The AI race goes to conviction, not to those with the most compute
Google invented Transformer, mesh tensor flow, and the entire modern AI stack, yet fell behind due to internal disbelief in AGI; what truly separates AI labs is conviction that AGI will arrive—not money, GPUs, or talent.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Belief, not technology, explains Google's fall
Kevin says the book is fundamentally not a technology history but a history of belief: the real dividing line between AI labs is not money, GPUs, or research talent, but whether they believe AGI will arrive—because only belief enables leaders to persuade others and secure resources and focus. Google is the textbook counter-example: it invented Transformer, mesh tensor flow, and nearly the entire modern AI stack, yet internally erupted in memo wars—one faction urging more AGI talk, another insisting it be avoided. Without conviction, it ceded first-mover advantage to those who had it.
— Kevin RooseTop lab CEOs are humanity's most competitive individuals
Kevin says if you ranked the world by competitiveness, Demis, Dario, and Sam would place in the top ten. He offers concrete evidence: Demis once vanished for a weekend—he'd secretly entered a professional poker tournament and won, despite being a novice player. He was a childhood chess prodigy, accustomed from early years to beating opponents far older than himself. Kevin sees this extreme competitiveness as the core trait driving them to persist for over a decade in an age when no one believed in AGI, rather than mere scientific idealism.
— Kevin RooseThe Manhattan Project model fails when applied to AI
Demis often compares DeepMind's work to the Manhattan Project; Dario imagines himself as Szilard. But Kevin identifies a critical difference: the atomic bomb, once built, needed no further revision—government officials could thank the scientists and hand everything to the military to manage. AI models demand continual post-training, iterative versions, and bug fixes—no one in government has the technical chops to take custody of a model checkpoint and continue training it. This means AI lab founders will not be sidelined the way Manhattan Project scientists eventually were.
— Kevin RoosePeople hate AI in words but depend on it in practice
Jordan asks whether people truly hate AI as much as they claim. Kevin judges it to be aesthetic dislike for now, but if unemployment ticked up two percentage points, that dislike would become real, politically destructive anger—at 10% unemployment, it might spark street riots or violence. Yet both agree that even then, programmers won't return to writing Python by hand, and lawyers won't abandon ChatGPT and Claude. True AI contraction would require organized professional resistance like legal licensing barriers, not sentiment.
— Kevin RooseNobody can fully explain why Dario is the hardest China hawk
Kevin notes that of the three leaders, Dario holds the most hawkish China stance, and Anthropic has a visibly lower proportion of Chinese employees than other labs—yet no one can fully explain why. The only thread is that Dario's first tech job after his PhD was at Baidu's Silicon Valley office; from that experience onward, he became convinced democratic nations must reach AGI before China. Chinese internet rumor points to a romantic heartbreak involving China. Kevin himself emphasizes he's speculating, not stating fact: Dario may be unsettled by not knowing how the systems he trains will ultimately be used.
— Kevin RooseNo scientific theory explains why scaling makes AI smarter
A paper written by Noam Shazeer after he left Google ends by attributing it all to ‘divine benevolence’—not as metaphor but as literal belief: no one can scientifically prove that training larger models on more data produces smarter outputs; it feels like a fundamental law of nature, like no one knows why hydrogen and oxygen bond to form water. This remains AI's deepest mystery.
— Kevin RooseAnthropic approached Trump Jr. for cap table equity but was refused
Jordan raises a detail he believes should have made the book: early on, before Pentagon restructuring and other Trump administration moves, Anthropic pitched to put Trump Jr. in its cap table, but he declined. The two discuss that Musk was the more pivotal variable at that moment—he was Sam's rival and served as an actual gatekeeper early in Trump's term, which explains why AI labs and the White House started off on delicate terms.
— Jordan SchneiderAI will automate its own research and trigger recursive improvement by 2027
Kevin reports that a thesis once considered fringe—that recursive self-improvement arrives by 2027—is now being told to him as plausible by serious people. Full automation of AI's own R&D could arrive as soon as next year. Even if progress stopped today, society and institutions would need roughly a decade to digest existing technology. If recursive self-improvement truly triggers, models might become significantly stronger roughly every six days—a pace of instability humanity has never weathered.
— Kevin RooseIn their own words · checked verbatim
Because only by believing and then only by selling that belief to others can you amass the resources and the focus to to get to AGI.
Kevin Roose9:06
Like they are some of the most competitive people who have ever lived.
Kevin Roose18:37
The generals really take over. They're like, thank you for your work. We'll take it from here.
Kevin Roose21:52
I think there's like an aesthetic hate. But if unemployment goes up by two points, that turns into an extremely real hate.
Kevin Roose26:58
We have no explanation for why this works.
Kevin Roose38:25
What could possibly be more important than teaching sand to think?
Jordan Schneider47:33
they offered to let uh trump jr on the cap table and then he said no
Jordan Schneider50:41
this guy cannot tell a lie to save his life
Kevin Roose52:48
Figures
| Company valuation mentioned | 2 trillion USD | 16:18 |
| Unemployment threshold for social upheaval | 10% | 28:04 |
Glossary
- RSI (Recursive Self-Improvement)
- A feedback loop where AI improves its own research capabilities, accelerating its progress.
- LessWrong
- An online community discussing AI risks and rational thinking; an early gathering place for AGI believers.
- cap table
- A record of a company's shareholders and their equity stakes.
- LARP
- An immersive role-playing game where participants act out historical or fictional characters through interactive play.
How to listen
Entrepreneurs and investors tracking AI lab power dynamics and interested in Silicon Valley leadership psychology and China-US AI narrative divides.
The closing Easter egg section mimicking Lenin rap has low information value and can be skipped.