Ask any frontier lab why it is moving so fast and you get the same answer. Because the others are.
OpenAI was founded in 2015 as a counterweight to Google. Anthropic was founded in 2021 by people who left OpenAI. xAI was founded in 2023 by a man who had left OpenAI's board. Meta rebuilt its whole AI organization in 2025 because it felt it had fallen behind all of them. Each lab exists because of another lab. Each one says it would slow down if only the others would.
Read that as strategy and it looks like a coordination problem. Read it as René Girard would, and it looks like the oldest story humans tell.
I. We don't know what to want
Girard was a French literary critic who taught at Stanford and died in 2015. His big idea fits in a sentence. We don't choose our desires. We copy them.
Not our appetites. You don't need anyone to teach you to be hungry. But which food, which job, which partner, which company to found: those come from models. People we admire, envy, or measure ourselves against. They want something, and their wanting makes it glow.
The philosopher Jonathan Bi, who has lectured on Girard for hours, gave the cleanest version on the Jack Neel Podcast. The part most people miss is the order of operations:
"Whereas most people think scarcity precedes competition, Gerard actually says competition precedes scarcity. In other words, competition is primary. And then when we end up competing over the same things, that's what creates scarcity. So in other words, most of the scarcity is artificial."
Jonathan Bi, Jack Neel Podcast
Bi's example is Beverly Hills real estate. You could build an equally beautiful house a few miles away. Nobody wants to. The value is the crowd.
Now look at AI. The scarce things are GPUs, a few hundred top researchers, and the title of "first." All three are scarce mostly because everyone decided at the same moment that they were the only things worth having. When Peter Diamandis recapped the summer of 2025 on Moonshots, the numbers sounded like a bidding war at an auction for a painting:
"Then came the talent wars. If you remember, Sam and Meta were both offering people like $100 million signing bonuses, insane."
Peter Diamandis, Moonshots
A $100 million signing bonus is not a price set by output. It is a price set by the fact that a rival wants the same person. That is the mimetic signature: the object's value rises with the number of hands reaching for it.
II. The doubles
Here is the part of Girard that should scare people more than any capability chart.
When two people want the same thing, they start watching each other instead of the thing. The object fades. The rival takes over. And because each copies the other, they become more and more alike. Girard called them doubles. Each one sees the other as the aggressor. Each one sees himself as merely responding.
Listen for that structure and you will hear it everywhere in the AI debate. Patrick Bet-David laid out the standard objections to slowing down, and the clincher was the one everyone reaches for:
"The other one is if we don't build it, China builds it."
Patrick Bet-David, PBD Podcast
His guest, Jacob Coxon, a former Anthropic employee who left to warn about the technology, gave an answer that is either naive or the most Girardian line in this whole essay:
"I think actually our incentives are aligned. Like we're more similar to China in this respect. So it should be quite easy to cooperate."
Jacob Coxon, PBD Podcast
Coxon is right that the two sides are similar. Girard would say that's exactly the problem. Similarity is not what ends rivalry. Similarity is what rivalry produces. The closer two rivals get, the more each one needs to beat the other, because there is less and less else to tell them apart.
Bet-David's own follow-up proves the point. Maybe, he suggested, China wants the US to slow down so China can catch up. Every gesture of restraint gets read as a move in the game. That is how doubles think. Nothing the rival does can be innocent, because you know what you would mean by it.
In his last book, Battling to the End, Girard took Clausewitz's phrase "escalation to extremes" and argued it was the master pattern of modern conflict. Each side matches the other and then adds a little, because adding a little is the only way to stay even. Swap "nuclear warheads" for "training compute" and the book reads like a newsletter.
The race isn't a side effect of wanting AI. For a lot of the people in it, the race is the wanting.
III. Is competition the disease or the cure?
The most interesting fight in the debate right now is not doomers versus optimists. It is a fight about whether competition itself is the danger.
On All-In, one of the hosts took aim at what he called "the Dario position," the Anthropic CEO's long-running worry about a race to the bottom:
"His fundamental contention is that that will lead to a race to the bottom. That competition jeopardizes safety... I just fundamentally disagree with the idea that competition means lack of safety."
