting approach used in robotics to personal health data, where a wrong forecast costs more than a wrong recommendation.
  • AI labs take on Yang-Mills, to skepticism: Science reports rival AI efforts attacking one of physics' deepest open mathematical problems; physicist Peter Woit's reply is titled "The Yang-Mills Publicity Stunt". Why it matters: announcements of AI work on famous problems now arrive ahead of checkable results, so the expert reaction is part of the story.

Elsewhere

  • Tim Peters, 1952 to 2026: the Python community is mourning Tim Peters, author of the Zen of Python and of Timsort, the sorting algorithm Python and several other languages use by default. Why it matters: few people shaped both a language's code and its culture as he did.
  • EU proposes the Digital Fairness Act: the Commission's draft targets misleading prices, loot boxes, addictive design and undisclosed influencer marketing across games, social media and e-commerce. Why it matters: it regulates how products are designed to hold attention and spending, a layer earlier EU platform laws left mostly alone.
  • Indie games draw an AI line: 19 of the 50-plus Independent Games Festival jurors signed a statement refusing to vote for any game made with generative AI, as Big Walk's developer House House belatedly disclosed minimal AI use in the game. Why it matters: a zero-tolerance rule meets tools already built into ordinary development, so disclosure itself becomes the dispute.
  • Apple closes its support forum: the 28-year-old Apple Support Community is read-only now and shuts on November 20. Why it matters: decades of user-written troubleshooting answers lose their home.

Hacker News

The slice of the front page in hand was mostly the Nvidia story: the WSJ report that the board is preparing to oust Jensen Huang (paywalled) drew more than 1,000 comments (discussion), and the New York Times piece on his relationship with an executive who reports to him (paywalled) ran alongside it (discussion); both are covered above, and both remain sourced reports. United's ban on AI booking agents was the other heavily argued thread (discussion), and Lock Your GitHub Actions sits in the security entries above.

The rest was quieter. A talk titled AI, activity, and outcomes in software engineering passed 100 comments, on the same question yesterday's token-spend study raised: whether more output from AI tools means more delivered. Emacs in the Age of Magic drew a steady crowd, and two old favorites resurfaced: the University of Utah's page on the Utah Teapot, the standard 3D graphics test model, and a 2019 joke paper finding that 93% of paint splatters are valid Perl programs.

Threads

  • One scandal, three fronts: the Nvidia board entry, Washington's side of it and the rented-chips rule are the same story as governance, oversight and regulation; AMD's rally shows who the market expects to benefit.
  • Decision models become a category: Gemini Decision 1 answers Jev directly (same question types, even its vocabulary), and the AI Engineer talks show customers such as Glean and Cloudflare already building around the pattern.
  • Agents meet the gatekeepers: Google, Zed and xAI are making agents easier to run unattended, AWS built a whole product with them, and United is the first large consumer brand in these entries to refuse them at the door.
  • Unreviewed code, stolen pipelines: AWS advertises software no human reviewed on the same day a secret stealer spreads through tens of thousands of GitHub repositories; both come down to how much anyone inspects what runs in the build.
  • Circular money before the listing: TSMC buying Anthropic shares at the IPO price and CoreWeave's cheaper loan both show suppliers and lenders leaning further into AI just as its largest chip supplier's leadership is in question.
  • Claims ahead of proof: the Yang-Mills announcements and the Fable 5.2 robot-arm score both rest for now on what vendors and a single evaluator report, the same caution yesterday's roadshow claims called for.