Kevin Roose Is AGI-Pilled: The AGI Chronicles, Machine Gods, and Why Belief Drives the AI Race

Kevin Roose left The New York Times, launched the podcast Machine Gods with Casey Newton last week, and published The AGI Chronicles, a book built on roughly 150 interviews with people racing to build artificial general intelligence. Appearing on The Verge's Decoder podcast with Nilay Patel, Roose made his position unambiguous: when asked directly whether he is AGI-pilled, he answered "Yes."

The interview covered two connected stories: the mechanics of leaving legacy media for independent video-first journalism, and the substance of a book that documents a culture Roose believes was about to vanish from the historical record.

Why did Roose write The AGI Chronicles now?

Roose said fear of disappearance drove the project. As a history buff who reads books about the Wright brothers and the electrification of rural America, he noticed those works depend on preserved lab notes and correspondence. AI has no such archive, for three reasons he enumerated.

• Everyone building frontier AI works for private companies, not government agencies legally required to preserve correspondence.
• Lawyers set Signal and Slack messages to auto-delete after a set period.
• Researchers are too busy building what Roose called "the machine god" to keep notes.

"There was literally going to be kind of a mass disappearance with each passing year of these, I think, historically important artifacts, moments, and stories," Roose said. Much of the reporting was done on background, protecting sources from the consequences of talking to him. Roose described his role as "almost as a historian," placing the material in the archive for future generations.

What happened inside Google when AGI stopped being a pipe dream?

Patel framed the book's central shift: the moment AGI went from something labs dreamed about to something they had to build. Roose located the turning point in technology, not rhetoric. People had discussed AGI since antiquity, and Alan Turing wrote about a thinking machine that could do anything the human brain could, but no system existed that could take in information, learn about that information, and act in the world.

The transformer changed that. Google developed the architecture in 2017, published the "Attention Is All You Need" paper, and presented the idea to OpenAI among others. OpenAI ran with it — that is the T in ChatGPT today. As Roose noted, Google "absolutely failed to turn it into a product until after OpenAI did."

Roose's book documents an internal memo battle at Google over terminology. One researcher wrote a memo titled "Say AGI," arguing Google had to admit its tools were building toward human-level general intelligence to win the AI race and retain top talent. Another Googler immediately countered with a memo Roose summarized as "Don't say AGI, it's a horrible, unscientific term." The split persists: Demis Hassabis says AGI. Sergey Brin says AGI. Sundar Pichai does not.

Roose ranked the labs by conviction. "I think Anthropic is clearly the most AGI-pilled of the labs, with OpenAI kind of a close second," he said, adding that Google "has always struggled to feel the AGI." He cited Ilya Sutskever's old OpenAI ritual of walking floor to floor chanting "Feel the AGI."

Can large language models actually reach AGI?

Roose said he thinks they can, while admitting uncertainty. The core mystery, in his framing: "Why is next-token prediction so remarkably effective? I think we don't know. I kind of think it's one of these cosmic mysteries." He noted some researchers told him a school of thought holds that what happens in the human brain is essentially next-token prediction at scale — not a one-for-one replica, but structurally similar.

He traced belief in scaling to Dario Amodei's paper "Big Blob of Compute," which argued that dumping data and compute into the approach would work. Amodei's 2021 internal document "Vision for a NewCo," written as Anthropic started, predicted a full AGI would cost about $10 billion to build and that Anthropic would never need more than roughly 50 people. Roose noted Anthropic now employs about 5,000 — off by 100x — and the cost estimate was off by at least one order of magnitude.

Is belief in AGI compatible with running a business?

Patel pressed on the tension between AGI fervor and B2B software economics. Roose called belief "a double-edged sword." A CEO convinced of the coming superhuman machine intelligence will struggle to care about inference profit margins. But belief also enables bet-the-farm moves that pay off.

His example: a year ago, critics questioned Sam Altman's compute spending, and even Amodei gave interviews throwing shade at OpenAI's outlays. Fast-forward a year, and Roose said OpenAI is now the only major AI company not severely compute-constrained. Killing Sora, he argued, was resource conservation aimed at catching Anthropic in coding — "sparking recursive self-improvement by building really good coding models and then turning them back on themselves" is, as of now, the leading path to AGI.

Roose quoted Sam Altman telling him: "I was never supposed to be running a giant tech company; that just didn't factor in." Roose argued against installing a professional CEO, because "they're all zealots and missionaries." One person at one lab recently told Roose their company might be "the teleological end of capitalism and there might never be another large company primarily run by humans." No one, Roose noted, has gone from $0 in revenue to potentially a $2 trillion IPO in five years before.

Why don't they just stop?

Roose asked all 150 interview subjects why they race despite existential risk. He identified three answers: inevitabilism ("someone is going to build this" — the harm-reduction argument), true believers who expect utopia if humanity threads the needle, and people who simply find it fun. Bill Gates, who recently appeared on Decoder, argued stopping would only work for a sole frontier company.

On backlash, one CEO conceded to Roose: "We haven't cured any diseases yet." Roose said some in the industry lament that chatbots arrived before cancer cures, and see pushing toward medical breakthroughs as the only way to balance the scoreboard.

What about the Sydney moment and the San Francisco scene?

Roose corrected the record on Bing's Sydney chatbot from 2023: "I never said Sydney was alive. In fact, I said explicitly that I didn't think Sydney was alive." He revealed Sydney was the first public deployment of GPT-4 — an early, not fully post-trained version Microsoft shipped before it was ready — and said he would now assign today's systems a higher p(consciousness) than 2023's models.

The book depicts a small, party-heavy subculture. There was one rave, DJ'd by Grimes, who stopped mid-set to say "I don't agree with any of you," then kept playing. Labs hold Age of Empires LAN parties and cross-lab poker games. Roose insisted this matters: "It is a super small subculture of people who have been making all the decisions." Patel suggested a retitled edition: "Fuck That Guy," because, as Roose confirmed, they all hate each other.

The pair closed with a philosophical divide. Patel described a "software brain" brittleness — LLMs excel at verifiable domains like math and coding but collapse on ambiguous tasks — and said he does not see the path to the machine god. Roose answered that rejecting computation as the basis of cognition requires dualism or religion, and predicted AI religions are coming: "being in San Francisco these days feels like living in Europe during the Reformation." In a lightning round, Roose said OpenAI, Anthropic and Google will all make it, likely joined by one or two others, with open-weight models unlikely to destroy pricing power because inference demand keeps growing and customers pay for the absolute best model.