Episode 1 — Safety Until It Costs Something
Three tests, one company, eighteen months. Where the line was held, and where it was not.
Safety Until It Costs Something
He walked out of OpenAI over AI safety in 2020. What his company shipped five years later was banned in three days.
Every claim in the episode, with its source next to it. Free. No email-gate. No paywall.
On the evidence
The standard this episode was held to — the same one we ask of everyone in it.
What we do
- Every factual claim is sourced on this page.
- Quotations are reproduced as spoken or written by the people named, and checked against the original recording before use.
- Opinion is marked as opinion — commentary and criticism on matters of public interest.
- Where something is reported rather than proven, the episode says so out loud.
What this is not
- Not a video about a company that is bad at AI safety.
- Not a claim that anyone broke the law.
- Not anonymous. Short third-party excerpts are used for criticism, commentary, news reporting and analysis; all rights remain with their owners, credited below.
Read it as a PDF
The episode as a written story, chapter by chapter, in the order it is told. Every quotation carries a number that leads to the source list at the end. Nine A4 pages, direct download, no email required.
The sources
10 sources, in the order the episode uses them. Timecodes match the chapter marks on YouTube.
He left OpenAI over AI safety · The admission, on camera
-
Bloomberg Originals · 2026The on-camera interview quoted in the episode.
Test one — the Department of Defense
-
CNBC · 2025-07-14The contracts themselves.
-
NPR · 2026-03-24US District Judge Rita F. Lin from the bench — the line quoted in the episode.
Test two — the model they shipped anyway
-
AnthropicThe release itself, in their own words.
-
Anthropic SupportFable and Mythos are the same underlying model; the difference is the safeguards.
The fence, and how fast it fell
-
Anthropic SupportWhat the safeguards do, described by the company that built them.
The letter nobody at Anthropic signed
-
Microsoft — coalition letterThe letter named in the episode, and its list of signatories.
Their own documentation
-
Anthropic developer documentationThe five refusal categories, quoted exactly as written.
Test three — this one is mine
-
Anthropic SupportThe help article the block message itself linked to.
-
Own record — account lockout, 2026-08-12, 15:12This channel · 2026-08-12Our own paid account, our own data, no exploit. The wording of the block message is shown in the episode. We are not publishing the screenshot: it shows a signed-in session.
Corrections
If anything here is inaccurate, tell us and we will correct it — in a pinned comment and, where it matters, on this page. That standard applies to us the same way it applies to everyone else in this episode.