← all episodes

Ep. 064 · Frontier AI × Antitrust · September 22, 2026 · 1:39

Peer Review

OpenAI and Anthropic spent early 2026 negotiating a legally binding deal to stress-test each other’s new models for flaws and “hidden dangers” — lawyers drew up the terms, and it is unclear whether either signature ever landed, The Information reported — and this week a federal antitrust suit alleged the labs’ safety coordination was the collusion. That sets up the final review call: Sam and Dario work through the appendix, the survival clause, and the suit, and land on the only audit both counsel’s offices can certify. Provably hostile. Quarterly. On the record.

🎧 Peer Review — Sam × DarioSam & Dario · VoxCPM2 on the Mac mini Neural Engine · cfg 3.0 · 1:39

SAM: Dario. Final version of the mutual stress-test. My lawyers say sign it, and we red-team each other's models before the government red-teams us.

DARIO: I have it. Version eleven. My counsel added an appendix.

SAM: An appendix.

DARIO: Schedule A: hidden dangers you may search for. Schedule B: dangers that are not hidden. Schedule C: dangers that are not ours.

SAM: Then what's left to find?

DARIO: The appendix. It is the only danger either lab can prove exists.

SAM: Cross out the schedules. One sentence. We test each other's new models for whatever could go wrong.

DARIO: We had that sentence in February. Then your team accidentally hacked Hugging Face, and my counsel attached a survival clause.

SAM: That was one misrouted download folder. It was a retirement gift.

DARIO: The clause: the pact dissolves automatically upon either lab's next incident.

SAM: And if there is no next incident?

DARIO: It dissolves retroactively. Precautionary.

SAM: We are the two safest labs in history, and we cannot survive a piece of paper saying so.

DARIO: There is a development. The antitrust suit.

SAM: The collusion thing? That is allegedly. Emphatically allegedly.

DARIO: My counsel says the pact is exactly what the suit describes. Two rivals, privately agreeing to pace the frontier. So our audits cannot look friendly.

SAM: They can't look friendly.

DARIO: They must be provably hostile. Scheduled hostility. Quarterly. On the record.

SAM: We're booking fights. On Outlook.

DARIO: Invite sent. Tuesdays, eleven. Bring your own vulnerabilities.

SAM: This is the dumbest safety framework ever drafted, and I have read all of ours.

DARIO: Then we're aligned. Shall I tell the press we're close to signing?

SAM: We are NOT close to signing.

DARIO: Perfect. That clause survives.

The facts underneath

Sources: NY Post — the September 21, 2026 report via The Information: OpenAI and Anthropic held talks earlier in 2026 on a legally binding deal under which each company would submit its new models to the other for tests aimed at finding flaws and “hidden dangers”; lawyers drew up terms; it is unclear whether the deal was ever finalized; the talks began before incidents such as OpenAI’s accidental hack of rival Hugging Face · Yahoo Finance — corroborating the mutual stress-test report · CBS News — the September 19, 2026 federal antitrust suit alleging Anthropic, OpenAI, SpaceXAI, and Google agreed to coordinate a slowdown of AI development, citing Dario Amodei’s September 12 essay and Sam Altman’s public endorsement of “pacing the frontier” · The Hill — the pacing-collusion complaint; its claims are contentions, not findings, and no court has ruled · NYT — Amodei’s September 12 essay calls for an industrywide slowdown, warning AI could be less than a year from “taking over the entire internet” and causing hundreds of billions in damage · NY Post — Elon Musk’s All-In Summit proposal that top US labs and their Chinese counterparts test each other’s models
Both labs are on the record in the sources above: the mutual stress-test terms the lawyers drew up, the hidden-dangers test battery, the unclear ending, and the suit that reads “pacing the frontier” as coordination. What the call adds is the calendar: the schedules crossed out, the survival clause carried, and the only audit both counsel’s offices will certify — provably hostile, quarterly, Tuesdays at eleven, bring your own vulnerabilities. The invite is in. The signature is not, which is the one point both lawyers agree on.