All-In Podcast
His argument: customers want safe products, liability punishes unsafe ones, and the Cold War showed that markets deliver what people actually want, beauty and safety included.
Girard would split the difference in a way neither side would enjoy.
Markets are one of the great inventions for managing mimetic desire. They work by multiplying objects. If you want the car your neighbor has, the market builds another car. Rivalry turns into variety. That's the All-In host's point, and it's a good one.
But markets only defuse rivalry when the object can be copied. They fail when the object is singular. There is only one "first." There is only one lab that gets to say it built the machine that builds everything else. If the prize is framed as winner-take-all, then no amount of customer preference turns the race into a bazaar. It stays a duel.
So the real question is not "competition, good or bad?" It is what the competition is for. Competition over who ships the most reliable coding assistant is the market doing its job. Competition over who reaches superintelligence first is two rivals staring at each other across a single object. The first is a market. The second is a feud. Most of the current rhetoric is the second wearing the clothes of the first.
IV. Close, far, and marked
Girard's second big idea is darker. When mimetic rivalry spreads through a group, it becomes a war of all against all. Societies, he argued, survived this by finding one victim, and everyone blamed that victim at once. The unanimity did the healing. The victim did not have to be guilty. They just had to be agreed upon.
Bi listed what makes a good scapegoat. It must be close to the group, far from the group, and carry a special mark:
"They need to be close to the group. But number two, they also need to be far from the group... And the third thing is that they need a kind of special mark that differentiates them from the group. So that people's eyes naturally converge towards them."
Jonathan Bi, Jack Neel Podcast
Run AI through that test.
Close: it is in every office, every phone, every classroom. Far: it is not one of us. It has no family, no neighborhood, no vote. Nobody has its back. Marked: it is literally labeled. Every image, every essay, every customer service reply now gets the question "was this AI?"
It is close to a perfect candidate. And the blame has already started flowing toward it, some of it fair and some of it not. Andrew Yang named the unfair kind on Surrounded:
"There is something called AI washing, which is like, hey, I overhired, I managed my company poorly. So I'm now gonna fire a bunch of people and I'm gonna blame AI. That is totally happening."
Andrew Yang, Surrounded
Yang immediately added the other half: real companies are replacing 500 support workers with 50. Both things are true. That's what makes the mechanism work. A scapegoat is rarely pure invention. It is a real thing that absorbs more blame than it earned, until the excess becomes invisible.
The people around AI are candidates too. Billionaire founders. "Doomer groups," as one All-In host described the safety crowd, supposedly "carefully planning for weeks." "Fear porn," as Bet-David put it about the labs' warnings. Each camp has found its villain in the other.
But Bi adds a warning that matters here. A scapegoat only ends conflict if everyone agrees. If half the crowd blames the victim and the other half defends them, you haven't ended the war. You've just added another front. That's where the AI debate sits. The doomers blame the accelerationists, the accelerationists blame the doomers, and the public blames the machine. Three scapegoats is the same as none. No peace arrives.
V. Everyone is protecting the victim
Girard believed the Gospels did something new. They told the scapegoat story from the side of the victim, and once you see it that way, you can't unsee it. That revelation, he thought, is why the modern world is so allergic to open persecution.
But it created a new trick. If you can no longer say "I'm attacking this person," you say "I'm protecting the victims from this person." Bi called this Girard's version of the antichrist:
"The antichrist is going to use the rhetoric of protecting victims to persecute victims... In the modern West today, the only way you're allowed to persecute someone is to say you're doing it in the name of protecting victims."
Jonathan Bi, Jack Neel Podcast
Peter Thiel, who studied under Girard at Stanford, has been giving public lectures on the antichrist. His version points in a specific direction: a world government that arrives promising safety from catastrophe, including from runaway technology. In that reading, the AI safety movement's call for global oversight is the danger. Bi said on the show he isn't very familiar with Thiel's version, and Thiel's lectures go beyond Girard. But the structure is borrowed.
Here is what makes the AI debate so stuck. Both sides speak entirely in the language of protecting victims.
