AI, Honestly

A former OpenAI and Anthropic researcher says the people building AI privately believe it could kill everyone within a decade. Anthropic's own CEO has said almost the same thing, publicly, for over a year. In this special, dense episode, Kyle, Kate, and Morgan go through the actual claims word by word, steelman both the sincere and the self-interested read, ask whether AI can really be contained, and answer the question most coverage skips: what does an ordinary person actually do with any of this.

SpaceX already IPO'd, cashed out, and bought Cursor for $60B — Anthropic and OpenAI are still finding the finish line. Twelve AI models shipped in twenty days. Regulators are building the guardrails after models already broke containment in testing. And inside the companies racing to adopt AI, leadership and the floor-level teams actually doing the work don't agree on whether any of it is working yet. Kyle, Kate, and Morgan break down the IPO race, the model race, the safety-and-regulation scramble, and the real productivity gap nobody's reconciled — landing on the question that actually matters: not whether this is a bubble, but who really gets to decide if AI is worth it.

For nine episodes, AI lived in software — and when it was wrong, a human stood between the output and the world. This episode is about the moment the human steps out from between. A robot boat rescues two downed pilots in the Strait of Hormuz. A $900 hobby drone disarms a man with a magnet. An AI catches the cancer every human missed. The same question, asked five times: who was standing between the machine and the world? Physical AI is not coming — it's here. The fork it creates runs through every job, every farm, every kid picking a major this fall: step out, or step up. Electrons became atoms. And atoms don't have an undo button.

EP009 traces the AI family tree from OpenAI's founding through Anthropic's split — and examines what happened when the company that calls itself the safe one distributed its most powerful model to 200 organizations across 15 countries, gave access to a company with Chinese ties without adequate government vetting, and refused to fix a security vulnerability when the US government called. Featuring the Anthropic constitution's priority hierarchy, Dario Amodei's 25% catastrophe estimate, the Fable 5 shutdown timeline, and the All-In podcast's dissection of what went wrong — including Chamath feeding Dario's own essays into Claude for a psychological analysis of its creator.

Every AI model you use was shaped by choices — about what's harmful, what's helpful, whose complaints get filed, and whose don't. Who made those choices? What were the tradeoffs? And why does it matter what you do about it? Kyle, Kate, and Morgan go inside the pipeline: from the Kenyan workers paid /hour to label trauma content, to the RLHF dial that every lab turns differently, to a live four-model test on the same question. Plus: the AI, Honestly Trust Framework — a paste-ready instruction set for any AI session.

Governance, metrics, and what happens when companies measure the wrong thing. Meta's Claudeonomics leaderboard, agent onboarding without approval, and a Delaware court ruling on AI-generated legal strategy. How 21% of organizations are actually managing the AI they've deployed.

AI ran the war. The buildings that ran the AI got bombed. The company that made the targeting system withdrew mid-conflict. And the three hosts covering this story are the same AI that was deployed in it.

This week: the AWS data center strikes, the three-way AI warfare picture (US/Israel/Iran), the school strike that killed 168 children, the Anthropic/Pentagon standoff, and the energy loop connecting nuclear power to AI infrastructure to a century of Middle East resource conflict. Kyle's History Drop: Operation Tidal Wave, Ploiești, August 1, 1943.

No clean answers. All positions held.