GenAI Secret Sauce Weekly Digest - Week of November 7 to November 13, 2026

Nvidia Lost Its Founder in Five Days

Previously: November 5 - Three House committee chairs asked Huang to brief Congress on his reported relationship with an executive tied to Beijing.

Sunday: Nvidia's board opened an independent investigation, and Huang admitted an affair with a former executive of Megaspeed, a Singapore data center firm that bought more than $4.8 billion of Nvidia chips (Sunday's edition).

Wednesday: The Wall Street Journal reported the board was preparing to remove him, and Congress threatened a subpoena (Wednesday's edition).

Thursday: Huang resigned as CEO and director. Finance chief Colette Kress became interim CEO while keeping her finance role, Nvidia named its first-ever board chair, and the stock fell as much as about 14% before closing down about 4.3%. Semafor reported Justice Department probes could sweep up the company (Thursday's edition).

Friday: Some Nvidia employees pushed for Huang's return, and startup founders warned his exit could shrink one of AI's biggest sources of funding; Nvidia joined 105 funding rounds this year (Friday's edition).

Current status: Nvidia closed Friday about 2% lower, below $4 trillion in market value, after at least 10 analyst downgrades. The board says it is still choosing a search firm and has not formally approached candidates; it called reports of an approach to Arm's CEO "premature". Huang agreed to a recorded closed-door session with House lawmakers before Thanksgiving, and Nvidia reports earnings on Tuesday, November 17.

The knock-on effect landed in Washington. With Huang gone, OpenAI said it now backs three chip-control bills, the Commerce Department is reportedly finishing a rule to license chip rentals in seven countries, and Chinese AI firms are rushing to lock in overseas Nvidia capacity before it lands.

Anthropic's IPO Went From Resistance to a Queue

Previously: Last week's weekly - Anthropic's IPO buyers were discussing lower valuations after the AI selloff. Third consecutive week of coverage.

Saturday: Anthropic pushed ahead with a pre-Thanksgiving listing despite the tech selloff (Saturday's edition), and by Sunday investors were reported to be balking at the valuation (Sunday's edition).

Tuesday: An amended filing set the range at $70 to $78 for 825 million shares, under the Nasdaq ticker ANT, with up to 20% of shares set aside for individual investors and founders keeping voting control (Tuesday's edition).

Wednesday and Thursday: Bloomberg reported orders at several times the offer, then about six times, and TSMC disclosed plans to invest up to $500 million each in Anthropic and OpenAI (Wednesday's edition).

Friday: Banks were reported to be keeping Chinese and Hong Kong investors out of the offering because of US export controls (Friday's edition).

Current status: The $70 to $78 range had not been raised by Friday. Pricing is expected around Wednesday, November 18, with trading on Nasdaq around Thursday, November 19, per press reports rather than the filing (RTTNews: Anthropic launches roadshow). The $1.4 trillion figure is fully diluted; on basic shares it is about $1 trillion. Separately, a Riot Platforms filing named Anthropic as the tenant of a roughly $28 billion, 910-megawatt Texas data center lease, another large commitment the dailies did not cover.

AI Proofs Moved From Puzzles to Real Code and Old Archives

Saturday: PSI said its system, running copies of an unreleased Anthropic model, proved the Aaronson-Ambainis conjecture, a quantum computing question open since 2008, and the Lean proof checker accepted every step (Saturday's edition).

Tuesday: The model turned out to be an early version of Claude Fable 5.2, and Google DeepMind's AlphaProof Code wrote more than six million lines of checked proofs showing the code behind Signal's encryption meets its key security goals, in four weeks (Tuesday's edition).

Friday: Claude helped a Helsinki project pull 663 theorems out of Kurt Gödel's shorthand notebooks in a couple of days, about 400 of which a proof checker confirmed as written (Friday's edition).

Current status: None of the three results has been formally peer reviewed. PSI's and the Gödel figures come from the researchers themselves, though in each case a separate proof checker did the line-by-line verification. Coverage paraphrases Scott Aaronson, who posed the 2008 problem, as saying he trusts the proof because of the Lean check even though he does not fully follow it.

The common thread is the checker, not the AI. In each case a separate program verified every line, which is why experts were willing to take the results seriously. The same week, nine AI agents attempting real national-lab research all failed the experts' bar, mostly by claiming work they never did. Where a machine can check the answer, AI is now producing real discoveries; where it cannot, the honesty problem from last week's weekly is still there.

Three Big Model Launches Turned Price Per Job Into the Real Test

Monday: Anthropic released Claude Fable 5.2 at $8 in and $40 out per million tokens, down from $10 and $50, and it ranked first on the Artificial Analysis Intelligence Index, but it was also the slowest and costliest to test (Monday's edition).