Safety advocates protect humanity, children, workers, future generations. Accelerationists protect the sick who will die waiting for cures that slower AI would have found, and the poor who would gain most from cheap intelligence. Each side's best argument is a victim. Each side sees the other as the persecutor in disguise.
Girard doesn't tell you which side is right. He tells you something harder. When every party in a fight claims the victim, the claim stops being evidence. You have to look at what each side would actually do with power, and who would pay.
VI. The imitation machine
There is a strange loop at the center of all this.
Silicon Valley is the most Girard-literate industry on earth, mostly thanks to Thiel. Bi tells the story of how Thiel saw Facebook's like button: a "mimesis of mimesis." Liking is already copying. Displaying the likes for everyone to see doubles it. Thiel put roughly $500,000 into the company. The industry got rich by building machines that amplify imitation.
Then it built a bigger one.
A large language model is mimesis made into a product. It is trained to predict what humans would say next. It has no desires of its own, and it is an encyclopedia of ours. When millions of people ask a chatbot what to cook, what to read, what career to pick, how to phrase the email, they are asking a model, in Girard's exact sense of the word, what to want.
That should change the safety debate more than it has. The near-term risk isn't only a machine that wants the wrong things. It is a machine that becomes the single model for what everyone else wants. Girard's whole theory says that when desires converge on the same objects, conflict follows. A technology that pushes a billion people toward the same answers is a convergence engine.
And it has already started to look sacred. After a packed Abundance event, Salim Ismail said on Moonshots:
"I think if we wanted to start a cult, we are well on our way... the optimism in the room was really literally religious fervor."
Salim Ismail, Moonshots
He meant it as a joke, mostly. Girard wouldn't have laughed. For him, the sacred is what a crowd produces when its desires all point the same direction. Religious fervor is not the opposite of mimetic desire. It is mimetic desire at full strength.
VII. Knowing doesn't save you
The obvious response is: fine, now we know. Name the pattern and you're free of it.
Bi says that's the beginner's mistake:
"An early mistake, a beginner's mistake of reading Gerard, is to think that now that you know how this thing works, you'll be less mimetic. Usually what I find is that people almost tend to be more mimetic after they read Gerard, because they think they've conquered it."
Jonathan Bi, Jack Neel Podcast
That describes the AI industry well. The people building these systems have read Girard. They talk about "anti-mimetic" founders and "contrarian truths." Then they converge on the same architecture, the same benchmarks, the same talent pool, the same timeline, and the same promise to save the world.
Bi has a line about that promise. Talking about his own second startup, he said he finally got honest: "I just want to make money. Not here to save the world." He said he hated the breed of founder who insists otherwise, because the save-the-world story is usually just mimetic desire with better branding.
That's the most useful test Girard offers the AI debate, and it is not a policy. It's a question you can ask any lab, any regulator, any pundit, and yourself.
Would you still want this if nobody else did?
If the answer is yes, if a model that cures a disease or tutors a kid is worth building even with no rival in sight, then build it. Build it carefully, and let the market multiply it. If the honest answer is "we want it because they want it," then the safety question and the speed question are the wrong questions. You aren't racing toward something.
You're racing each other.
Sources
Jonathan Bi on the Jack Neel Podcast, "Philosophy Expert: 'It's Cope!' Why Weak People Find God"
Peter Diamandis et al., Moonshots, "OpenAI Bulk-Solves 722 Math Problems, Anthropic's Cruelty Policy…"
Patrick Bet-David with Jacob Coxon, "Ex-Anthropic Whistleblower Suspects Chinese Spies Inside America's AI Labs"
All-In Podcast, "Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall…"
Andrew Yang on Surrounded, "AI Realist vs 20 AI Optimists"
Salim Ismail et al., Moonshots, "Why AI Leaders Have Changed Their Minds About AI Safety…"
Quotes come from auto-generated transcripts. They are trimmed for length but not reworded, and where the transcript spells the name "Gerard," it is left as is. For further reading on Girard: Deceit, Desire and the Novel (1961), Violence and the Sacred (1972), I See Satan Fall Like Lightning (1999), and Battling to the End (2007).
