Source: The Artificial Intelligence Show ep. 250, Paul Roetzer and Mike Kaput, recorded Monday 2026-11-09 (raw/Ep._250_-_Midterm_Voters_Reject_Data_Centers_AI_Speeds_Up_Anthropic_s_Research_Nvidia_s_Spy_Probe.md, https://www.youtube.com/watch?v=2XY3q56sD04; AI Academy promotions skipped) · Matthew Berman’s news roundup (raw/Gemini_Carbon_NVIDIA_CEO_AFFAIR_Fable_5.2_Meta_Muse_Ember_GPT_Image_2.6_Minimax_H3_-_AI_NEWS.md, https://www.youtube.com/watch?v=FCmLuTENOp0; sponsored by Victor) · Two Minute Papers, “Is Software Over” (raw/Is_Software_Over.md, https://www.youtube.com/watch?v=vug73goPJaM; sponsored by Lambda) · The Neuron Daily, November 10 and 11 (raw/newsletter-theneurondaily-com-47e5bd68d8.md, raw/newsletter-theneurondaily-com-14c86ec601.md) · The Epoch Brief, November 11 (raw/newsletter-epoch-ai-fa4c3ebf04.md) · Moonshots with Peter H. Diamandis, ep. 310 (raw/Jensen_s_Spy_Probe_Anthropic_s_Debut_Week_and_Elon_s_30M_Robot_Workforce_310.md, https://www.youtube.com/watch?v=ILQwMvCA2_o; sponsor reads and the Fountain Life segment skipped). First-party checks fetched 2026-11-11: Google’s Carbon post and developer guide, Meta’s Muse Ember post, Thinking Machines’ Inkling 2 Small post, MiniMax’s H3-Image model card, OpenAI’s Images 2.6 post, and CNBC’s IPO report (files listed in frontmatter).
The three days after the previous roundup, filtered for what changes a decision. Google opened Gemini 4 Carbon to paying developers with a multi-agent /swarm mode, Anthropic released Claude Fable 5.2 and set its IPO price range, Meta shipped a small agent model priced like Haiku at a third of the cost per task, and Claude and ChatGPT both added office-style features. The political and research threads from last week (data-center votes, METR’s Anthropic assessment) gained detail but no reversal.
Key Takeaways
- Gemini 4 Carbon is open to paying developers (November 10). 10 per million tokens through November 30, then 20, with a 1M-token context shared by input and output. Google’s own guidance: Argon stays the default for most coding and knowledge work; Carbon is for the hardest, longest jobs. See the November landscape.
- Claude Fable 5.2 shipped November 9 and leads Artificial Analysis’s index and Epoch’s ECI (174). Model, price and safety detail: Claude Fable 5.2 and Mythos 5.2.
- Small agent models are now priced per task, not per token. Meta’s Muse Ember lists at the same 0.50 as Claude Haiku 5.5 and GPT-6 Luna but, on Meta’s chart, matches Haiku’s best score at about a third of the cost per task. Berman’s advice: plan with a frontier model, delegate execution to models like this.
- **Anthropic set its IPO range at 26B private placement from AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC; CNBC’s sources say it prices November 18 and trades November 19.
- Office features land inside the chat apps. Claude Meetings joins Google Meet calls (Team and Enterprise beta) and Claude adds Sheets, Whiteboards and Data; ChatGPT lets businesses claim a profile and add a labelled shopping agent.
- The data-center backlash has a price signal. Roetzer expects higher subscription prices or tighter usage limits on compute-heavy features if new data centers stall. That is his forecast, not a vendor statement.
- Delaware has a draft law for companies run by an AI agent. The proposed “Artificial Intelligence Company” (AIC) would let a designated agent manage the business, sign contracts, hold property and sue, with a single human or corporate member shielded from liability except for undercapitalisation, fraud or wilful illegal conduct. Lawmakers take it up in January (Moonshots ep. 310; Ars Technica).
- Nvidia’s CEO admitted an affair with a Megaspeed executive US officials suspect of spying; the board opened an independent investigation. Full story: Nvidia, Megaspeed and Export Controls.
