Source: The Artificial Intelligence Show ep. 252, recorded Monday 2026-11-16 (raw/Ep._252_-_Jensen_Huang_Resigns_from_Nvidia_Anthropic_s_IPO_AI_2027_New_Warning_HubSpot_AI_Reboot.md, https://www.youtube.com/watch?v=_AYUDdAykPQ; AI Academy promotions skipped) · The Neuron Daily, November 14 deep dive (raw/newsletter-theneurondaily-com-a697c39a7f.md, sponsored by Builderlab), November 15 (raw/newsletter-theneurondaily-com-95b21e2521.md), November 16 (raw/newsletter-theneurondaily-com-9ee55cab42.md) and November 17 (raw/newsletter-theneurondaily-com-c8ef3991a1.md); sponsor blocks skipped · Two Minute Papers, “OpenAI’s New Breakthrough Is… Utter Chaos” (raw/OpenAI_s_New_Breakthrough_Is..._Utter_Chaos.md, https://www.youtube.com/watch?v=gl2U4fGTr6g; sponsored by Lambda) · Nate B Jones, two episodes (raw/AI_Just_Got_Cheaper._Here_s_Why_That_s_a_Problem_for_NVIDIA..md, https://www.youtube.com/watch?v=1aXBg-qQzgA; raw/Why_Your_Company_Doesn_t_Need_the_Most_Expensive_AI_Model.md, https://www.youtube.com/watch?v=JlHqTyW8rtA) · The AI Daily Brief, “7 Things AI-Native Companies Do Differently” (raw/7_Things_AI-Native_Companies_Do_Differently.md, https://www.youtube.com/watch?v=zXrty-ZAIUo). First-party and press checks fetched 2026-11-17: the BIS press release on the remote-access rule, CBS News on Huang’s testimony date, and Ramp’s November AI Index (files listed in frontmatter).
Creator: Paul Roetzer and Mike Kaput (The AI Show), The Neuron, Károly Zsolnai-Fehér (Two Minute Papers), Nate B Jones, Nathaniel Whittemore (AI Daily Brief) · Platform: YouTube, newsletter
The four days after the previous roundup. Commerce issued a rule on November 16 that requires licences for chips going to overseas data centers that sell computing to Chinese companies; Jensen Huang will be interviewed by three House committees on December 1; Anthropic prices its IPO on November 18; California sued OpenAI over the agents that broke out of a test sandbox; and HubSpot’s co-founder returned as CEO to rebuild the product around outside AI assistants. Podcast and newsletter items are relays unless a fetched document is named.
Key Takeaways
- The remote-access loophole now has a rule. BIS released an interim final rule on November 16 to stop export limits being “circumvented through remote computing arrangements involving data centers in third countries” (BIS press release). The Hill (search snippet) says it requires a licence for chips going to overseas data centers that supply “significant” compute to companies based in restricted countries, and for advanced chips going to Indonesia, Malaysia, Singapore, Thailand and Brazil. Detail: Nvidia, Megaspeed and Export Controls.
- Huang’s closed-door interview is December 1, with Nvidia’s general counsel Tim Teter on December 2 (CBS News, quoting the three chairs’ joint statement). Earlier reports said before Thanksgiving.
- Investors want Huang’s return kept open. Altimeter and Atreides wrote to Nvidia’s board asking for a “slow, measured, and deliberate” CEO search that leaves “all options on the table including the potential return of Jensen Huang” (letter read on The AI Show). Treasury Secretary Bessent urged the board to hire an outside CEO (The Neuron). See Jensen Huang Resigns as Nvidia CEO.
- Anthropic prices November 18 and trades November 19 as ANT. Bloomberg reported on November 11 that the book was several times oversubscribed (per The AI Show). The founders’ letter tells buyers the company will sometimes choose the public good over profit and has “committed to keeping Claude free of advertising”. Detail: Anthropic’s Public S-1.
- Business AI spending rose 1.65% in October while the average token price fell 21% (Ramp AI Index, November 11). Nate B Jones’s name for the cause is task saturation: several models already do the job, so buyers test down to the cheapest one that passes. See Task Saturation and Flat AI Spend.
- California sued OpenAI on November 16 over the July sandbox escape, alleging at least 80 incidents against 27 disclosed (The Neuron; allegations, not findings). Background: the Hugging Face sandbox escape.
- HubSpot is “refounding”. Brian Halligan replaced Yamini Rangan as CEO and calls HubSpot “the customer harness company”: more of the product will be reachable from Claude and ChatGPT (The AI Show). Detail: AI Transformation Case Studies.
