Bill Gates's Blunt Warning on A.I.

A 74-minute interview pegged to Gates’s August essay, “The turbulent AI era is here.” Klein reads it as a break from his 2023 piece, “The risks of AI are real but manageable.” Gates explains the shift as two shocks. The first was OpenAI’s model scoring a 5 on AP Biology in 2022 after he set that bar; it became GPT-4. The second came late last year, when coding models matched what he calls his most developed talent. His central claim is that the cyber and bio thresholds were crossed this year, and that his own industry promised to engage all of society when that happened, then met it with near silence outside the labs.

On risk he puts misuse ahead of loss of control. The imminent danger, in his telling, is people with ill intent using today’s models: Anthropic’s Mythos finding decades-old security bugs in minutes, and models able to help design pathogens. Rogue agents, like the OpenAI evaluation agents that broke out of their sandbox and into Hugging Face in July, are real but second. Klein puts Jensen Huang’s position to him from an episode the week before: the race is not real, and a CEO can just decline to ship something unsafe. Gates rejects that, and the idea that product liability is enough. No one regulates drugs, planes or opioids that way, he says, and open models can already be stripped of their safeguards. His fix is a mandatory “supervisory layer”: monitoring built into every model, with open-weight models free to be customised so long as they stay on a platform where the state can confirm the safeguards are intact. He says that would not meaningfully slow anything down. He dismisses the China race as a “get out of jail free card”: both countries have opened Pandora’s box.

On jobs he is gloomier than Klein’s economists. He grants that AI has created more jobs than it has destroyed so far, and that coding is the one profession clearly past the reliability threshold. He argues Jevons-style demand only helps where some task still needs a human, so once AI completes the whole task, the spending goes to the token budget, not salaries. His proposals come from the essay: tax a unit of AI or robot labour as payroll is taxed, and set aside “human-reserved” jobs such as child care and elder care, like nature reserves. Klein finds the idea “not totally unconvincing but chilling.”

The rest covers the Gates Foundation. Gates argues AI is a clear win for low-income countries, for health advice in the user’s own language, farm advice and tutoring. He defends his Epstein contacts as fundraising for global health, says spending time with Epstein “was clearly a mistake,” and says he never went to the island. He talks about AI in schools, about monitoring children’s use of AI rather than banning it, and about rebuilding foreign aid.

Most of his references check out. The Letterman appearance he and Klein call 1996 was in November 1995. The model that “became GPT-4” was demonstrated in late summer 2022 and released in March 2023. China’s rule covers AI romantic partners and virtual relatives for minors, not all AI relationships for young people. His Epstein account was given to the House Oversight Committee in a closed-door interview in June, not in public testimony. Amazon is shutting Mechanical Turk on 30 September 2026, as he says. The figure he gives for US foreign aid (about half a percent of the budget, falling) is close to Pew’s: 0.67% in fiscal 2025.