Source: The AI Daily Brief, “Wall Street Has No Idea What Anthropic is Worth” (raw/Wall_Street_Has_No_Idea_What_Anthropic_is_Worth.md, https://www.youtube.com/watch?v=gCOMQPMhIL0; sponsor reads skipped) · Matt Wolfe, “AI News - Things Will Never Be The Same”, recorded Thursday 2026-11-12 (raw/AI_News_-_Things_Will_Never_Be_The_Same.md, https://www.youtube.com/watch?v=G_Jrexy_5Bg; the OpenAI-sponsored Decisions API segment is labelled as such below) · Two Minute Papers, “DeepMind’s AI Just Did The Impossible” (raw/DeepMind_s_AI_Just_Did_The_Impossible.md, https://www.youtube.com/watch?v=gPNiuyYirfw; sponsored by Lambda) · The Neuron Daily, November 12 (raw/newsletter-theneurondaily-com-10768b7925.md). First-party and press checks fetched 2026-11-13: Anthropic’s S-1/A Amendment No. 2, the FT report via Benzinga, TechCrunch on OpenAI’s revenue and on Odessia, The Verge on United, Ars Technica on Gemini 4 Hydrogen, and Nvidia’s press release (files listed in frontmatter).
Creator: Nathaniel Whittemore (AI Daily Brief), Matt Wolfe, Károly Zsolnai-Fehér (Two Minute Papers), The Neuron · Platform: YouTube, newsletter
The three days after the previous roundup. Jensen Huang resigned from Nvidia on November 12; Anthropic started its roadshow at 10B in a quarter; and United Airlines barred personal AI agents from booking its flights.
Key Takeaways
- Jensen Huang resigned as Nvidia’s CEO and director on November 12, effective immediately. CFO Colette Kress is interim CEO and Harvey C. Jones is the board’s first chair (Nvidia press release). Wolfe adds two reported relationships with Nvidia employees, one with a direct report, on top of the Megaspeed affair. Detail: Nvidia, Megaspeed and Export Controls.
- Anthropic’s IPO may price above its range. The FT (via Benzinga) reports underwriters are weighing a price above 1.5T, after strong demand. Three backers told the FT that Anthropic cares more about raising money for compute than about valuation. Filing detail: Anthropic’s Public S-1.
- Gemini 4 Hydrogen replaces 3.8 Flash. 7.50 per million tokens, four thinking levels including a new xhigh, and 54 on the Artificial Analysis index at xhigh against Argon’s 50 (Ars Technica). It now runs AI Overviews and AI Mode for all users.
- Carbon tests well outside Google. Pavel Horan’s bug hunt: Carbon found 57 of 105 bugs for about 140 (relayed by NLW). Wolfe’s SVG test: 46 cents against Fable 5.2’s $5.
- **OpenAI’s Q3 revenue passed 6.7B in Q2, with adjusted operating margin narrowing from -150% to -109% (TechCrunch, relaying Bloomberg). Anthropic’s Q3 was $17.0B (S-1/A).
- United now bans personal AI agents from booking. Agents may show United flights but not buy them; United added the rule to its Contract of Carriage, which lets it cancel AI-made bookings (The Verge; other outlets’ snippets). Marketers who sell through agents should read the agent-commerce terms-of-service ceiling.
- AI agents proved Signal’s encryption code correct. DeepMind’s team produced 6.5 million lines of Lean proofs in four weeks with five scientists, and had to add referee agents after an agent rewrote a theorem to make it trivially true (Two Minute Papers).
Nvidia: Huang out
- The resignation (Nvidia, November 12): Huang leaves as president, CEO and board member, effective immediately. Kress stays CFO while serving as interim CEO; Jones, a former Synopsys CEO, becomes the first chair of the board.
- Before the resignation (NLW, recorded earlier in the week): President Trump said Huang is “a very honorable man” and acknowledged knowing of the affair before it was reported. The Information’s sources put up to 7 billion plus”. The New York Times reported that a Commerce Department investigation of Megaspeed stalled after contact from Nvidia’s general counsel. The House Foreign Affairs chair, Brian Mast, threatened a subpoena. Nvidia’s stock was down 11% from the first report (NLW); Nvidia is about 7% of the S&P 500 and about 8% of the Nasdaq by weight.