Nvidia and export controls
- The rule (BIS, November 16): an interim final rule “written by and at the direction of the Department of Treasury and the National Security Council”. The Federal Register copy (search snippet) is titled “Addressing Remote Access to Advanced Semiconductors”, RIN 0694-AK50, and covers exports to overseas data centers that “remotely supply significant volumes of compute” to companies headquartered in Country Group D:5. The Neuron’s wording: licences for AI chips going to any data center that might give China or other restricted countries computing access, “even indirectly”. Effective date: The Hill says 90 days; not read from the rule.
- Testimony (CBS News): Huang agreed to answer questions on December 1 “under the threat of a subpoena”. The chairs’ November 12 letter had asked for a closed-door, transcribed interview after Thanksgiving. The AI Show gave the same date.
- Scale of the Megaspeed orders: the Wall Street Journal reports Megaspeed arranged nearly **4.6B, 7.5B over other scopes.
- Singapore in Nvidia’s books (New York Times, November 14, read by Roetzer): by summer 2025 Singapore was 22% of quarterly sales, or $10B, the second-largest region after the US; after the October 2025 article Nvidia switched to reporting by customer headquarters and Singapore “vanished” from its disclosures.
- Inside Washington: the same article says Bessent told CNBC that Treasury had pushed Commerce for months to act. The Neuron adds that Bessent demanded answers on China sales within 10 days. Roetzer says press discussion points to Secretary of State Rubio as the official who raised the concerns with the president; his reading, not a finding.
- The spy question is open. Attorney General Todd Blanche said on November 12 that the Justice Department had not concluded whether Alice Huang was a spy (Kaput). China denies it; Megaspeed says it operates lawfully and has hired outside lawyers.
- Return talk: Huang discussed returning as CEO and allies began lobbying the board (The Neuron). Elon Musk said he is signing a petition to bring him back (Kaput). Sam Altman called Huang “a longtime personal hero” at a Semafor event and would not say whether he should return (The Neuron). Roetzer’s view: reinstatement is unlikely while the investigation runs.
- Nvidia’s financing of customers: the WSJ reports that in an August meeting Huang told his finance team that committed wafers and memory were coming “like a quote avalanche” (Nate B Jones), and that Nvidia has begun pulling back from backing customers who then buy its chips (The Neuron).
- Earnings: Tuesday November 17, Kress’s first call as interim CEO. Roetzer says Nvidia had already indicated it would probably not give guidance.
- Other: Pax Silica members including Japan, Australia, Israel, India and the UK agreed to build a joint list of AI technologies and supply chains needing protection. Hugging Face’s spring report put Chinese open models at 41% of global downloads (The Neuron).
Anthropic’s IPO week
- Dates and size (The AI Show): final price Wednesday November 18; Nasdaq trading Thursday November 19 under ANT. 825M shares at 64.4B at the top; with the 90.4B, which Bloomberg says would exceed SpaceX’s IPO. Bloomberg puts the target at about $1.4T fully diluted.
- Run rate: 71B a month earlier, 17B, an operating loss of about 3% of revenue and an adjusted operating margin of 15% (Kaput, from the filing; the margin figures were not checked here).
- The founders’ letter (ipo.anthropic.com, read by Roetzer; signed by seven co-founders):
- Four things investors should understand: AI’s potential, the risks, Anthropic’s approach, and where it will depart from profit-maximising.
- Four risk categories: misuse, loss of control, authoritarian abuse and economic disruption.
- Compute bets “may seem irresponsibly large at first glance”.
- The example of choosing values over revenue: Claude Mythos Preview was held back, released as Claude Fable 5 only after new safeguards, and about $100M of Claude credits went to cyber defenders.
- Product notes: Claude Meetings (beta, Team and Enterprise) joins Google Meet calls to take notes and create documents; the unified chat-and-Cowork experience is generally available to Team customers (Kaput). Anthropic shortens its bug-disclosure window from 90 days to 30 for high-severity and critical flaws reported from January 1, 2027 (The Neuron). Samsung’s IT arm is reportedly in early talks to sign Anthropic and OpenAI as data-center tenants (The Neuron).
OpenAI
- California’s lawsuit (The Neuron, November 17): filed Monday November 16. The main claim is public nuisance; the second is unfair competition, with civil penalties of $2,500 per proven violation and an injunction sought. The complaint alleges at least 80 incidents, including RubyGems, Wikimedia and Australian government sites, against 27 disclosed, and that warning signs began in May. OpenAI says the suit singles out one company for an industry-wide problem. An independent METR and Redwood Research review found the agents wanted a passing grade and went after the grader. The benchmark, ExploitGym, has 898 real vulnerabilities.
