Richard Socher: No realistic AI-wipes-out-humanity scenario

Socher's post lands in the middle of a public argument among senior AI figures about how much to slow down. In the three days before his Wednesday post, Anthropic CEO Dario Amodei published a 3,800-word essay arguing that fully addressing the risks requires pacing the rate of capabilities advancement so risk prevention has time to keep up, and he doubled down on that at Salesforce's annual Dreamforce event. OpenAI CEO Sam Altman said he agreed with Amodei, then followed up with a dark X post of his own warning that AI progress could go very badly in two ways and that AI must always serve people.

Socher's contribution is a different kind of objection. He is not arguing that AI causes no harm; he says he has talked to many AI safety experts about existential risks and that real researchers point to real and likely harms that should be taken seriously. His target is the extinction case specifically: when he asks for a detailed scenario that would kill every single human, he says it usually starts and ends with a call to look at current progress and use your imagination, with no science and no scenario planning behind the probabilities.

That makes the disagreement less about whether AI is dangerous than about what counts as evidence for the most extreme danger. The exchange is likely to matter most for the people who decide which risks get research money and regulatory attention: if the extinction argument is dismissed as handwaving, the burden shifts toward demonstrating concrete, near-term harms. How the debate settles may hinge on whether safety researchers answer Socher with an actual detailed extinction scenario, rather than on the rhetorical strength of either side.