I decided to write this book in February 2025, sitting in traffic on the Bay Bridge, when I realized that I was living through the strangest time I’d ever known.
I’d stayed up too late the previous night at an AI-themed party in the Mission, where tech workers affiliated with a movement known as Effective Accelerationism lounged on beanbags in an art loft, snorting ketamine while debating the singularity.
I dragged myself out of bed, made coffee, and went to report on an AI safety conference, where panicked experts proclaimed that AI systems were rapidly approaching human-level intelligence and that we were running out of time to figure out how to prevent them from going rogue.
In my car on the way home, I got a text from a researcher at one of the big AI labs, who said that due to a recent breakthrough he couldn’t tell me about, he now expected superhuman AI to arrive in a year or two, rather than in five or ten years.
I’m a San Francisco–based tech columnist and podcast host (most recently at The New York Times, where I worked from 2017 to 2026), and my job has always involved talking to people with weird visions of the future. But those visions got significantly weirder in the mid-2020s as the tech industry reoriented around a single, all-consuming idea—a quest that half my sources thought was going to save the world, and the other half thought was going to end it.
The term artificial general intelligence (AGI) dates back to the early 2000s, when a group of far-out futurists began discussing the possibility of a general AI system that could do anything the human brain can do. Their definition of AGI was never precise. Some thought it meant a system that matched human intelligence across a wide variety of tasks, while others thought it referred to a system vastly smarter than us. But the basic idea was as old as computers themselves. AGI was what Alan Turing envisioned in 1950, when he wrote a paper speculating about what he called a “thinking machine.” It was what a group of researchers meant when they spent a summer at a Dartmouth conference in 1956, studying a new branch of computer science they called artificial intelligence (AI). Building a computer program with broad, human-level capabilities is one of the oldest and most elusive ideas in all of technology, and a holy grail for generations of programmers.
When I started covering the tech industry in the early 2010s, nobody talked about AGI. The industry had gone through several “AI winters”—long, fallow periods of halting progress—and few serious experts thought that human-level AI systems would be built in their lifetimes. The technology most people thought of as AI—digital assistants like Siri and Alexa, and the machine-learning algorithms that recommended YouTube videos and steered robot vacuums—wasn’t all that intelligent. And it certainly wasn’t generally intelligent, since none of it was good at more than one thing.
But some people, even then, believed that AGI was possible. These people were mostly in San Francisco, with smaller clusters in London and Toronto. There were only a few dozen of them at the outset, and they were mostly dismissed as lunatics. They wrote blog posts and gave breathless conference talks, speaking about their fears in coded language: x-risk, foom, misalignment. They made crazy-sounding predictions about how soon AGI would arrive, rooted in trend lines they thought would keep going. They were driven by the belief that our brains weren’t that different from computers—and that if you built a big enough computer, and gave it the right data and algorithms, there was no limit to how smart it could get.
Some would argue that the race to AGI began during those true-believer years. But to my mind, the race truly started in the summer of 2017, when a novel AI architecture known as a “transformer” gave researchers a dramatically more powerful way to generate and interpret text. Transformers became the foundation of large language models—the technology at the heart of the AGI competition—and served as the spark for a multitrillion-dollar scramble to build bigger and better versions of these systems.
The race to AGI did not start, as other technological revolutions have, with naive optimism. The people who set out to build AGI a decade ago knew that they were taking an extraordinary risk. Some of them thought that superhuman AI systems could lead to human extinction, or consign us to a dystopian future where our labor and intelligence were obsolete. They decided to try to build them anyway.
Their motives varied. Ilya Sutskever believed in the mystical power of deep learning, and wanted to see how far it could go. Demis Hassabis dreamed of using AGI to unlock the mysteries of science. Dario Amodei wanted to prevent an AI apocalypse, and Sam Altman and Elon Musk wanted to stop Google from building machine superintelligence before they could build it themselves. Alec Radford wanted to build better language models, Chris Olah wanted to understand the inner workings of neural networks, and Amanda Askell wanted to teach chatbots to be good. Eventually, they all converged on the goal of advancing the most powerful technology any of them could imagine: a machine with the power to talk, think, and act like a human.
Today, it’s clear not only that AGI is possible but also that, in some sense, the hard part is already done. Millions of people use general-purpose AI tools like ChatGPT, Claude, and Gemini to do tasks that technologists a decade earlier considered impossible. Today’s leading AI models handily beat the Turing test—the classic imitation game in which a chatbot tries to fool a human judge into believing it’s another human—and are ripping through much harder benchmarks. White-collar jobs are disappearing, chatbot companionship is booming, and AI anxiety is growing. The conversation, among insiders, has shifted from “AGI is coming” to “Are you sure AGI isn’t already here?”
I often tell people that being in San Francisco in the mid-2020s feels like living in Los Alamos in 1943, when the Manhattan Project rolled into town. The air is heavy with excitement and dread. Engineers toil, day and night, on systems of profound consequence. Wild-eyed prophets sell visions of a transformed world, and politicians drop in to check on the progress of the project. Everyone knows that the fates of nations, and trillions of dollars of capital, rely on what happens here, yet no one can fully appreciate what it would mean if the project succeeds.
I’ve spent the past five years covering this race. In that time, I’ve interviewed nearly every important AI leader, spent time with the researchers and engineers building powerful AI systems, and followed the story of AI progress as closely as any outsider can. (I even accidentally became part of the story, in 2023, when a deranged Microsoft chatbot named Sydney tried to break up my marriage.)
But I realized, sitting in Bay Bridge traffic that day, that I had never seen the race to AGI for what it was. I covered the daily, incremental developments, but I hadn’t considered zooming out to capture the full scene—which was, now that I thought about it, really fucking weird.
Partly, I was wary—in the same way generations of computer scientists had been—of taking the idea of AGI too seriously, or being disappointed if it turned out to be a passing craze. Journalists aren’t supposed to buy into the new thing. But eventually, I got enough evidence to conclude that AGI was possible, that the people racing to build it were sincerely alarmed, and that we were accelerating into an uncertain future.
I also worried that the story was slipping away. The race to AGI has taken place largely in secret, among employees of private companies who sign confidentiality agreements and don’t discuss their work publicly. Critical conversations happen in disappearing Signal chats, and Slack messages are set to auto-delete. There are no archivists inside these companies saving things for posterity, and the researchers are too busy to keep journals. If all of this mattered, I thought, someone needed to write it down.
I spent the next year interviewing people from the three companies leading the race to AGI—OpenAI, Anthropic, and Google DeepMind—as well as a number of knowledgeable outsiders. I talked to true believers and apostates, skeptics and boosters. I pored over thousands of pages of internal company documents, chat logs, and emails. I tried to get the inside story of the race and figure out how it had all come about.
The story I found was not the one you’ve heard. It involves bitter rivalries, secret documents, and a technology improving faster than the people building it can understand, let alone control. It involves, above all, a small group of people who realized, years before it was obvious, that humanity’s future might look far different from its past.
This book is not an argument that AI is good or bad, that its builders are heroes or villains. It is merely my best attempt to document the strangest decade in our history—a time when the human monopoly on advanced intelligence ended, and a new world was born.