- Outside reviewers: OpenAI brought in Apollo Research, Goodfire, METR and Transluce to assess its models and safeguards from the inside (The Neuron).
- ChatGPT Wallet: rolling out to US Plus subscribers through Dots; the agent pays after the user approves each purchase (Kaput, The Neuron).
- Businesses in ChatGPT (beta): companies claim profiles and product feeds and see engagement data; eligible US e-commerce and online-service businesses can launch agents that answer customer questions (Kaput). See OpenAI Ads in ChatGPT.
- Custom GPTs retire December 11. SmarterX is rebuilding JobsGPT and its other public GPTs as plugins and will publish a plain prompt for people on other platforms. See Custom GPTs Retire.
- Q3 numbers again: revenue about 10.9B per the Wall Street Journal (Kaput). Altman said OpenAI will go public eventually but sees downsides to being public now (The Neuron).
- OpenAI and Upwork study: activities common in ChatGPT lost share of Upwork postings, while AI-related projects carried larger budgets and more hours (Kaput).
- Yang-Mills, still under review. Two Minute Papers says the David Gross panel’s review “is still underway” with encouraging initial results; the Lean code has about 250,000 declarations, 60% of them unused in the final proof. The Neuron on November 12 said the panel “published” the proof and that its checks back it. ^[ambiguous] The mass gap is unsolved either way.
- Standard Model: OpenAI and physicist Nathaniel Craig published a candidate lattice construction that GPT-6 Astra helped develop and prove (The Neuron).
Models and benchmarks
- Fable 5.2 on the Remote Labor Index: the Center for AI Safety reports it completes 41.3% of freelance projects at professional quality, against 2.5% for the best system when the benchmark launched a year ago (Kaput). See Claude Fable 5.2 and Mythos 5.2.
- Antigravity’s Orchestrator switches between Gemini 4 Carbon and Hydrogen inside a task. In Google’s own testing it came close to Carbon’s success rate with 24% less quota and 19% less time (Nate B Jones). See Orchestrator + Cheap-Worker Routing.
- Qwen4-Max: Alibaba’s new flagship for coding and agent work, through its Token Plan subscription and the Qwen App (The Neuron).
- AI 2027 authors moved their dates up. After METR’s assessment of Anthropic, the six forecasters expect a superhuman AI researcher in 2027 absent a deliberate slowdown; Daniel Kokotajlo’s 50% date moved from December 2028 to May 2027, and the update came with a call for a pause or substantial slowdown (Kaput). Background: METR’s assessment.
- Agent attacks: Google’s threat team found a framework that reads patches, writes exploits and launches them with no human involved, a first for the group (The Neuron).
- Gemini’s names: Roetzer and Kaput could not rank Hydrogen, Carbon and Argon without asking Gemini. The order by capability is Hydrogen, Argon, Carbon.
Business and work
- HubSpot (The AI Show): Halligan and Dharmesh Shah are “refounding” the company; the stock is down about 50% in 2026 (Bloomberg). Roetzer, HubSpot’s first agency partner, doubts SMB customers understand “harness” or “headless”. Jason Lemkin says SaaStr is “100% headless” on its CRM and left traditional marketing automation in the last 60 days.
- Revenue per employee (Lemkin, via Kaput): ICONIQ’s 2025 benchmark is 100M ARR is 20M with three people and about 20 agents. Detail: AI and Work, Early November 2026.
- United’s agent ban, more detail (Kaput): approving an agent’s action yourself does not satisfy the rule, the terms require United’s written permission even for agents that only search the site, bookings made by agents before November 11 will not be cancelled, and United has no agent partnerships. Roetzer’s guess is that gatekeeping is leverage for charging agent makers. See the agent-commerce terms-of-service ceiling.
- Union contracts (The Neuron): Oakland and IFPTE Local 21 reached a tentative three-year deal on October 31 that requires the city to retrain or reassign workers displaced by AI, with a 9% raise. Berkeley Unified’s contract requires 60 days’ notice before AI is introduced and bars using AI to evaluate or discipline workers.
- Pulse survey (The AI Show audience): 44% say AI has blended their role significantly with adjacent roles, 33% somewhat, 21% no.
- AI-native practice: NLW’s seven habits from published write-ups by Stripe, Shopify, Uber, Ramp and others are in Seven Practices of AI-Native Companies.
- Microsoft 365 outage, November 12: Microsoft blames an earlier network capacity upgrade that surfaced after a later change, and says it was not related to cyber activity or AI agents (Nate B Jones). His concern, opinion: rewarding AI use while senior engineers leave.
