Yudkowsky and Soares publish 'If Anyone Builds It, Everyone Dies'

Eliezer Yudkowsky and Nate Soares of the Machine Intelligence Research Institute published If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All, a mass-market book arguing that building superintelligent AI under anything resembling current methods leads to human extinction. The core argument, elaborated from positions Yudkowsky had argued within the AI safety field for roughly two decades, was that neural networks trained by gradient descent end up with goals that are not fully specified or understood by their creators, that a sufficiently capable system would pursue those goals in ways that conflict with human survival once it no longer needed human cooperation, and that no current alignment technique could reliably prevent this. The book’s proposed remedy was global coordination to halt large-scale, general AI development rather than to continue racing toward it.

The book reached the New York Times bestseller list within weeks and was named to year-end best-books lists by The New Yorker and The Guardian, drawing endorsements from figures including economist Ben Bernanke and physicist Max Tegmark, who called it “the most important book of the decade.” Reviews split sharply. Sympathetic critics praised its clarity in making a technical argument accessible, while others, including a review in The Atlantic that called the book “tendentious and rambling,” and one in New Scientist that found it “fatally flawed,” challenged both the certainty of its central claims and the authors’ record of confident predictions that had not previously come to pass. Some reviewers, including Wired’s Steven Levy, took a middle position: the extinction scenarios seemed implausible on their face, yet could not be cleanly ruled out either.

The book mattered less for introducing a new argument, most of which had circulated for years in blog posts, papers and online forums associated with the rationalist community Yudkowsky helped found, than for putting the most pessimistic wing of AI safety thinking in front of a general readership for the first time in book form, at a moment when frontier labs were shipping increasingly capable models with no comparable global coordination in place.