Podcasts and talks about AI, 15-Nov-26

• Nvidia CEO Jensen Huang has resigned, and Colette Kress, the company's CFO, is interim CEO, according to Bloomberg reporting.
• Nvidia's stock is down about 14 percent on the week, its worst since January 2025, with earnings due Tuesday.
• Anthropic's IPO is reported to be multiple times oversubscribed, with demand for at least twice the shares available.
• Anthropic seeks up to 90.4 billion dollars through the IPO and a private placement, valuing it near 1.4 trillion dollars.
• Anthropic released new versions of its Mythos and Fable models, with Mythos restricted to users the company allows.

BLOOMBERG TECH

* Nvidia CEO Jensen Huang resigns; stock down about 14% on the week
  – Jensen Huang has resigned as Nvidia CEO, and CFO Colette Kress is interim CEO, according to Bloomberg's Ian King.
  – King says the board had retained a headhunter and considered an interim CEO, and Huang resisted leaving until Wednesday night.
  – Nvidia's stock is down about 14 percent on the week, its worst since January 2025.
  – A congressional request also summons Nvidia's general counsel and Commerce Department officials.
  Bloomberg Tech

* Anthropic IPO Said to Be Multiple Times Oversubscribed
  – Bloomberg's Riley Griffin reports that sources describe Anthropic's IPO as already multiple times oversubscribed.
  – Canvas Ventures co-founder Rebecca Lynn says the offering validates AI-first venture portfolios.
  – Lynn expects an OpenAI listing next year, noting Anthropic nears profitability while OpenAI differs.
  – Bloomberg also reports CoreWeave's third-quarter sales topped analyst estimates.
  Bloomberg Tech · at 33:51

* Anthropic releases Mythos and Fable models, with Mythos restricted to certain users
  – Anthropic has released new versions of its Mythos and Fable models, Bloomberg's Rachel Metz reports.
  – Mythos is aimed at cybersecurity and limited to users the company allows.
  – Fable is generally available with more guardrails, faster at complex operations and cheaper.
  – Anthropic would not say what the U.S. government concluded after reviewing the models, saying it offered no feedback.
  Bloomberg Tech · at 4:14

* Menlo Ventures report puts enterprise AI spend at $177 billion by September
  – Didi Das, a Menlo Ventures partner, says enterprise AI deployment spend reached 177 billion dollars by September.
  – That is up from 33 billion dollars a year earlier and 1.7 billion dollars in 2023.
  – Das says 80 percent of AI usage comes from agents rather than chatbots.
  – He says Anthropic, OpenAI and Google take 69 percent of total spend.
  Bloomberg Tech · at 34:09

THE INFORMATION'S TITV

* Adobe's Deepa Subramaniam says Adobe is exposing local MCP endpoints for its creative apps
  – At Adobe MAX 2026 in Miami Beach, Adobe's Deepa Subramaniam tells The Information's TITV that Adobe is exposing local MCP endpoints.
  – The endpoints let users drive Adobe's creative applications with an agent of their choice.
  – This is alongside the Adobe Creative Agent inside Photoshop and Premiere, and earlier connectors into LLMs.
  – Subramaniam says MCP endpoints give higher-fidelity output than connectors.
  The Information's TITV · at 9:09

* Amy Dockser Marcus describes interviewing Ray Kurzweil alongside his AI double, RAI
  – The Information's Amy Dockser Marcus describes interviewing futurist Ray Kurzweil next to RAI, his AI double.
  – RAI runs on a Google large language model trained on Kurzweil's books, interviews and archive.
  – Marcus says Kurzweil sees RAI as preserving his legacy and continuing after his death.
  – Her account says his children prefer the living Ray.
  The Information's TITV · at 36:29

* Jing Yang: Jensen Huang's ties to Alice Huang helped facilitate at least $7.5 billion in Nvidia chip shipments to China
  – The Information's Jing Yang says Jensen Huang and Alice Huang, who are not related, have known each other for years.
  – Alice Huang, her sister and her husband built a network around Shanghai's Century Litong, founded in early 2023.
  – That network procured at least 7.5 billion dollars in Nvidia chips since 2023, roughly 240,000 mostly A and H generation chips.
  – Yang says it is unknown how much Jensen or Nvidia knew of the arrangement.
  The Information's TITV

MOONSHOT

* Jensen Huang resigns from NVIDIA after 33 years; Colette Kress named interim CEO
  – Peter Diamandis reports that Jensen Huang resigned from NVIDIA after 33 years, with Colette Kress as interim CEO.
  – Diamandis cites New York Times reporting on alleged affairs, including with a Singapore executive under U.S. intelligence watch.
  – NVIDIA shares fell more than 13 percent at the open, closing 4.3 percent down.
  – Trump suggested Tim Cook should take the job.
  Moonshot · at 5:21

* Anthropic releases Claude Whiteboard, Sheets and Meetings; Microsoft shares fall 3.5 percent
  – On November 10, Anthropic put Claude Whiteboard and Claude Sheets into beta and made Claude Data generally available.
  – Sheets builds a spreadsheet from a prompt, and Data connects to a warehouse or CRM to build dashboards.
  – The same day, Anthropic launched Claude Meetings, placing Claude in Google Meet as a visual participant.
  – Microsoft shares fell 3.5 percent on the news, according to the MOONSHOTS podcast.
  Moonshot · at 1:12:33