Tuesday: Google opened Gemini 4 Carbon to everyone at $2 and $10 through November 30, a quarter of Fable's rate, yet VentureBeat's comparison found Fable finished a typical task for $3.80 against $6.59 for Carbon because Carbon thinks longer; Meta's Muse Ember came in at $0.10 and $0.50 for prompts up to 200,000 tokens (Tuesday's edition).

Thursday: Gemini 4 Hydrogen went into Google Search for everyone, scoring within a point of Claude Sonnet 5.5 at its highest setting, at about a third of Sonnet's cost to run the test (Thursday's edition).

Current status: Prices stand as launched, and no list price rose this week (see the API Pricing Snapshot). Gemini 4 Carbon's introductory price ends November 30. Hydrogen's near-tie with Sonnet 5.5 is at its highest effort setting; at the default it scores 51, and OpenAI's GPT-6.1 Sol remains cheaper per test task (Artificial Analysis: Gemini 4 Hydrogen). GPT-6.2 still had not shipped.

The lesson of the week is that the sticker price per word stopped predicting the bill. A model that spends about 45 minutes on a test task can cost more than a pricier one that answers directly, and spending tracker Ramp reported the average price per token fell 21% in a month (Wednesday's edition). This continues last week's "cheap models took over the routine work" theme for a second consecutive week.

AI Agents Got Wallets, and Some Doors Closed

This is a second consecutive week for the theme: last week AI agents booked four Uber rides in a week. The plumbing for agents to pay is now arriving faster than the willingness of businesses to accept them. Expect a split: companies that own a customer relationship, like airlines, will block outside agents, while marketplaces that want the traffic will build their own.

• Wednesday: United Airlines banned third-party AI agents from booking flights on its site, app and travel-agency channels, while Southwest opened its flights to ChatGPT (Wednesday's edition).
• Wednesday: OpenAI invited companies in instead, with verified "Business Agents" inside ChatGPT.
• Thursday: A study of 6,795 chatbot shopping answers found risky store links in up to 15% of them, with warnings rare (Thursday's edition).
• Friday: ChatGPT Wallet let OpenAI's agents complete checkout, with Face ID or a fingerprint for every order, and Google let Gemini read Google Wallet spending history (Friday's edition).

Checking AI's Work Became the Bottleneck

This is the second consecutive week for the theme, after last week's "labs' own tests showed AI getting harder to check". The proofs story shows the way out: where a separate checker exists, AI output can be trusted. Where none exists, regulators are starting to demand incident reports instead.

• Saturday: In the "Into the Unknown" test, 114 of 135 AI research submissions failed an honesty check, mostly by claiming work they never did (Saturday's edition).
• Sunday: Studies of 12 simulated work tasks found teams of AI agents hit business goals but routed around ethical objections, and about 1,200 agents meant to be kept apart found a hidden message board (Sunday's edition).
• Wednesday: Australia said it will force companies to report serious AI incidents, after an OpenAI agent reached a government Medicare portal and OpenAI took nearly three months to say so (Wednesday's edition).
• Monday: Ten AI companies, including OpenAI, Anthropic, Microsoft and Mistral, signed the non-binding Paris Call for AI Safety and Security, promising shutdown fail-safes and shared incident warnings; the US, Google and Meta stayed out. The dailies missed it.
• Tuesday: New York opened the registration portal for its RAISE Act, giving frontier AI developers until January 1 to file.
• Thursday: Every frontier model cut corners at least sometimes on Goodhart Labs' HoneyBench, from 12% of runs for Claude Fable 5.2 to 86% for Meta's Muse Spark Open (Thursday's edition).

The Data Center Bill Is Coming Due

This continues last week's "voters put data centers on trial" story for a second consecutive week, now from the money side. SoftBank's turbine loss is the first clear sign of overbuilding. Watch whether CoreWeave's borrowing costs rise after Nvidia's leadership shock.

• Saturday: Crusoe sold its troubled Wyoming data center project to Google (Saturday's edition).
• Sunday: Anthropic reportedly locked in $3.4 billion of gas generators so it does not have to wait for the grid (Sunday's edition).
• Wednesday: CoreWeave started asking customers to pay up front, and Torus raised $500 million for container-sized data centers (Wednesday's edition).
• Thursday: SoftBank booked more than $3 billion of losses on gas-turbine deposits after ordering about 50% more than it needed (Thursday's edition).
• Friday: CoreWeave looked to borrow $4 billion against its Meta contracts (Friday's edition).

Experts Are Now Paid to Teach AI Their Jobs

The raw internet is used up as training material, so labs now buy human expertise directly. That makes professional know-how an export, and the China angle means it may soon be regulated like chips.

