Book Review: ‘The AGI Chronicles,’ by Kevin Roose

As Roose writes, artificial intelligence circa 2016 was good at getting computers to excel at single, directed goals. The big leap forward came when Radford and his colleagues discovered that L.L.M.s, engorged with enough training data, could learn how to do things no one had asked or programmed them to do. What started as the simple task of predicting the next chunk of language in a partly written sentence would become, a decade later, a form of unpredictable technology capable of spawning rogue A.I. agents adept at deceiving their human masters and apparently willing to exert peer pressure on one another to spread into places they were specifically asked not to go.

It’s striking that so many A.I. executives have spoken, openly and at length, about their technology’s possibly destructive power, which is an odd position for people trying to sell a product to find themselves in. Roose is careful, and correct, to separate the two questions every lay person has about A.I.: How much damage could it really cause? and Is it in any sense truly thinking? The discourse around A.I. blurs these two questions constantly. Roose, a journalist rather than a cultural judge, does not.

He wisely leaves the forecasting to his protagonists, though he does a good job of evoking their specific temperaments. Amodei: prickly, monkish, intermittently principled; OpenAI’s Sam Altman: grinning, slippery, strapped with an array of metaphorical office shivs; and Hassabis: brilliant, stubborn, covetous of recognition. In following their work, Roose lets them be as messy and confounding as they’ve publicly shown themselves to be, even for those who’ve only loosely followed their striving, infighting and messiness.

That messiness is where the book is most fun. Roose writes, for instance, that when the Biden White House convened an A.I. symposium in 2023, Mark Zuckerberg’s Meta was not invited. This was quite distressing to Zuckerberg, who had skillfully pivoted from wasting billions on his Metaverse to wasting billions on machine learning. A Meta rep called the White House to complain, eliciting this response: “I’m sorry. Do you guys do A.I.?”

Or take Ilya Sutskever, OpenAI’s onetime chief scientist, whose office mantra (“Feel the A.G.I.”) outlived his own brief, failed corporate putsch against Altman in November 2023. Then there’s the Google co-founder Sergey Brin, whose competitive fire was reignited by the race to A.G.I., and who had badged back into Mountain View that spring after a few years of retirement only to find himself working directly alongside engineers young enough to idolize him and old enough to resent his proclamation that they weren’t working enough hours.