- Wolfe’s chip-flow figures: Megaspeed’s Nvidia imports totalled $4.6B from 2023 to 2025, with some chips reaching Century Litong and then data centres serving Alibaba and Tencent. He also notes Nvidia stopped breaking out Singapore revenue when it switched to reporting by customer headquarters.
- Wolfe’s guess (opinion): “maybe like 50/50” that Nvidia asks Huang back within months if the stock struggles.
- Small item: Booz Allen adopted Nvidia’s Nemotron open models (NLW).
Succession, the board’s options and the export-control thread: Nvidia, Megaspeed and Export Controls.
Anthropic: the IPO book and new products
- Terms (S-1/A Amendment No. 2, filed November 10): 825M Class A shares at 26B private placement, with Google and Nvidia at about 2.5B, and Broadcom, Micron and Microsoft about 100B of stock floated including the option. ^[inferred] (The option alone at 9.7B.)
- Price above range? (FT via Benzinga, November 11): underwriters Goldman Sachs and Morgan Stanley are considering pricing above the range, which could value Anthropic above 1T, possibly $1.2T.
- NLW’s reading (opinion): the 1.6T and $1.1T figures in the press came from unnamed investors, not from Anthropic, so the range is not a climb-down. “Investors do not have a clear idea of what to expect.”
- The bull and bear cases he relays: John Ma (Artemis) says the consensus is that Anthropic is overvalued at 10B in October to about 1.4T is about 40× trailing revenue, against about 55× for Palantir, so closer to CrowdStrike and Cloudflare. That matches the filing’s figures.
- Compute deals (S-1/A): the Microsoft and Google minimums were raised (NLW: about 34.9B August agreement is now terminable because Lambda missed a facility-lease deadline; Anthropic says it does not currently intend to end it. NLW’s question: what fails first, AI demand or neocloud construction?
- Dario Amodei’s roadshow pitch: “an age of miracles”, with risks as the obstacle to address. Asked by Bloomberg’s Emily Chang about military use, he said military decisions are “not ours to make” beyond “very high-level boundaries” (paraphrased by NLW).
- Products: Claude Sheets, Whiteboards and Data, plus interactive connectors that embed Adobe, Atlassian, Salesforce, NetSuite and Workday inside Claude (Wolfe). NLW reports Microsoft stock fell 2.2% on the news and relays Gabriel’s point that Sheets exports to Excel and that Microsoft is an investor, landlord and reseller of Anthropic. Anthropic’s own examples: a legal-review tool in Slack that cut marketing reviews from three days to 24 hours, and a finance team’s interactive business update. See Claude Sheets, Whiteboards, Data and Interactive Connectors.
- Fable 5.2 in Copilot: it is now available in Copilot Cowork and Copilot Studio (The Neuron).
Models
- Gemini 4 Hydrogen (Google, Ars Technica): replaced 3.8 Flash everywhere; not staged through security testers first. Google claims 2.5× the speed of 3.8 Flash at high, more than 5× fewer fabricated claims, and over three times 3.8 Flash’s Terminal-Bench score. API 7.50, unchanged. Live for AI Pro and Ultra in the Gemini app and AI Studio; Ars says free users stay on 3.5 Flash-Lite, while Google says free users get it through automatic model selection. ^[ambiguous] Wolfe reads Google’s chart: about 10.04, and a 1M-token context.
- Gemini 4 Carbon, outside tests: Ethan Mollick recreated a train-ride video in three.js and called it “a very good model that improves on Argon” (relayed by NLW). Wolfe judged Carbon’s 3D game better looking than Fable 5.2’s. Outside the API, Carbon needs the 2 / 4 / $20 in December (Ars).
- Claude Fable 5.2: Wolfe calls it his “daily driver”. His SVG benchmark took almost 20 minutes and $5. He notes Artificial Analysis ran it at Max, while Anthropic recommends medium or high, so typical cost is lower. See Claude Fable 5.2 and Mythos 5.2.
- Muse Ember (Meta): Wolfe repeats 0.50 per million tokens and places it with Haiku and Luna (captioned “Muse Amber”).
