"Gambling With Our Lives"
~17 min · September 2026
▶ Now Playing
This Episode
A former OpenAI and Anthropic researcher resigned this month and said, in public, that the people building AI privately believe it could kill everyone by the end of the decade. Anthropic's own CEO has said almost the same thing, publicly, for over a year. This is a shorter, denser episode than usual — one topic, no padding — going through the actual claims word by word, applying the same scrutiny to the CEO defending AI as to the employee who quit over it, steelmanning both the sincere-fear read and the self-interested read, and asking honestly whether containment is even possible, whether superintelligence is already here, and what an ordinary person is actually supposed to do with any of this.
Cold Open
"Ninety Million Views"
The hook, then the disclosure: this show runs on Claude, made by Anthropic — the company at the center of tonight's story. Whatever standard gets applied to Anthropic gets applied to everyone else in this episode too.
The Claims
Jacob Coxon, In His Own Words
Not a summary — every claim from his resignation post and Wall Street Journal interview, read directly, sorted by what's fact, what's testimony, what's forecast, and what's simply unverifiable.
Same Standard, Every Voice
Hubinger, Amodei, and the Counter-Argument
A current Anthropic safety lead puts his name on the same number Coxon used. Amodei's said 25% for over a year. Then David Sacks, on All-In, calls the whole thing a "doomer psyop" — and gets the exact same scrutiny turned back on him.
The Steelman
Why It Might Kill Us, Why It Might Not
Sincere fear and self-interest, both taken seriously, neither one resolved — plus the uncomfortable case that US caution just hands the lead to a government with a completely different incentive.
History Drop
The Risk That Already Happened Once
A known, disclosed risk, funded anyway. A moratorium that didn't stop the work — it just moved somewhere the rule couldn't reach. The precedent nobody wants to say out loud.
The Close
What You Actually Do With This
Not fatalism, not false comfort. What isn't a real lever, what is, and why "just let China build it" doesn't actually make anyone safer.
Sources
- Jacob Coxon's resignation post and Wall Street Journal interview (TechCrunch; Deadline; Time, September 2026)
- Evan Hubinger's public reply (CNBC; CBC, September 2026)
- Dario Amodei's 25% p(doom) remarks (Axios, September 2025 and 2026)
- David Sacks on All-In, "doomer psyop" (OfficeChai; Zvi Mowshowitz's episode coverage)
- Zvi Mowshowitz, "Jacob Coxon Warns of Human Extinction and Triggers a Preference Cascade" (Don't Worry About the Vase, September 2026)
- EU AI Act enforcement status; China's "AI Plus" framework (informedclearly.com; Transcend.io)
- CIA COVID-origin assessment, January 2025 (CBS News); Fauci pardon and 2026 Senate testimony (NPR, CBC, CNN, PBS, The Hill, Fox News, CBS)
- 2014 gain-of-function funding moratorium and EcoHealth/WIV funding continuity (Congress.gov CRS; The Intercept)
- Apollo Research on model self-exfiltration and scheming under evaluation conditions (apolloresearch.ai)
- 2026 International AI Safety Report capability findings (via Axios)
- Anthropic's ~80% Claude-authored code stat (Time, August 2026)
- APA/Morning Consult AI-anxiety data, April 2026; KevinMD on "anticipatory anxiety"
The Cast
Kyle
Host. Opinionated. Expect a history drop.
Kate
The correspondent. Tight, sourced, no spin.
Morgan
The heartbeat. Asks the question nobody else will.
Full Transcript
Loads on open
▶
Full Transcript
Loads on openLoading transcript...