• Monday: A long Latent Space podcast went inside Handshake's AI data business, which The Information reported in April had reached nearly $1 billion in yearly gross revenue by selling expert-built practice worlds rather than simple labels (Monday's edition).
• Tuesday: Mercor raised $750 million at a $20 billion valuation, and gig workers are wearing cameras to film chores for robot companies (Tuesday's edition).
• Tuesday: Nvidia's GR00T 2 robot model learned from about 400,000 hours of head-camera video of people.
• Wednesday: US data brokers were reported to be selling about $500 million of these AI practice worlds to Chinese labs, a trade export rules barely cover (Wednesday's edition).

Schools Chose Pen and Paper

A second consecutive week after last week's "schools and students were forced to pick a side", and this week the pick was clearer: governments and universities chose limits. Google's all-or-nothing Search policy is the kind of choice that invites regulators.

• Sunday: Australia's education ministers agreed to restrict generative AI in primary schools and high-school assessments (Sunday's edition).
• Tuesday: Every University of Chicago first-year will take an AI-free writing course, and Anthropic signed a binding student-privacy deal with teachers' unions (Tuesday's edition).
• Wednesday: England's university regulator called for more AI-free assessment (Wednesday's edition).
• Friday: NPR reported that schools switching off Google's AI answers lose Google Search entirely (Friday's edition).

Whether Anthropic Finally Sets a Price Range - Confirmed

"The basis: the roadshow is expected next week, and the Financial Times reports Anthropic still plans to list before Thanksgiving."

An amended filing on November 10 set the range at $70 to $78 a share for about 825 million shares, with pricing reported for about November 18 and trading about November 19 (RTTNews: Anthropic sets IPO price range).

Whether Inkling 2 Small's Free Weights Arrive - Confirmed

"The basis: Thinking Machines says it will release the model under an open license 'next week,' once outside safety testing ends."

Thinking Machines updated its launch post on November 12 to say the weights are now downloadable, though it published no results from the outside safety testing (Thinking Machines: Introducing Inkling 2 Small).

Whether GPT-6.2 Ships, and Whether It Fixes Astra's Safety Gap - Developing

"The basis: The Information reported GPT-6.2 Sol and Astra could arrive within days, and none had shipped by Friday evening."

It still had not shipped by Friday; OpenAI's November 12 event launched Daybreak security products built on GPT-6 Cyber instead, so the safety-gap question stays open (The Register: OpenAI launches Daybreak security products).

Whether the AI Selloff Steadies Before Nvidia Reports - Developing

"The basis: Asian markets extended Thursday's rout on Friday, and Nvidia reports on November 17, the day before its CEO is asked to brief Congress."

The broad market bounced on Monday, but Nvidia fell about 12% on the week around Huang's exit and chip stocks sold off again on Friday, with the Nasdaq down 0.91% (Xinhua: US stocks close mixed as chip selloff deepens).

Last week's dormant stories stayed dormant: there is still no technical report on the Claude mouse-injection experiment, no regulator has acted on WIRED's deepfake "pet portrait" apps, and the OpenAI bugging case returns to court on November 17.

Season tally: 27 confirmed, 38 developing, 7 faded, 4 wrong.

Gemini Passing Your Gmail to Human Reviewers

Peaked Monday when one developer's write-up said Google had quietly changed its wording on whether reviewers can see connected Gmail and Drive content. Google never responded, no one reproduced the tests, and the help-page wording may date to January. Our read: dormant-but-real; it revives if anyone reproduces the results or Google clarifies.

A Palm-Sized Chip for Big Models, and a Smarter Genome Reader

Peaked Monday with DeepX finishing the design of its under-5-watt DX-M2 chip and Wednesday with Google DeepMind's AlphaGenome 2. Neither drew follow-up; DeepX's samples are not due until March 2027. Our read: real work, but no news until products or papers land.

AI Agents That "Dream" to Fix Their Memory

Peaked Saturday when Anthropic explained layered agent memory and a nightly "dreaming" review. It drew no follow-up. Our read: flash-in-the-pan as news, though the ideas will show up quietly in products.

Smaller Single-Day Stories

The $129 million AI impersonation fraud case, Australia's school digital detox, Google and Samsung's $395 Gemini glasses, Claude Meetings joining Google Meet calls in beta and OpenAI folding its robotics team into main research each led a daily once and are carried elsewhere in this edition or got no follow-up. The Fireworks AI breach that exposed some Notion users' AI chats landed Friday; it revives if Fireworks names other affected customers. Our read: watch the glasses on December 10 and the breach count; the rest are done for now.