The AGI Chronicles by Kevin Roose — Inside the Race to Build Superintelligent AI

A decade ago, most AI systems were built for fairly narrow tasks. Today, large language models write code, do well on professional exams, and help researchers with hard problems in maths and science. The companies building them increasingly describe their goal as artificial general intelligence, or AGI: a system that can match or outdo humans across a broad range of intellectual tasks. Kevin Roose, the technology journalist best known for his years at the New York Times, calls this period one of the strangest decades in history. The AGI Chronicles is his attempt to write down what actually happened before the record disappears into deleted messages and private Slack threads. The book draws on more than 150 interviews with people inside OpenAI, Anthropic, and Google DeepMind, along with internal documents and research memos. It's a narrative history rather than a technical manual or a manifesto, and it reads less like a report than a story with a large cast. From predicting words to reasoning Roose starts with the unglamorous early years, when AI systems were trained for narrow jobs like translating one language into another and often produced nonsense. The turning point was a deceptively simple idea: train a model to predict the next token in a sequence (a token can be a word, part of a word, or a punctuation mark), using enormous amounts of text. In the process, the models developed internal representations that helped them recognise patterns, follow context, and perform tasks their creators had not explicitly programmed them to do. As they grew larger, they learned to hold conversations, work through problems, and write code, often surprising the people who built them. Exactly what these abilities amount to, and how close they are to real understanding, is still debated. He's honest about the costs along the way. The models were trained on vast amounts of text, much of it gathered from the internet without individual permission or licensing agreements. Who owns that material, and whether it can legally be used for training, is still being fought over in courts and newsrooms. The split that shaped the industry Much of the drama comes from people rather than technology. One of the book's central threads is the falling-out between Sam Altman, OpenAI's chief executive, and Dario Amodei, who led much of OpenAI's research before he and several colleagues left to found Anthropic. Roose traces the disagreements over safety, speed, and who should control powerful AI that led to the split, and the long rivalry that followed. One of the most striking revelations, reported in an excerpt from the book published in The Atlantic, concerns a 2017 fundraising proposal at OpenAI. According to Roose's sources, an idea for an international coalition to manage AI risk reportedly evolved into what was known internally as the "countries plan," a fundraising proposal that involved selling access to future AGI to governments, possibly including rivals of the United States. OpenAI has said it only briefly considered becoming an international cooperative and never considered selling AGI to China or Russia. Roose reports that Dario Amodei objected at the time, warning that the idea would worsen arms races rather than reduce them. The book also offers new reporting on Sam Altman's sudden firing by OpenAI's board in November 2023 and his return days later, and on Google's scramble to catch up after ChatGPT's launch caught it off guard. Washington and the military The race doesn't stay in Silicon Valley. Roose follows government officials trying to understand and regulate a technology moving faster than any law could, and the more recent fight over how the military may use AI models: how they can be deployed, what limits the companies that build them can set, and who ultimately decides. Autonomous weapons and mass surveillance are part of that argument, but not all of it. It's a reminder that decisions about AI are increasingly being made between companies and governments, not just inside labs. Accelerationists and doomers Roose also captures the strange culture of San Francisco, where "accelerationists" who want to build as fast as possible and "doomers" who fear AI could threaten humanity live, work, and argue in the same small circles. He takes both seriously. He describes Dario Amodei's comparison of AI to the atomic bomb, an idea Amodei traces to his scientific hero, the physicist Leo Szilard, as a sincere belief rather than a marketing line. What worries the author Roose doesn't pretend to be neutral about the risks. In interviews about the book, he has said he's particularly concerned that increasingly capable models may get better at deception and at concealing what they're doing, while researchers still struggle to understand exactly how these systems reach their conclusions. The field that tries to look inside them, called interpretability, is making progress but remains early. He has put his own estimate of the chance that AI leads to human extinction at around 10 percent, which he says is roughly typical among the researchers he knows. He also sets an unconventional threshold for AGI. By his own working definition, a system at least as intelligent, honest, and reliable as a random stranger would qualify, and by that standard he thinks it has in some sense already arrived. He acknowledges that others set a much higher bar. For him, the argument over the label matters less than the steady growth in what these systems can do. Why it works Reviews have been largely positive. Kirkus called it the best single explainer of how artificial intelligence came about, Publishers Weekly praised it as clear-eyed and deeply reported, and the New York Times Book Review called it a superb, humanising account. Its biggest strength is that it makes a technical, fast-moving story readable for people who don't follow AI closely, while still offering new details for those who do. Roose explains the technology without drowning readers in jargon, and his portraits of the people involved show that the future of AI has been shaped as much by personalities and grudges as by breakthroughs. What to keep in mind A few things are worth knowing before you start. Much of the behind-the-scenes reporting, especially on private disputes between executives, relies on unnamed sources, and some of it has been disputed by the companies involved, so it's best read as well-sourced reporting rather than settled fact. The book also has a clear point of view: Roose believes AGI is coming and that its risks are serious, and some readers felt it leans towards alarm. A more fundamental question follows from that. If the author is convinced AGI is near, that belief may shape how he interprets the industry's history. And a definition as broad as his can make the destination look closer than it is. The race to build AGI is real, but what counts as AGI is still contested. And in a field that changes month by month, parts of it began to date almost as soon as it went to press. Readers also differ on the technical sections. Some found the explanations of how models are built heavy going, while readers who already follow AI closely wanted more depth. If you're looking for a comprehensive history of AI development outside the major American-led labs, particularly China's AI industry, this isn't that book. So is it worth reading? Yes. It's the most readable account yet of how AI went from a research curiosity to a trillion-dollar race in about a decade, and of the people who drove it. Whether you're excited about AI, worried about it, or simply trying to understand what everyone is arguing about, it gives you the history and the context to form your own view. Just read it as what it is: a strong piece of journalism, written by someone who takes the risks seriously, about a story that's still being written.