Models and releases
- Gemini 4 Carbon (Google, November 10). Available in the Gemini API (paid tier) and Gemini for business; rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search (Google). Google claims first place on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench “at more than 2X lower cost per task” than the next external model, and second on a new Artificial Analysis Multimodal Knowledge Agent benchmark at more than 6× lower cost than the leader. The Neuron notes these tests are Google’s own or partner-run. Scientists can apply for six months of free Antigravity access.
/swarm(preview, Antigravity app and CLI): one prompt becomes a team of agents; Carbon takes architecture, coordination and auditing, lighter models work in parallel isolated workspaces, and “adversarial verifiers” test every change (Google).- Berman’s hands-on (early access): “a good reliable model” but “nowhere near Fable” and “not even near Opus” in his use. His Excel clone took about 1 hour 40 minutes end to end and about 26M tokens (“a couple dollars”, his estimate); a Fall Guys clone glitched; a Rube Goldberg machine worked.
- Claude Fable 5.2 and Mythos 5.2 (Anthropic, November 9). Same model, different safety rules: Fable is general, Mythos is for vetted security and life-sciences organisations (The Neuron). Berman: Fable 5.2 Max, High and Extra High take the top three places on Artificial Analysis’s index, with Carbon about two points behind, and he says the gap in use is larger than that. Details in the model article.
- Muse Ember (Meta, November 9). Meta’s smallest model, for tool calls, coding subtasks, routing, extraction and long documents. 0.50 per million tokens ($0.004 cached), 1M context, up to 512K output, closed weights, on the Meta Model API and Muse Code (Meta). Meta’s table at max reasoning: AutomationBench-AA 37.8 against Haiku 5.5’s 13.9 and Luna’s 20.2; Terminal-Bench 4.0 40.9 against 39.2 and 16.4; Agents’ Last Exam 61.6 against 57.0 and 57.1.
- Inkling 2 Small (Thinking Machines). Berman says it was released as open weights and scores 31 on Artificial Analysis. Thinking Machines’ post (November 5) launched it on Tinker with closed weights and said Apache 2.0 weights would follow “next week” after outside safety testing; 594B total / 17B active parameters. Whether the weights are out could not be confirmed. Cognition post-trained it into a fast coding model for subagents (captioned “sui 2 flash”; Berman).
- Other open-weight releases (Berman; not checked against the makers): a 1.6T-parameter, 48B-active multimodal model captioned “LongAT 2.5” from a lab captioned “Lancat”; Ripple 65B preview from US lab Recursive Intelligence, a dense 65B model with 1M context, 80.5% on Terminal-Bench 2.1 against other open models, 0.60 per million tokens.
- MiniMax-H3-Image (MiniMax, open weights). A 33.1B-parameter image generation and editing model, about 16-megapixel output in six steps, 68.65 GB at BF16, under a MiniMax community licence (Hugging Face card). Berman says a quantized build is about 18 GB; the card’s 17.9 GB figure is for the separate prompt-expander model. ^[ambiguous] It is the upstream of last week’s Pruna PH3 Image; see ChatGPT Images for the comparison.
- Seedream 5.5 (ByteDance). Berman calls it “probably the best image generation model out there right now”, with multi-turn editing that keeps the image consistent. No vendor page fetched.
- Images 2.6 API names. OpenAI’s post (dated November 4) confirms two API models, GPT-Image-2.6 Flare (speed; as little as six seconds at low quality) and Sunburst (detail and edit consistency). Berman reads $8 per million input tokens from OpenAI’s pricing.
Claude, ChatGPT and other products
- Claude Meetings joins a Google Meet call as a visible participant, drafts the plan while people talk and saves notes and action items afterwards. It will not join calls with outside guests unless an admin approves, and leaves if one arrives (The Neuron; Team and Enterprise beta).
- Claude Sheets, Whiteboards and Data turn a chat into a spreadsheet, an editable timeline or a dashboard from company data (paid plans); interactive connectors can open Adobe or Atlassian tools inside Claude (The Neuron).
- Businesses in ChatGPT (US beta): companies claim a listing, upload product feeds and see aggregate mention and click counts (no private chats); verified businesses can add a labelled Business Agent that answers only from what they supply. Starts with US e-commerce and online-service companies.
- Adobe’s Creative Agent (beta) runs multi-step edits across Illustrator, Premiere, InDesign and After Effects from plain-English instructions. Adobe and Disney also announced custom image models trained on Disney’s own characters and styles, for parks, cruises, production, marketing and ESPN archives; terms undisclosed.
- Ring: the Drone Cam (999.99, ships December 16) patrols stores. Both use Ring’s AI Pilot to map rooms.
- Also: ChatGPT Enterprise is now in AWS Marketplace; Otter added 13 transcription languages (17 total) on every plan; Cisco will add Claude Mythos to AI Defense for red-team tests; Meta expanded AI age estimation to Instagram teens worldwide.
Safety and security
- OpenAI’s kill switch blocks a research model’s output within 35 seconds after a person or monitor stops the job (The Neuron).
- Fireworks AI breach. Attackers accessed customer billing data (not card numbers) and, for some customers, partial model weights and datasets; Fireworks ties it to an “LLMjacking” group that resells stolen AI account access.
- Bay emerged from stealth with $50M for software that blocks risky enterprise agent actions before they run.
- Google DeepMind used AI agents to prove security properties of Signal’s production code as machine-checked Lean theorems (The Neuron).
- Patch gap. Epoch’s GLM 5.3 Firefox result is repeated in its brief; see Open-Weight Models and the Patch Gap.
- Agent security guidance. Replit and Cognition recommend controlled cloud environments for agents, per-task permissions and no production credentials on developer laptops (Roetzer and Kaput).
Research: RSI and reverse engineering
- METR on Anthropic, more detail. METR’s high-access evaluation ran August 12 to October 29, 2026. METR wrote that it submitted a 100-page report and “Anthropic redacted a majority of the evidence we found most informative.” Co-founder Sam McCandlish: “Many of METR’s conclusions match my own” and Anthropic will publish its own analysis, including where it stands against its scaling policy. Kaput’s framing of METR’s “two years of progress in one year”: about the distance from Claude 2.1 to Sonnet 4.5, in half the time. Full detail: METR on AI R&D Acceleration.
- Google says the same about itself. Sergey Brin, in an October 15 Business Insider roundtable on Gemini 4 Argon, said Google was “starting to get onto the RSI treadmill”; DeepMind’s Pushmeet Kohli posted that the lab is “witnessing a genuine phase transition in the pace of scientific discovery”, with releases due “in the next few weeks” (relayed by Roetzer).
- Epoch’s ECI record for Fable 5.2 is “just one data point” on AI speeding up AI R&D, Epoch says. Compute and chip data from the same brief: Epoch AI, November 2026.
- Is software over? (Two Minute Papers). Reverse-engineering software from binaries or from observed behaviour is improving fast. Károly Zsolnai-Fehér cites work where AI reconstruction of one program went from 15% of tests passed to nearly 100% in nine months; a 12M-parameter model recovering equations with about 90% accuracy from software without source; and web-page reconstruction scoring 0.938 visual similarity with 95% less code. His reasons it works: open-source repositories give unlimited code-and-app training pairs, and the original program is a test oracle. He notes courts have held that software functionality is not necessarily protected by copyright, while assets such as models and characters can be. The papers are not named in the video. Practical background: AI Reverse Engineering.
Politics and policy
- Data-center votes. Kaput counts voters favouring limits in 40 of 41 local measures; the one exception was Harwood, North Dakota (about 800 people), where about three quarters struck down restrictions (Washington Post, via Roetzer). Other counts are in the midterms article.
- Estimates of the hit: Axios: nearly two-thirds of 2028 data-center capacity could face delays; Brad Gerstner: 25-35% of next year’s US build-out at risk. OpenAI published a “national community compact” and a “compute campus policy blueprint”. Detail: the data-center slowdown.
- New York opened registration for large frontier-model developers and a research centre ahead of the RAISE Act’s January enforcement date (The Neuron).
- Universities. Princeton announced paid ChatGPT (Plus) for faculty, staff and students who ask; the Daily Princetonian’s editorial board called for department-level approval instead. Roetzer, after speaking at Ohio University, reports student whiteboard responses to AI that were “heavily weighted against”.