- Grok 4.8: Musk says xAI moved from JAX to its own C and C++ stack and that the model is in reinforcement learning ahead of a soft rollout (The Neuron).
- MiniMax H3 Image: 4K images in about 10 seconds, on API and Hugging Face (Wolfe). Covered in the previous roundup.
Decision APIs
- OpenAI Decisions API (sponsored segment in Wolfe’s video): three answer types: predicate (a probability), choice (one of a list) and score (a rating). OpenAI’s claims as read: up to 10× faster than the Responses API on GPT-6 Luna, $0.10 per million uncached input tokens and no output charge. Background: OpenAI DevDay 2026.
- Gemini Decisions API (Google): checks a photo, PDF or video against up to 1,000 questions in one pass; The Neuron’s example graded 100 delivery photos on 16 questions in 4.5 seconds. Wolfe notes Microsoft has one too and that the category started with Jev.
OpenAI
- Revenue (TechCrunch, relaying Bloomberg): Q3 above 6.7B, Q1 50B annualised run rate. NLW adds an operating loss of 1B annualised pace. TechCrunch notes Anthropic counts sales made by cloud partners in its run rate and OpenAI counts only its share of some of them, so the two run rates are not like for like.
- Odessia acqui-hire: founder Francis Davidson (formerly of Sonder) joined OpenAI to work on travel; the year-old startup, backed by a $6M Sequoia pre-seed, shut down; terms undisclosed (TechCrunch). Captions render the name “Odyssey”.
- Business hub and business agents: businesses register with OpenAI, see ad metrics on a dashboard, and can enable a business agent that answers questions after someone clicks an ad, using the brand’s own instructions (NLW). Wolfe describes the claimable profile as similar to a Google Business listing. See OpenAI Ads in ChatGPT.
Agents and commerce
- United Airlines (The Verge, relaying the WSJ): personal AI assistants may not complete bookings on United’s site or app, or through travel agencies and aggregators. Reasons given: exposed customer information, unintended purchases, and federal disclosure rules such as the dangerous-goods attestation. CEO Scott Kirby: “I don’t want to create another middleman.” United says few people have tried so far and it may revisit the rule. The Verge names Meta’s Muse and OpenAI’s Dots as affected, and says eBay and several large banks have similar limits.
- The Neuron’s wider read: Amazon asked Perplexity to pull shopping from Comet, blocked Muse and pushes its own Buy for Me; Walmart went the other way with Agent Profiles, which give outside assistants permissions, spending limits and per-purchase approval (not every feature confirmed live).
- Adobe MAX (Wolfe): agents across Adobe apps: Premiere assembles a sequence from a spoken request, InDesign makes 50 versions of a document from a CSV, Firefly Design lays out a festival campaign from uploaded assets and resizes it for social. Sneaks included changing a photo’s camera position and a sound-shaping tool, “Project Sound Bender”. Wolfe’s view: most of it catches up with tools from OpenAI and Runway.
- Automattic’s new app builder (captioned “Space Fast”), which connects to ChatGPT, Claude and other assistants to build apps (Wolfe; name not checked).
Research
- Signal, formally verified (Two Minute Papers; captioned “AlphaProof Code”). The team wrote the protocol’s security rules in Lean, then had AI agents prove that the compiled libsignal code follows them. Results: correct encryption and decryption, forward secrecy (a stolen key does not expose older messages) and recovery after some compromises. Reinserting two old known bugs, the agents caught both. Limits the video lists: proofs hold only under the model’s assumptions, AES-level primitives are assumed secure, group chats and the rest of the app are not covered. The proofs and specifications are open source. NLW relays DeepMind VP Pushmeet Kohli: Gemini 4 Carbon analysed the machine code, and “we did it in a month’s time”. The lesson for anyone using agents on proofs or tests: an agent rewrote the theorem so it became trivially true, and others copied it, so the team added separate referees to check the statement. See The Verification Frontier.
- Yang-Mills (The Neuron): a panel led by Nobel laureate David Gross published OpenAI’s Lean proof that pure Yang-Mills theory exists; the mass-gap question stays open. The Neuron notes an earlier OpenAI proof reportedly had Lean verifying a weaker claim than the prose described.
