Welcome, humans.

So apparently, three of OpenAI's most senior national security people just spent 17 minutes on a video call with who they thought was Ukraine's spy chief. It was two Russian pranksters.

The duo, Vovan and Lexus, have been prank-calling famous people since 2011 (past victims run from Elton John to John McCain). Here's how The Intercept says they pulled this one off. They called a senior Ukrainian official in San Francisco while posing as Rustem Umerov, who runs Ukraine's Foreign Intelligence Service. That official was already in touch with OpenAI's head of national security policy, Sasha Baker, so he helped set up a meeting. Baker, Caroline Zier (OpenAI's strategic initiatives lead), and Ben Crampton (who handles European national security work) all hopped on.

The fake spy chief then asked what OpenAI's military help for Ukraine could look like, including offensive AI, and steered the talk to Elon Musk. One OpenAI exec reportedly allowed that Musk isn't a fan of OpenAI's CEO. The pranksters posted the recording online. Seventeen minutes is a long time to not notice your spy chief is a bit.

OpenAI says it opened an investigation weeks ago, concluded the call was a prank, and took steps to prevent a repeat. It didn't say what those steps are, which is the part worth watching.

Here’s what happened in AI today:

• 😸 Nvidia CEO Jensen Huang resigned after 33 years.

• 📰 Nonprofits sued five AI labs to freeze new models.

• 📰 Anthropic's IPO was reportedly six times oversubscribed.

• 🍪 Google launched Gemini 4 Hydrogen, a cheaper agent model.

• 🎓 Let a big model plan and Hydrogen execute.

…and a whole lot more that you can read about here.

😸 Nvidia CEO Jensen Huang resigns after 33 years, and his CFO steps in

Jensen Huang resigned as Nvidia's CEO on Thursday, effective immediately, and left the board too. Nvidia's SEC filing says CFO Colette Kress is now interim CEO while the board searches for a permanent one. Director Harvey Jones becomes the board's first-ever chair.

Huang co-founded Nvidia in 1993, so this is the first CEO change in its history. Thirty-three years, one CEO, and four days from first admission to exit. Nvidia hasn't given a reason.

Here's what happened:

• Nov 8: Huang told employees he'd been unfaithful to his wife, and the board opened an independent investigation into reports he was targeted by a foreign intelligence effort.

• Nov 11: The Wall Street Journal reported the board was preparing to remove him, and the filing says it accepted his resignation that day.

• Nov 12: Nvidia announced it, shares fell more than 13% in early trading, and Huang agreed to testify before Congress, per TechCrunch's timeline.

Why a national security story? New York Times reporting says U.S. officials believe the woman works for Chinese intelligence. She's also reportedly an executive at Megaspeed, a cloud company investigated over whether it helped Chinese firms get around chip export controls.

Two caveats. The board says it has no reason to believe national security or company secrets were compromised, and her alleged spy status isn't officially confirmed. Still, governance professors told TechCrunch they know of no precedent for a CEO caught up in a suspected honeypot, which is spy-speak for a relationship used to get close to someone with influence. Not a sentence we planned to write today.

U.S. officials say separate Justice Department investigations involving Nvidia, Megaspeed, and Supermicro cover potential antitrust and chip-smuggling issues. The Atlantic argues the damage could spread because Nvidia sells chips to AI labs and also helps financially back some of them, so one shaky link could wobble the whole loop. That's an argument, not a forecast.

Why this matters: Our own Corey Noles thinks Nvidia will probably be fine long term, but the transition could be rough. His case: a CEO can leave in a day, while the ecosystem built around his company can't.

The big piece is CUDA, Nvidia's software layer that lets developers write programs for its chips. Nvidia's 2026 annual report says more than 7.5M developers use it. Picture a finance team living in one spreadsheet program. Switching means rewriting every formula and retraining everyone mid-close. Translation: the boss left, the plumbing didn't.

The slower risk is customers who keep buying today while planning to depend on Nvidia less tomorrow. Corey keeps the business question separate from the conduct one, and calls what Huang admitted unacceptable. His watch list: product delivery, leadership retention, licensing terms, and the pace of Nvidia's investments.

The first real test is Tuesday, November 17, when Nvidia reports quarterly earnings with Kress as interim CEO. Listen for whether orders hold, what she says about China, and any hint of a permanent pick.

FROM OUR PARTNERS

Fingerprint live build: spot Meta's Muse without blocking real customers

AI agents now sign in and buy for real customers, and some, like Muse, don't identify themselves. Fingerprint’s live build will show you how to let them in without letting abuse through.

On Nov 17 at 9 AM PT, Keshia Rose codes a server-side policy layer, then tests it on a real Muse session. You'll learn:

• Who's who: Tell humans, AI agents, and automation tools apart. 

• How much to trust: Signed, verified, unknown, and spoofed identities. 

• What to do: Allow, limit, verify, or block each visitor. 

🎓 AI Skill of the Day: Let a big model plan and Gemini 4 Hydrogen do the work

Google's Gemini 4 Hydrogen launch post describes it as the fast, lower-cost model for everyday coding and agent work, built to execute quickly rather than reason deeply. The heavier Gemini 4 models, Carbon and Argon, take the long, hard thinking. Google's own coding tool, Antigravity, already splits the job that way. Its Orchestrator switches between Carbon and Hydrogen in one session, and Google reports near-Carbon results with about 24% less quota and 19% faster completion on its own coding tests. Vendor numbers, so season to taste.

