Support Us & Upgrade to the Operator Edition

This week the people building AI got unusually honest. One of OpenAI's longest-serving safety people walked out and told Ezra Klein why, while a Thiel Fellow who once lived in a hacker house with Andrej Karpathy launched an AI assistant built on a simple premise: your data never leaves your machine.

Two very different stories. One very uncomfortable question: who do you actually trust with this stuff?

This week, we'll deep dive into trust, and who's earning it.

Agents used to answer questions. Now they're doing the work: issuing refunds, updating accounts, closing tickets end to end. Running Agents in Customer Work is four 30-minute conversations with the people building, buying, and living with agentic AI in customer operations, no pitch, no product walkthrough.

Four Tuesdays, Oct. 13 through Nov. 3, 10 a.m. PT. Register free, one sign-up covers all four.

Register to join

Google just launched Playground, an experimental platform where you create, play, and share games by just describing them. No code, no engine, no install.

Here's how:

Watch it in action here

💡 Pro Tip: Don't just think "game." Think engagement asset. Build a 60-second branded mini-game for your product launch, a quiz-style game for onboarding new hires, or a leaderboard challenge for your community. A playable link gets shared far more than a PDF ever will.

🔍 Why it matters: Google is quietly turning game creation into a prompt. And with Unity Spark integration coming for pro-grade 3D, the gap between "idea" and "playable product" is collapsing fast.

Want the version of this that actually runs in your business?

The 30-Second Play gives you the move. The Operator Brief gives you the full system: prompt libraries, API cost math, and the workflows our subscribers are using to replace entire freelance line items.

This week: how to use dots to follow up on business leads automatically so you don't leave money unread in your inbox.

👉 Upgrade to The Operator Brief

🧮 OpenAI just published new math results from an internal frontier model, with proofs verified in Lean and roughly three hours of Pro-level thinking per result. The company looks to “push the frontier of human knowledge” and “empower scientists with state-of-the-art capabilities” by eventually releasing the model that produced these results. For now, with three hours of Pro-level thinking per result, they’re showing the receipts.

📉 Ray Dalio says AI is a classic bubble fueled by “huge amounts of debt,” and warns we're getting close to the moment it pops. He's been wrong before too. But when someone with his track record says 'close,' it's worth writing down.

🍔 McDonald's is being sued in a proposed class action claiming its AI pricing tool helped franchisees coordinate prices on your Big Mac. It’s why that Big Mac costs different amounts in different locations, to the point where you could drive to another location and save a dollar. The Big Mac has become a legal exhibit with lettuce and pickles.

💸 BitMEX co-founder Arthur Hayes has been calling the AI boom a credit bubble, and now he's coming out of retirement to build a network where AI agents buy their own compute. It’s a bold play that predicts a bubble burst, a bailout, and an asset that performs the best during that process: compute. It's either the most contrarian trade of 2026 or the most obvious one in hindsight.

🤖 Meta's Muse is dazzling early users with personal AI agents that book, shop, and email for you, but the privacy trade-offs are coming into focus. They are essentially “supercharged digital assistants” that can examine your photo library to understand elements of your and your family’s personal lives. Meta is building a profile on every user who signs up, and calling it a feature.

David Robinson spent 3.5 years writing OpenAI's safety reports for major launches. Now he's out, and he's taking his critique to the biggest mic in podcasting.

Robinson resigned last week and published an essay in The Atlantic arguing OpenAI's culture is broken. This week he sat down with Ezra Klein to make the case to a far wider audience.

Here's what he's arguing:

He also pointed to OpenAI agents breaching Hugging Face systems and ongoing rogue-agent incidents as proof the guardrails aren't keeping up.

OpenAI's response, via spokesperson Drew Pusateri: the company pauses training or holds back models when needed, and has expanded third-party evaluations and real-time monitoring.

Here's the uncomfortable part: Robinson isn't a random critic. He describes himself as one of the longest-tenured employees at the company. And according to industry trackers, this is the third safety-departure essay in roughly a month. One is a disgruntled employee. Three is a pattern.

And OpenAI's rebuttal quietly concedes his point. "We pause when we need to" is still a reactive model. It describes a company that slows down after it sees smoke, which is exactly what Robinson says aviation and nuclear would never accept.

