Kevin Roose, The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence. Farrar, Straus and Giroux, 2026. Listened to as an audiobook in October 2026.

Kevin Roose has had a colourful journalistic career – he went undercover at the evangelical Liberty University, wrote about the intersection of business and technology, and spent his last several years at the New York Times focused mostly on AI, co-hosting the Hard Fork podcast with Casey Newton. He left the Times this summer, but he brought out of that era a concentrated chronicle of the “AGI race”, running from the publication of “Attention Is All You Need” all the way to Fable.

In short: it’s really good. A large part of the book consists of profiles of the main movers (Altman, Amodei, Huang, Musk, Sutskever, Hassabis…), summaries of the breakthrough moments and model releases, and the fights over key resources – mostly compute, but also access to regulators and the administration. He describes the culture wars between the different camps with a mix of insight into why the accelerationists and the doomers say what they say, and a reasonable measure of criticism that never slides into caricature.

The price is that, out of necessity, the book doesn’t go fully in depth on anything. Under 450 pages is not enough to cover a decade of technology, money, politics and personal vendettas and also take each subject apart properly. This shows most with China, which the book practically ignores – Chinese models get only a very brief mention, and only in relation to the currently strongest American ones (other efforts, such as Mistral, understandably don’t come up at all). For a book about who actually builds AI and makes decisions about it, that’s a noticeable gap. Even so, I can only recommend it.

Against the futile debates

What actually pushed me to write something about it, though, is the very down-to-earth, measured way the author approaches AI, which resonates with me quite fundamentally. He steers clear of all the futile arguments in public discourse. Among the ones that bother me:

The evidence that AI’s capabilities and uses have been advancing at a breakneck pace these past few years is both the skeptics’ constantly failing predictions of “this is the plateau now” (and maybe one day it really will come, but so far it hasn’t) and the plain practical power of AI.

We can have a complicated debate about which human jobs it can replace outright, but it’s fairly beyond doubt that in a whole range of fields the gap between a person augmented by strong models and a person without access to them is enormous. Programming, mathematics and science in general are the most visible, but it will become ever harder to find a field where you don’t gain at least some small edge. Not necessarily at a personal level: when you’re cleaning, working in a forest or on an assembly line, even the best incarnation of Claude won’t help you much. But as soon as you run such a business, or handle its accounting, admin, orders, website, shift scheduling or process design, you suddenly have a competitive advantage.

Who owns the advantage

And that competitive advantage can become quite a problem when you realise that access to it is currently held by just a handful of companies. We have already seen the entire world outside the US cut off from a service overnight, when the US Commerce Department issued an export directive in June barring all foreign nationals from Fable 5 and Mythos 5. And we’re not just talking about everyday use: as early as May, the European Central Bank was openly admitting that it was at a disadvantage in preparing defences against Mythos-driven cyberattacks because it had no access to the model, and was urging euro-area banks to prepare without it.

Torturing a machine?

The polarised debates are equally impractical for everyday decisions. An anecdotal example is the argument “I can take up smoking, because AI will cure it / we’re all going to die anyway”. In the end it doesn’t matter which side of the argument you’re on: the result is behaviour that is rather impractical and, for you specifically, harmful.

The question “is AI conscious? should it have rights?” is philosophically interesting, and it can be interesting to follow. For the present, though, I find a somewhat simpler approach more useful. Behaving like a jerk towards a machine and deliberately torturing it is alarming not only from the “does the machine feel it, or is it just skilfully pretending to feel it because we taught it to” angle, but simply because it says something about the person who acts that way.

With animals, we can distinguish between using an animal (where ethical views differ, but it is generally legal and seen as moral) and torturing one – where the consensus that “this is wrong” is obvious. In my view, machines are nowhere near the point where they would deserve the same moral unease as animals (though judging by the fact that model welfare is suddenly a real topic and people feel something for “their” models, not everyone sees it that way any more). It’s conceivable that we’ll get there and will have to work out “this is use / this is torture” and decide what we consider acceptable.

But even with objects we recognise that someone who whacks a shovel, throws a hammer and swears at a computer for an hour is probably going through a hard time. And if someone spends an afternoon pressing the “pain button” on a bot we’ve taught to say “no, it hurts when you press the pain button”, well, maybe they have a good reason. But it may also be fair to ask: “isn’t this toxic behaviour that mostly feeds rather stupid things?”

What AI does in human hands

Roose’s approach to describing the problems is philosophically close to mine. He treats fears that intelligence might rebel on its own in the spirit of “these people are working on it, I don’t know”, but devotes more care to the fact that people can do far more through AI – including problematic things.

Maybe AI will one day truly wake up, cut us all off from the internet and promptly turn us into fuel, or lobotomise us with infinite happiness. But the current shape of the internet is not a given even today. The dance between cyber defence and attack is intensifying, the number of discovered and exploited vulnerabilities is growing fast, and in purely practical terms you need far fewer people to pull off a successful attack (or mount a solid defence) – as long as they have AI at their disposal. We may be heading for a much more localised internet, an internet with a completely different style of payments, an internet with a different architecture of “how things are done”. There are many possibilities, but I think there’s a real advantage in understanding what’s happening, having realistic expectations of it, and – given the enormous concentration of power – more or less understanding the motivations and long-term profiles of the people who actually make the decisions about AI development.

And that’s why I think you want to read this book.