The post gives no prompting advice, so here's our version of the pattern for any chat. Think architect and contractor. The architect decides what to build and in what order. The contractor follows the blueprint quickly. Paying the architect to hammer every nail is slow and pricey, and asking the contractor to design the building is how you get a door that opens into a wall.

Run it in four moves:

1. Plan with your strongest model. Ask it to turn the job into numbered steps, each written to stand alone, with a "done looks like" check you can verify in 30 seconds.

2. Run one step at a time on Hydrogen. Pick it in the Gemini app's model dropdown (free users may get it through auto model selection instead). Paste the full plan for context, then tell it to do only step 3. Google says it follows tone, format, and specific rules closely, which is what you want here.

3. Match the thinking depth. Hydrogen has four thinking levels: low, medium, high (the default), and xhigh (new). Google says you pay and wait only for the depth a task needs, so if your setup lets you choose, save xhigh for the one tricky step.

4. Send only the failures back upstairs. If a step misses its check, hand that step and its output back to the big model.

Here's the planning prompt:

I'm going to hand this job to a fast model one step at a time, so write the plan for it.

Job: [describe the job and what finished looks like]

Break it into numbered steps. For each step, give me:
- an instruction that makes sense on its own
- the input it needs from earlier steps
- a one-line "done looks like" check

Mark any step that needs real judgment, not just execution.

The thing to test is Google's speed claim. Google measured Hydrogen's 2.5x speedup at its high default setting against an older model, Gemini 3.8 Flash. So time one of your own tasks at a lower thinking level before deciding where Hydrogen earns its keep.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Analytics on Live Data Without Leaving Postgres

When analytics on Postgres slows down, most teams add a second database. Then they manage pipelines, sync lag, and drift forever. TimescaleDB extends Postgres instead. Analytics run on live data, in the database you already have. No pipeline. No migration. No new query language.

Get $1000 Credit

📰 Around the Horn

• Four nonprofits asked a San Francisco judge to freeze new frontier models at Anthropic, Google, Meta, OpenAI, and SpaceXSI.

• OpenAI reportedly told lawmakers it will back bipartisan bills restricting China's access to AI chips and chipmaking equipment.

• Anthropic's IPO was reportedly about six times oversubscribed, and its roadshow claimed Claude ported a model to AMD GPUs in three days, down from months of engineer work.

• Cerebras reported core Q3 revenue of $230.9M, up 70%, and raised its full-year outlook.

• The FBI disrupted North Korea's FaceStealer campaign, which used fake Zoom-style meetings with recordings of people victims knew to steal more than $400M in crypto.

• SpaceX announced $74B in AI computing sites, including a new $50B Arizona site and a $24B gas-powered complex near Starbase.

Want absolutely EVERYTHING that happened in AI this week? Click here!

🎓 The Neuron Academy New Course Drops!

This week, we added two new lessons and two new courses to The Neuron Academy!

Together, these releases help answer two practical questions: Which AI tool fits the work, and how can you use it without losing control of the process?

• AI Tool Strategy for Business Leaders

• Intro to Mistral/Vibe: What It Is and When to Use It

• AI Coding for Non-Coders

• Codex for Practical Workflows

🍪 Treats to Try

*Asterisk = in collaboration with (only the first one!). Advertise to 700K+ readers here!

1. *Assess your organization's AI governance readiness and identify critical gaps in a live 90-minute workshop with MIT faculty. Register today.

2. Gemini 4 Hydrogen is Google's faster, cheaper model for coding and agent work, available in the Gemini app dropdown, the Gemini API, and AI Studio.

3. Preorder Google's Gemini glasses from Warby Parker and Gentle Monster for hands-free directions and 70+ language translation, from $395 (US and South Korea only, in stores December 10).

4. Dots, ChatGPT's personal agents, now follow custom instructions, so you can tell yours to write in lowercase or show you a draft before it sends any email.

5. Click the Gemini icon in Chrome's toolbar to open a side panel that summarizes, compares, and answers questions about the page you're already reading.

6. Yogurt records your calls without a bot joining the meeting, then writes up a summary with next steps by person—free/open-source.

7. Pomerado turns a website with no API, like a supplier invoice portal, into a tool your coding agent can call, built from a single prompt—free plan, then $50/mo.

8. Kiln runs chat, transcription, and image models right in your browser straight from their original Hugging Face files, with no server or conversion step—free/open-source.

UPCOMING EVENT: The Neuron IRL in San Francisco

The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.

It’s free, but space is limited. Save your spot!

💡 Intelligent Insights

• Nick Clegg argued in The Atlantic that Silicon Valley's political alignment with Trump failed on its own terms, citing China's narrowing AI gap, U.S. public resistance, and hostile foreign markets.

• The Atlantic looked at people trying to preserve their AI companions across model changes, where a product update can feel like a forced personality change.

• Sahan Paliskara and colleagues at Anthropic found per-user AI agents sharing one resource did far worse than one coordinator, with Claude Opus 5 scoring 30% versus 64% of the best outcome in one test.

• Susanna Loeb and colleagues at Stanford's SCALE Initiative reviewed 52 causal studies of AI in education and found that using AI to finish work can hurt later unaided performance, while learning-support results were mixed.

• Simon Willison argued that price per token no longer tells you much because reasoning models use different token counts, so cost per task is the better yardstick (Meta's Muse Ember scored well on it).

ICYMI: New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.