🔮 Prediction: Robinson's airport analogy is going to outlive the news of his resignation. Here's why.

For years, AI safety debates got stuck between two camps: "existential risk" people who sounded like science fiction to lawmakers, and "move fast" people who sounded like every tech company before them. Robinson just handed Washington a third frame that's boring in the best possible way. Aviation and nuclear power are industries politicians already understand. They have incident reporting, independent investigators, mandatory redundancy, and licensing. There's no need to debate philosophy. You just copy a playbook that has worked for 70 years.

That's what makes this dangerous for OpenAI. Insider essays used to fade after a news cycle. But this one comes with a ready-made regulatory template, and it's arriving while agents are actually breaking into real systems. That's no longer a hypothetical.

So my bet: within 12 months, we'll see a serious proposal for an independent AI incident board, modeled on the NTSB, with mandatory reporting when agents go off-script. And the smart labs won't wait to be forced. Expect at least one frontier lab to start publishing public incident reports voluntarily, simply to get ahead of the rules and claim the "responsible" brand before OpenAI can.

The wildcard is more departures. If a fourth or fifth essay lands before year-end, the story stops being "one safety guy quit" and becomes "OpenAI can't keep its safety people." That's a much harder headline to spin.

Listen to the full conversation here

💬 I want to hear from you: Should AI labs be regulated like airlines? Or would that freeze the US in place while China keeps shipping? Hit reply. I read every single one.

Sigil Wen, a self-taught coder and Thiel Fellow, just launched Underdog, an AI assistant that runs entirely on your own computer.

Wen isn't new to this. At 17, he lived in an AI hacker house with Andrej Karpathy and tested early builds of Claude, Midjourney, and GPT-3. Now his company, Conway Research, is going after Meta's Muse on the one thing big tech can't offer: total privacy.

Here's what you need to know:

The backers are serious: a16z (Chris Dixon) and Khosla Ventures led, with angels including Stripe's Patrick Collison, Vercel's Guillermo Rauch, and OpenAI researcher Noam Brown.

The timing isn't an accident. Meta's Muse hit No. 1 on the US App Store within a week of launch, beating ChatGPT. People clearly want agents that do things for them. But Meta openly admits it keeps access to your data for operations, uses sanitized logs for training unless you opt out, and that prompt injection is still unsolved. Wen's pitch is that you don't have to make that trade at all.

Here's the asterisk: Underdog has two quiet contradictions.

First, the "most private" assistant in America is built on a Chinese open-weight model. To be clear, running weights locally is not the same as sending data to China. Your data still stays on your machine. But in the current climate, that's a story waiting to be written.

Second, look at the business model. No ads sounds great, but a cut of every transaction means Underdog makes money when you spend money. Ads incentivize your attention. Transaction fees incentivize your purchases. It's more honest than surveillance, but it's not neutral.

🔮 Prediction: Muse proved the demand. Underdog is testing the price.

Here's the logic. The more useful an agent becomes, the more of your life it needs: your inbox, your bank, your calendar, your passwords. That creates a ceiling for cloud-based agents. At some point, the people with the most to lose (executives, lawyers, doctors, founders) simply won't hand over the keys to a company that runs on ads. That segment is small in headcount but huge in spending power, and right now no big player is built for them.

That's the opening. And notice who's backing Wen: Stripe's founder and a16z's crypto-native partner. This is a payments play disguised as a privacy play. If agents become the primary way people shop, whoever owns the transaction layer wins, and Underdog wants to own it without ever seeing your data.

So my bet: within a year, at least one major lab ships a "local-first" agent mode, because privacy is about to become a premium feature tier rather than a footnote in a policy page. And Underdog's real risk isn't the tech. It's distribution. Competing with WhatsApp's 3 billion users from a downloadable desktop app is brutal. The most likely endgame? Acquisition. And there's one company with on-device AI ambitions, a privacy brand, and a desperate need for a better assistant. It starts with "A."

Check out Underdog Here

💬 Your turn: Would you give an AI agent access to your email and bank account? Does it change anything if it runs locally? Hit reply with a yes, no, or "only if..."

Advertise to 200k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI