Today in AI — November 19, 2026

No model launch owned the day. The money did, and so did the paperwork: the largest IPO in history, a congressional document dump on chip export licences, and a lawsuit that puts a school system's AI gun detector on trial. Here is what else moved.
Industry
Anthropic is a public company, and its first trade priced it above $2 trillion. Shares opened at $111.50 on the Nasdaq under the ticker ANT, roughly 34 percent above the $83 the offering priced at, on a raise of $81.3 billion that tops SpaceX's June listing as the biggest IPO ever, CNBC reported. At the offer price the company was worth about $1.5 trillion; by the opening bell it had cleared $2 trillion. The S-1 disclosed $16.3 billion of first-half revenue against an $80.7 billion net loss that is mostly an accounting charge, and a governance structure that keeps voting control with the founders rather than the new shareholders.
Dario Amodei used the moment to restate the safety posture that the prospectus also warns about: the company will "pace the training and deployment of AI systems, taking as long as is necessary to have confidence that they are safe," and "will not bend to either business nor political pressure," he said at the Nasdaq, per Reuters. That promise is now a promise to public shareholders, which is a different kind of promise.
Google put your AI chat history on the search homepage. A collapsible sidebar on the Google homepage now lists the threads you have started in AI Mode, lets you start a new one, search old ones, or connect a Gemini Notebook, rolling out globally on desktop, per The Verge. Product VP Robby Stein said the change came from watching people arrive with questions that "aren't just one and done."
The strategic point is smaller than it looks and bigger than it reads: search used to be a transaction you completed, and the homepage was built to forget you the moment you left it. A chat list makes the front page a place you return to, which is the shape of an app, not a search box.
Meta's Muse Charm looks like a $300 gadget. The screen-and-fingerprint puck that wakes Meta's Muse agent with a tap is planned at $300 with a free 12-month Verizon 5G plan, shipping in December, according to The Information. Meta's spokesperson gave the same non-answer to both outlets — "we're charmed by the interest in our AI devices" — and the company's smartwatch has slipped to 2027 to get the Charm out first.
At $300 with data included, Meta is pricing the Charm as an accessory rather than a device, which is the only way a screen you talk to instead of unlock a phone can find a market. Whether people want a third thing to charge is the question no launch deck answers.
Policy
House Republicans published how many AI chips Commerce let go to China, and the number is 879,000. Documents released by the Foreign Affairs Committee show the Bureau of Industry and Security approved sales of more than 879,000 AI chips to Chinese companies including Alibaba and Tencent as of May 2026, out of 3.4 million requested, per the committee's release. Some approvals went to firms the Defense Department lists as Chinese Military Companies. Chairman Brian Mast named Maginfra and New Matrix — both with empty offices, per a November New York Times report — as suspected smuggling fronts: "Both are fake companies." Commerce did not comment.
The release lands on top of a scandal the committee did not need to explain: Jensen Huang's exit, the affair with a suspected Chinese spy, and the diversion concerns that followed. What the documents do is move the argument from one man's conduct to a licensing pipeline that approved a million chips while Congress was told the rules were being enforced.
Four Baltimore students sued their school system and its AI gun detector after a bag of Doritos triggered an armed police response. In October 2025, four Black students waiting for rides after football practice were flagged by Omnilert's gun detection as one of them held a bag of Cool Ranch Doritos; staff reviewed the footage and cancelled the alert within two minutes, but the principal — untrained on the system, per the state inspector general — texted a school resource officer anyway, and six patrol cars arrived with guns drawn. The students were handcuffed on the concrete before officers found no weapon, the complaint alleges. Filed November 17, the suit names Omnilert, the school system and responding officers, and asks for an injunction barring the district from using the system until protocols catch up.
The inspector general found no bias in the detector and a great deal wrong with everything downstream of it: more than 99 percent of the district's alerts were false positives requiring no police response, and the alert that was supposed to be cancelled stayed live in a text thread. That is the pattern worth watching as AI gun detection spreads — the classifier is not the failure point, the handoff is, and the handoff has no vendor.
Tools
NVIDIA and Microsoft spent Ignite arguing that the model should run where the data already is. Nemotron 3.5 arrives in Foundry Local on Azure Local, with RTX PRO 6000 Blackwell GPUs supported on-premises and Jetson Thor and Orin added at the edge; Windows 365 GPU Cloud PCs get Direct Stream previews early next year, and BioNeMo's agent toolkit is now inside Microsoft's Discovery research platform, per NVIDIA. Microsoft's own AI team is using NVIDIA's Dynamo to run reinforcement-learning rollouts during post-training.
Read the announcements together and the pitch is one sentence: the same stack from the rack to the desktop, so an enterprise can keep weights and data inside its own perimeter and still get the tooling. For buyers in regulated industries that is the only argument that has ever closed.
ChatGPT transcribes as you talk now. Live Dictation shows the transcript building while you speak, on web, desktop and mobile across all plans, and OpenAI removed the 10-minute cap on voice dictation in the same update, per the release notes. You can tap the waveform to read what it has written so far, then edit it in the prompt box.
Dictation is unglamorous and it is where voice input actually gets used — meetings, notes, the long prompt you would rather say than type. Removing the cap matters more than the streaming text: a limit you hit mid-thought is the reason people stop using voice for anything real.
PostgreSQL 19 shipped two months late, and AI bug reports are why. The release arrived Thursday instead of September, delayed by the volume of security issues the team had to work through during what core developer Robert Haas called the AI-reported vulnerability wave, Simon Willison noted. The headline feature is WAIT FOR LSN, which lets an application read from a replica only after that replica has replayed the write it just made — read-your-writes consistency without hand-rolled polling.
The delay is the more interesting artifact. Frontier models are now good enough at reading code to generate plausible vulnerability reports at a volume no maintainer team can triage, and Postgres is the first project where that shows up as a shipping date. Every open-source project with a security policy is about to find out what that queue looks like.
One question worth arguing about tonight: if AI bug reports can delay a database release, who is supposed to pay for the triage? Tell us via our contact page.







