Ezra Klein calls for stopping recursive AI self-improvement, not just slowing down

Columnist Ezra Klein argues in an episode of his podcast of 20 September, based on his New York Times opinion column, that slowing down isn't enough. What should be stopped, he says, is recursive self-improvement: AI training the next AI. He proposes reversing the burden of proof: labs should have to show that what they are doing is safe, with narrow exceptions worked out with Congress.

Klein says he is agnostic on extinction and presents loss of human control as common ground between the US and China. He cites Sam Altman, who told Fortune it is hard to define what a ban would mean and that it probably wouldn't be enough. Klein finds that absurd. This is an opinion, and he says himself that stopping self-improvement would be only the beginning of an agenda. The Times is suing OpenAI and Microsoft, and discloses it.