THE AGI CHRONICLES

A record of “the strangest decade in our history,” when computational statistics morphed into AI—for better or for worse.

Roose, a technology columnist for The New York Times, examines the period between 2015 and the present (with forays into past and future), when “the human monopoly on intelligence ended, and a new world was born.” He charts that birth, beginning with machines that spat out nonsense words in their infancy, training for highly specific tasks (say, translating English into French). They were further trained to predict what word might come next in any given sentence, with researchers assuming “that a model forced to get very good at next-­word prediction would have to develop a rich understanding of the world along the way.” With improved models and a vast corpus of literature to devour, if acquired without respect to intellectual property rights, we are approaching AGI, or artificial generative intelligence, by which machines become autonomous to some degree. As the author writes, this prospect represents a very real danger; one of his principals expresses fear that AI/AGI “could lead to death and destruction” if in the wrong hands. On that note, pressured by the Trump administration, several of the leading firms have bowed to the demand that their models be made available for military use. And, although CEOs, such as OpenAI’s Sam Altman, have held out for restrictions—“no domestic mass surveillance, no fully autonomous weapons”—the Pentagon has insisted on applying it to “all lawful purposes.” In a dense but generally accessible narrative—albeit stuffed with head-spinning journeys into how AI/AGI models are constructed and trained—Roose arrives at a worrisome conclusion: “I am worried about the harms these brilliant machines may cause, and I am concerned that we are running out of time to align them with our values.”

The best single explainer of how artificial intelligence came about, and what might happen next.