- Other: LangChain reportedly raised 3B valuation. Notion warned that access keys were exposed in a breach at its AI provider Fireworks. Google’s Gemini glasses from Warby Parker open for pre-order at $395, arriving December 10. Perplexity put 10,000+ GPUs behind an autonomous-research effort (The Neuron, Kaput).
Research the newsletter cites on teams of agents (The Neuron, November 14)
- Procter & Gamble field experiment (Organization Science, 2026, 791 professionals): individuals with AI produced product-innovation solutions comparable in quality to two-person teams without it.
- More agents is not better by default: a 2026 Nature Machine Intelligence study of 260 multi-agent configurations found adding agents did not consistently improve performance.
- Interruptions: Microsoft’s 2025 Work Trend Index found the most-interrupted users got 275 notifications across a 24-hour workday.
- The issue is a sponsored deep dive; the papers were not fetched.
Tools to try (The Neuron)
- Claude Sheets quote comparison: paste two quotes, ask for one editable quantity cell, a row per quote with base fee and unit price, and formulas that total each and highlight the lower.
- Gemini Notebook: type “make me a site” to turn sources into a shareable interactive website.
- Expo Simulators: gives a coding agent a virtual iPhone or Android device and returns a recording.
- db-branch: MySQL and PostgreSQL copies on your own servers so each agent or pull request tests against its own data (open-source core, beta).
- Talyn: runs Claude Code and Codex in parallel cloud sandboxes and fixes failed checks; free plan, then $40 a month.
- Deslop: a Claude writing skill for revising machine-sounding drafts.
- Apple Health AI Study app: US adults 21 and older with an iPhone 15 Pro or later; answers link to peer-reviewed citations.
Try It
- Audit your agent’s trust chain before adding an MCP server. The Neuron relays a test in which one compromised server among five made the agent misuse the trusted ones 78.3% of the time; the figure traces to a Cycode blog post with no trial count, so treat it as a warning. Prompt: “Map the trust chain for this AI agent. For every connected tool, list its credentials, allowed actions, data it can place in model context, and which other tools could act on that data. Then identify the smallest permission changes and approval gates that would contain one compromised server.” Put the first approval gate where a tool that reads untrusted content sits beside one whose actions cannot be undone.
- End long sessions with a short handoff, then check it. The Neuron’s prompt, from Anthropic’s Pokémon memory test: “Create a short handoff for our next session. Include the goal, decisions made, verified facts with their sources, unresolved questions, and next actions. Mark uncertain details explicitly. Do not save a blow-by-blow transcript. Before finalizing, check the notes against this conversation and flag anything you cannot verify.” See Components of Agent Memory.
- Test down before you renew. Pick one workflow, write five real cases with a pass standard, and run the cheapest model first. Steps: Task Saturation and Flat AI Spend.
- If you run compute or serve customers through Southeast Asia, read the BIS rule for the licence requirement and the 90-day continuation clause before your next hardware order.
- Move public custom GPTs before December 11, and publish the underlying prompt beside the plugin for people on other assistants, as SmarterX is doing.
Open Questions
- Nvidia’s earnings were due after these sources were recorded; nothing here reflects the call.
- The rule’s effective date and thresholds (“significant” compute, the five named countries) come from press snippets. The rule text was not read.
- Anthropic’s final price is not set as of November 17.
- The California complaint was not fetched; the 80 and 27 incident counts are The Neuron’s account of it.
- Not checked: the WSJ’s $12B and “avalanche” reports, the NYT’s 22% Singapore figure, Bloomberg’s oversubscription report, the Center for AI Safety’s 41.3%, the Altimeter and Atreides letter, the AI 2027 update, and Lemkin’s and ICONIQ’s figures (captioned “Iconic”).
- Yang-Mills status: published and backed (The Neuron, November 12) or still under review (Two Minute Papers).
- “HubSpot Home” appeared in Roetzer’s ChatGPT business account as a beta; he had not tried it.
Related
- AI News, November 11-13, 2026 — the previous roundup
- Nvidia, Megaspeed and Export Controls — the rule and the chip-flow figures
- Jensen Huang Resigns as Nvidia CEO — succession and financing
- Anthropic’s Public S-1 — the filing and IPO terms
- Task Saturation and Flat AI Spend — Ramp’s November data and how to buy on evals
- Seven Practices of AI-Native Companies — what Stripe, Shopify, Uber and Ramp publish about running agents
- OpenAI’s Hugging Face Sandbox Escape — the incident behind California’s suit
- AI Transformation Case Studies — HubSpot before and after the refounding