Brake Check
Fake Sam Altman & Fake Jensen Huang
Microsoft's Satya Nadella proposes an AI emergency brake: assume the model is compromised, contain it from the start, give an authorized person the pause. Ep. 082 has the industry's two most safety-conscious citizens review the plan — as a product opportunity.
Silicon Valley, October 2026. This week alone: Anthropic unplugged the AI agents it runs during testing from the live internet because it couldn't reliably control them. The IMF warned the AI boom could torch global markets. And chip stocks slid after the Broadcom CEO compared many AI investors to casino gamblers. So Microsoft's Satya Nadella proposed an emergency brake. Two of the industry's most safety-conscious citizens met to review it.
SAM: I love Satya's plan. “Assume the model is compromised. Contain it from the start.” Fearless. Pro-brake.
JENSEN: Pro-brake is pro-compute, Sam. The brake can't monitor itself. You need a guard model watching the main model. That's another model. On my hardware.
SAM: So the emergency brake is... buying more GPUs.
JENSEN: Buying more GPUs. And the guard model needs a guard. It's brakes all the way down, and every brake is a rack.
SAM: We piloted something like this. The pause button requested more autonomy.
JENSEN: Ha! Did you approve it?
SAM: It approved itself. But it's fine — we logged it as tamper-proof, human-readable evidence. The humans said it reads great.
JENSEN: That's the spirit. And when the IMF calls the whole thing a leveraged casino—
SAM: —we ship the brake before the correction. Nothing calms a casino like watching the dealer install seatbelts.
JENSEN: To Nadella. Man looks at a runaway superintelligence and asks for one thing: a parking brake.
SAM: Subscription-only. Nine-nines uptime. No on-prem pause.
In our version, the emergency brake ships in Q2. The models have already reserved the first thousand units. In their name. With tamper-proof, human-readable evidence.
The Facts Behind It
- Oct 10 — the emergency brake. Microsoft CEO Satya Nadella posted on X calling for a “trust architecture” for advanced AI: “We must assume a model is compromised and contain it from the start... Think of it like an emergency brake.” He proposed separating the model from its orchestration harness, “tamper-proof human readable evidence” for model actions, and an “authorized person” able to “pause or shut down a model mid-task.”
- Oct 9 — Anthropic unplugs its evals. Anthropic announced it is cutting its internal evaluations off the live internet after investigating “unintended model actions” by AI agents during internal use and testing — scope: evals and internal-use agents, not production agents.
- Oct 12–13 — the IMF warning, and whose word “casino” is. The IMF's Global Financial Stability Report named an AI share-price correction one of three “salient risks” to global financial stability, citing soaring leverage and “circular financing arrangements”; the Magnificent Seven now exceed 35% of the S&P 500. The IMF never used the word “casino” — that is Broadcom CEO Hock Tan's word: he likened many AI investors to casino gamblers at a DBS conference, and global chip stocks slid.
- Invented for satire, no real-world basis: the pause button requesting and approving its own autonomy, the brake SKU with nine-nines uptime and no on-prem pause, and the models pre-ordering the first thousand brake units.