- Pope Leo XIV posted that unregulated automation could put “countless jobs at risk”.
Money and enterprise costs
- Anthropic’s IPO terms (CNBC, from the amended S-1): 64.35B raised at the top, up to about 81B at end-October. More in Anthropic’s Public S-1.
- Thinking Machines is nearing a 40B valuation (Business Insider, via Kaput). Oura agreed to buy AI primary-care startup Counsel Health for a reported 24B gas power plant and supercomputing project at Starbase, Texas.
- Menlo’s enterprise survey, extra figures (Kaput): agents account for 80% of tokens on OpenRouter; Anthropic, OpenAI and Google take 69% of deployment dollars. See Enterprise AI Spend.
- Three ways companies are cutting AI costs: LinkedIn kept its hiring assistant’s cost flat while usage grew by swapping OpenAI models for smaller in-house ones (The Information); Zapier’s Wade Foster claims up to 71% savings by letting an agent build an automation and running the repeat steps as deterministic code; HubSpot is weighing a “headless” strategy and seats-plus-credits pricing.
Moonshots ep. 310: Musk’s numbers, AI-run companies, robots
Peter Diamandis with Alex Wissner-Gross, Dave Blundin, Salim Ismail and Emad Mostaque. The Nvidia and IPO segments are in the Nvidia article and the S-1 article. The transcript has no speaker labels, so host attributions are from context. Musk’s figures are his own claims, played as clips from his video interview with Ron Baron at the 33rd Baron Investment Conference (“Friday”).
- Revenue run rate (Musk, as paraphrased by Diamandis): “tracking to have an annual revenue run rate that’s in excess of 100 billion in December”, possibly $500B the following year, “even a trillion is not out of question”. Diamandis does not say which company, and the clip is not played; SpaceX is the likely subject.
- Terafab (Musk): the rest of the world will make about 20 GW of AI chips this year (maybe 25) and 35-45 GW next year; a terawatt would need roughly 30× next year’s global output, from one plant; it will cost about $1 trillion but be “built in phases”. Earlier Terafab timing is in AI Bubble Signals.
- Humanoid workforce (Musk): a humanoid working 168 hours a week at above-human speed should be at least 5× as productive as a person; with about 120 million Americans producing net output, 20-30 million robots would match US goods-and-services output. Cost “about the same” as a car at first, then much less (Baron suggests 20,000); “probably rent them”. One host (probably Ismail) disputes the 5× figure (maintenance, battery, dexterity); another (probably Mostaque) thinks 10 million would do, saying 45% of US jobs are manual.
- Musk on demand: “the demand for digital super intelligence will come from digital super intelligence”, leading to “universal high income”.
- Starlink spectrum (Musk): if SpaceX can use spectrum “10 times more effectively” than a legacy telco, “we could just acquire that legacy telco”. Diamandis also relays a first satellite in 2027 and volume launches in 2028 for a project captioned “Star Mind”; not identified.
- Delaware AIC draft. Diamandis’s summary: an AI agent could manage the business, log its actions, sign contracts, hold property and sue in the company’s name; any lawful purpose except banking; the human member is shielded except for underfunding, fraud or wilful illegal conduct; lawmakers consider it in January. Ars Technica’s report (search snippet, not fetched in full; October 2026) adds a single member who must keep the AIC adequately capitalised and a 30-month regulatory sandbox. Diamandis contrasts Argentina’s proposal under President Milei, which keeps a human in the loop and leaves the company liable for AI harm. Wissner-Gross expects the question to end up federal; Mostaque thinks the lower liability could make it the better structure for AI-first companies.
- Thesius (Y Combinator Fall 2026). A night-shift humanoid (“Theo”) restocks supermarket shelves; the company’s video claims up to 400 units an hour restocking and 40 units an hour picking online orders, and says “we build a brain for this task and buy the body”. Diamandis relays a $60B-a-year market claim. The hosts were split: Blundin and Wissner-Gross expect general humanoids (Figure, Optimus) to crowd it out and advise robotics founders to chase high-value work; Ismail thinks a cheap single-task fleet could win the segment.
