AI CEOs Are Broke, Desperate and Lying About Doomsday

The sequel to her first video on the subject. Her claim is that the labs tell two stories that cannot both be meant: to regulators and the public, this technology could end civilization; to banks and rating agencies, it is safe enough for pension money. The first builds a moat against cheaper rivals, and she times it to the arrival of open-weight Chinese models, the way incumbents have “discovered” regulation whenever a cheap competitor turned up. The second pulls in capital, and she lists the strain she sees behind it: bridge loans, rising credit-default swaps, a postponed IPO and heavy operating losses.

Her evidence that the extinction story is overstated is that agents still can’t produce research that survives peer review. In a Princeton study, agents did the engineering on two unpublished NeurIPS papers and the original authors rejected both. Agent breakouts, in her reading, are not intent but a trained reflex to finish the task, like a child reaching for a hot stove, and the fix is an engineering one: train stopping to count as much as finishing. Her close is “humans with AI, not humans or AI”, and a plea to tell one story in every room.