TBPN

* TBPN hosts: GPT-6 Cyber launches today as unreleased OpenAI cyber model hits 77% on Exploit Gym
  – The TBPN hosts report that GPT-6 Cyber launches today, citing an X post showing an unreleased OpenAI cyber model.
  – That model scored 77 percent on Exploit Gym, above GPT-6 Cyber at 52.6 percent and Claude Mythos 5.2 at 39.2 percent.
  – The hosts say the benchmark tests whether a model can hack while staying aligned and not cheating the test.
  – They advise companies worried about attacks to get access to such models quickly.
  TBPN · at 53:53

* Bloomberg reports Nvidia CEO Jensen Huang admitted to affair with woman identified as suspected Chinese spy
  – The TBPN hosts discuss a Bloomberg report that Jensen Huang admitted to an affair with Alice Huang.
  – Reports identify Alice Huang as a suspected Chinese spy.
  – The hosts call it the first real crisis in Nvidia's three-decade history.
  – They note they had previously flagged her state-media background and her role running Malaysian hyperscaler Mega Speed.
  TBPN

THE COGNITIVE REVOLUTION

* Cisco's Outshift open-sources Scope, Accord and Concord protocols for multi-agent AI
  – Vijoy Pandey, SVP and GM of Outshift by Cisco, tells The Cognitive Revolution that Cisco has open-sourced three protocols for multi-agent systems.
  – The protocols are Scope, Accord and Concord, covering knowledge graphs on agent traces, intent alignment and negotiated outcomes.
  – Cisco has deployed multi-agent systems with Swisscom, ServiceNow, SoftServe and Webex.
  – The AGNTCY project now belongs to the Linux Foundation.
  The Cognitive Revolution

* Goodfire's Curt Tigges says ablating up to 90% of neurons leaves performance largely intact, with ARC-AGI gains
  – Goodfire's Curt Tigges discusses neuron ablation results on The Cognitive Revolution.
  – Removing the middle 50 percent of neurons barely changes performance, while cutting 80 to 90 percent has more impact.
  – On ARC-AGI, more aggressive ablation improves performance, which Tigges attributes to models reaching answers faster.
  – Fine-tuning restores reasoning performance, he says.
  The Cognitive Revolution

ALL-IN

* Jensen Huang resigns as Nvidia CEO after spy report and affair disclosures; stock falls about 15 percent in a week
  – The All-In hosts report that Jensen Huang resigned as Nvidia CEO after 33 years.
  – The sequence began on Friday, November 6, with a New York Times report that Alice Wang, a Megaspeed executive, was a CCP spy.
  – Huang then sent a company-wide email admitting an affair, and a board investigation followed.
  – The CFO is interim CEO, and the stock fell about 15 percent in a week.
  All-In · NVIDIA's own announcement; leaked memo reported by The New York Times

AI IN MARKETING (PAUL ROETZER)

* METR's high-access evaluation of Anthropic found no proof of recursive self-improvement, Kaput reports
  – Kaput reports that AI safety nonprofit METR published a high-access evaluation of AI R&D acceleration at Anthropic.
  – The evaluation ran from August 12 to October 29, 2026, examining research data and interviewing staff.
  – METR found no proof of recursive self-improvement but said Anthropic made roughly two years of progress in one year.
  – Anthropic redacted much of the most informative evidence, saying it could accelerate other developers.
  AI in Marketing (Paul Roetzer) · at 24:54

THE MOST INTERESTING THING IN AI (NICHOLAS THOMPSON)

* Battsengel says Mongolians avoid Chinese and Russian AI models, favoring ChatGPT and Claude
  – Bolor-Erdene Battsengel, head of AI Academy Asia, tells Nicholas Thompson that Mongolians use no Chinese or Russian AI models.
  – Most use ChatGPT, with many companies moving to Anthropic Claude.
  – She cites data-privacy concerns, censorship and low awareness of Chinese models.
  – Her plans include training 1 million people and a second office abroad.
  The Most Interesting Thing in AI (Nicholas Thompson) · at 50:40

AI ENGINEER

* Artificial Analysis co-founder says model choice also means choosing an inference provider
  – George Cameron, co-founder of Artificial Analysis, says picking a model also means picking an inference provider.
  – The inference provider affects price, speed and cache discounts.
  – He describes the Artificial Analysis Intelligence Index, which combines ten evaluations, with Opus 5.5 leading while over 500 models were listed.
  – He says open-weights models trail proprietary ones by roughly three to nine months.
  AI Engineer

CHINATALK

* McGuire: Nvidia's China-threat data and CEO's spy ties damage its export-control credibility
  – On a ChinaTalk podcast, Chris McGuire of the Council on Foreign Relations says Nvidia's China-threat data conflicts with Huawei's actual chip production.
  – He says that conflict now damages Nvidia's credibility on export-control policy.
  – He says CEO Jensen Huang never internalized his product's strategic weight.
  – He expects the bipartisan Match Act and Overwatch bills to pass, likely through the NDAA.
  ChinaTalk · at 15:28 · author has a stake: policy advocate for stricter export controls

LATENT SPACE

* Handshake AI Builds Benchmark Showing Most Frontier Models Score Below 20 Percent on Life Sciences Tasks
  – In a Latent Space episode, Handshake AI's Nick Magazine and Jonas Mueller introduce Vials, a life sciences benchmark.
  – Magazine says most models score under 20 percent out of the box and all score at 50 percent or worse.
  – The benchmark tests how well agents interpret assay outputs, a prerequisite for wet-lab work.
  – Models are not yet up to the task, they say.
  Latent Space · at 30:42

---
The Observatory · https://observatory.mgorsky.net/archive/387/