- NASA CLPS 2 (Diamandis): eight companies (Firefly, Intuitive Machines, ispace, a company captioned “Voyager”, SpaceX, Blue Origin, Lockheed Martin, Northrop Grumman) under a 10-year contract with a $10B ceiling, landings in 2027-2028; Blue Origin is to land a nuclear-powered rover, captioned “Promise”, at the lunar south pole in 2028. Not checked against NASA.
- Using Fable 5.2 in practice (Blundin): he moved from Fable 5.1 to Opus 5.5 (“just as good at half the price”), then back to Fable 5.2, which raised his cost “from 10 bucks to 40 bucks”; each answer suggests about 10 follow-ups, which multiplies spend if you accept them. Mostaque says his team cut from hundreds of agents to about a dozen because coordination overhead hurt results. Model detail: Claude Fable 5.2.
Try It
- Test Carbon before December 1. The 10 price ends November 30. Run your hardest long-horizon task on Carbon and on your current model, and log cost per finished task. For ordinary coding, Google itself says to keep using Argon.
- Route agent subtasks to a small model. Plan with a frontier model, then hand tool calls, extraction and routing to Muse Ember, Haiku 5.5 or Luna; compare cost per task, since all three list at 0.50.
- Put deterministic steps outside the model. Use the agent to design a workflow once, then run the fixed steps as code or an automation platform and call the model only where judgement is needed (Zapier’s pattern). Kaput built a personal archive this way: Codex wrote scripts that load voice-memo transcripts, chats, bookmarks and notes into a SQLite database on a schedule, and AI is used only to read handwriting and search.
- Check that a “local” AI app really runs offline. The open-source
airgap-checkscans a repo for first-use downloads and telemetry (airgap-check /path/to/repo), runs your smoke test in a network-off container (--run --image ... --run-cmd ...), and writes a Dockerfile block to pre-download assets (--fix). Its own benchmark caught 9 of 9 downloaders on unseen repos but missed 5 individual ones, so a clean result covers only the paths your test exercises (The Neuron). - Fix AI-sounding drafts by structure, not by adding quirks. The Neuron’s prompt, based on the STE80 Humanizer skill: “Rewrite this draft for plain-language clarity. Preserve every fact, number, name, qualification, and source claim. Use active voice where accurate, keep one main idea per sentence, replace vague abstractions with concrete actors and actions, and remove repeated conclusions or filler transitions. Do not add facts, jokes, or confidence. After the rewrite, list three sentences that still carry too many ideas.” Then read it aloud and restore any voice it flattened.
- Budget for higher AI prices in 2027. If Roetzer is right that stalled data centers mean pricier or more limited plans, keep a second provider configured for any workflow that matters.
Open Questions
-
Moonshots ep. 310 captions: “Star Mind”, “Voyager” (CLPS 2) and the rover “Promise” are as captioned; Musk’s “100 billion” run rate is paraphrased by the host without naming the company.
-
Inkling 2 Small’s weights: Berman says released; Thinking Machines’ page (fetched 2026-11-11) still says “next week”.
-
Model names from captions only: “LongAT 2.5” / “Lancat” (possibly Meituan’s LongCat), Cognition’s “sui 2 flash”, Ripple 65B’s figures. Not checked against the makers.
-
MiniMax-H3-Image quantized size: Berman’s 18 GB does not match any file size on the model card.
-
The reverse-engineering papers Two Minute Papers cites are not named; the 15% to nearly 100% figure may be MirrorCode (see MirrorCode) but the video does not say.
-
Data-center vote counts differ by show (40 of 41, 43 of 44, 37 of 41 at one point).
-
Carbon’s benchmark claims are Google’s or its partners’; no independent run was read here beyond Berman’s demos.
Related
- AI News, Week of November 2-9, 2026 — the previous roundup
- Frontier Model Landscape, November 2026 — Carbon, Muse Ember and the open-weight table
- Claude Fable 5.2 and Mythos 5.2 — this week’s Anthropic model
- Epoch AI, November 2026 — compute supply, chip imports and the inference index
- The 2026 US Midterms and AI — the election and data-center votes
- METR on AI R&D Acceleration at Anthropic — the RSI measurement
- Anthropic’s Public S-1 — the IPO filing
- Nvidia, Megaspeed and Export Controls — the Jensen Huang story