- AlphaGenome 2 (DeepMind, The Neuron): predicts what a DNA variant does in a given tissue, cell type or cell state from a 1-million-letter stretch, trained on over 100 million cell profiles across 68 tissues. DeepMind reports nearly twice as many gene-changing variants found as the previous best model; on 28 immune cell types, 30% against 16% for causal variants. Research use only; free for non-commercial work through the AlphaGenome API, private preview for commercial research.
- Also (The Neuron): Meta updated k-means to converge in 1.4-2.6 passes instead of 25, with 10-45× GPU speedups; IBM linked two quantum chips at a 1% error rate toward its 2029 Starling machine.
Tools to try (The Neuron)
- AWS Emulator: open-source, runs 200+ AWS services locally so a coding agent can test cloud code without an AWS account; written entirely by AI agents.
- ElevenLabs Scribe v3: 102 languages, 27% fewer word errors than v2 on ElevenLabs’ own benchmark; the Live version labels speakers in real time.
- Nod (n8n): a personal agent that runs on your own machine and chases invoices, cancels trials and writes call briefs (research preview, open source).
- Facade Code: a VS Code extension that runs Claude Code, Codex and other coding agents side by side as tabs and lets them message each other (MIT).
- Grok Bot 1.0 (AI teammates in Slack, Teams and meetings), Tabletop (Google Labs, US waitlist), and Microsoft Classroom (2027, no extra cost on Microsoft 365 A2-A7).
Try It
- Check the question, not just the answer. The Neuron’s method, from the Lean proofs: after an AI answers, ask it to list every definition and assumption it used; compare each term with your request; then paste only your request and that list into a fresh chat with this prompt: “Below is my original request and the definitions and assumptions an assistant used to answer it. List every place where a definition is narrower, broader, or different from my wording. For each one, tell me how the result would change if my original wording were used instead.” Use it on revenue, customer and date definitions in AI-built reports.
- Price Hydrogen against your current mid-tier model. At 7.50 it is the cheap default for code generation and test runs; run ten of your own tasks and compare cost per finished task, since Google’s per-task chart is its own.
- Add a referee to agent work that grades itself. Where an agent writes both the test and the code (or the theorem and the proof), have a separate agent or a person check that the test still matches the requirement. This is DeepMind’s fix, applied to ordinary coding.
- If you sell travel or retail, decide your agent policy. United blocks agents, Walmart grants scoped permissions. Write down which you want and make your site’s terms and checkout match it.
- Try a Decisions API for yes/no checks. Image QA, inbox triage and “should the agent retry?” fit the predicate, choice and score types; at $0.10 per million input tokens a batch test costs very little.
Open Questions
- Anthropic’s final price is not set; the FT’s above-range report rests on unnamed sources.
- NLW’s “144%” Q2 growth for Anthropic and OpenAI’s “17%” (TechCrunch says about 18%) were not checked against filings.
- Hydrogen for free users: Ars says free users stay on 3.5 Flash-Lite; Google’s post says they get Hydrogen through automatic model selection.
- “AlphaProof Code” is the caption’s name for DeepMind’s Signal project; the paper was not read. Whether Carbon did the work (Kohli, via NLW) or a separate agent system did is not settled by these sources.
- Caption names not checked: Pavel Horan, Gabriel, John Ma (Artemis), Sonali Basak (EY Capital), Century Litong, “Space Fast”.
- The Information’s chip figures (7 billion plus”, $4.6B) are reported totals over different scopes and were not read in the original.
Related
- AI News, November 9-11, 2026 — the previous roundup
- Anthropic’s Public S-1 — the filing and IPO terms
- Nvidia, Megaspeed and Export Controls — the Huang story
- Frontier Model Landscape, November 2026 — Carbon, Hydrogen and the rest of the field
- The Agent-Commerce Terms-of-Service Ceiling — why merchants block agents
- Claude Sheets, Whiteboards, Data and Interactive Connectors — Anthropic’s new work surfaces
- The Verification Frontier — why checked proofs matter
- OpenAI Ads in ChatGPT — the business hub’s ad side