The Doomsday Machine

THE NUMBER: 10%. That is the floor that Anthropic’s own alignment-science lead, Evan Hubinger, put this week on the odds that AI kills every human being on earth within the next decade. Not a critic. Not a doomer with a Substack. The man whose actual job is to find the ways his company’s safety work fails before a shipped model does. He said it out loud, on the record, the same week his company is being talked about as a two-trillion-dollar IPO. Ten percent. Put it the way Jim Cramer put it the next morning: if a surgeon told you an operation carried a 10% chance of killing you, you would not have the surgery. Ten is twice the five that already makes you walk out of the office. That is the number. Now watch what the people who know the most do with it.

Cramer flipped in twelve hours

Tuesday night on Mad Money, Cramer told his audience the Anthropic IPO could be so big that investors would sell everything else they owned just to raise cash and get a piece of it. Wednesday morning on Squawk on the Street, he was doing surgery math and backing away. Nothing about the company changed in those twelve hours. One number changed the conversation.

Here is what dropped in between. On Monday, a researcher named Jacob Coxon quit Anthropic and posted why. Three years of pretraining research, first at OpenAI, then at Anthropic, the two most important labs in the world, and his parting words were that neither of them is acting responsibly, that they are racing straight at self-improving superintelligence, and that they are, his phrase, “gambling with our lives.” The post has done something like 124 million views. It is, by a wide margin, the most-seen thing in AI this week, more than the model launches, more than the math proof I’ll get to. Then on Tuesday, Hubinger, his boss’s boss on the safety side, backed him publicly and added the 10% number, along with the detail that should have been the headline: Anthropic has no plan to solve alignment for superintelligence and is not on track to get one.

So the market is now trying to price a number that the only people qualified to set it will not confirm and will not deny. That is a strange and important situation, and it is worth slowing down to read it correctly, because almost everyone is reading it wrong.

Everyone hit their marks

The first thing that happened is that every actor in this drama played exactly the part their incentives wrote for them. You could have scripted it a week in advance.

The politician called for control. Representative Ro Khanna put out a five-point plan within a day: a new federal agency modeled on the ones we have for nuclear power and aviation, mandatory pre-certification with kill switches and human permission before a model can rewrite itself, liability and compulsory insurance for agents let loose on the open internet, criminal penalties for shipping an uncertified model, and oversight hearings with whistleblower protection. He also said we urgently need an agreement with China so the race doesn’t become “devastating for humanity.” That is a serious, thoughtful list. It is also exactly what a serious, thoughtful congressman is for, and it tells you nothing you didn’t already know.

X did what X does. Within hours the theory hardened that the whole thing, the resignation, the number, the WSJ interview, was a coordinated influence operation, a well-funded push to hand one political party a reason to regulate AI into the ground. Jason Calacanis, who is nobody’s idea of a naif, asked Dario Amodei point blank whether Anthropic was running “a regulatory capture playbook.” The replies filled in the cynical version: step one, become fantastically wealthy; step two, regulate your future competition out of existence. Others fought back that the researchers are credible and the warning is real. Nobody resolved it, because on X nothing resolves.

And Washington will hold a hearing and do nothing. Not because Washington is stupid, but because Washington can do the game theory. You do not disarm unilaterally in a two-player race, and this same week the NSA, CISA and the FBI put out a joint warning that six China-based firms are aggressively distilling American frontier models, while Semafor reported that the United States and China cannot even agree on what the words “AI safety” mean. Khanna’s call for a deal with China is dead on arrival the same day he makes it, and everyone in that War Room knows it.

All of that is theater. I don’t mean it’s fake. I mean it’s predictable, and predictable behavior carries no information. The signal is somewhere else.

The tell is what you can’t see

The signal is that Dario Amodei said nothing.

Sit with that, because it’s the whole issue. This is the man who wrote “Machines of Loving Grace.” The one who has spent years as the designated grown-up of the industry, the CEO most willing to go to Washington and describe, in vivid detail, exactly how his own product could go wrong, the one who asked to be regulated. His entire brand, and a good chunk of his company’s two-trillion-dollar story, is built on being the careful one. And in the week his own alignment lead told the world there’s a better-than-10% chance this kills everybody, Dario did not issue a reassurance, did not offer a counter-number, did not push back. “Anthropic did not comment,” is what the Journal reported. There is even a piece going around this week titled simply “Dario Must Speak.” He hasn’t.

Now, the leaker talked, so let’s be honest about the leaker, because this is where I part company with the easy read. A pretraining researcher who just quit either has very little vested equity, in which case an IPO windfall is not a thing he’s protecting and talking costs him nothing, or he is, as the X crowd insists, some kind of plant. Either way, his candor is cheap. That doesn’t make him wrong. It makes him uninformative in the specific way that matters to a market: he has some of the information and none of the exposure. He is not in the room for everything, these are not big companies but they’re not that small either, and he has no billion-dollar reason to weigh his words.

The two men who have all the information and all the exposure are Dario Amodei and Sam Altman. And here is the asymmetry that is the actual story: the people who know the most and have the most to lose have either gone silent or gone vague, while the only plain talk is coming from the people with nothing to protect. Altman at least said something, but it was the cheap version. “A good time to slow down,” he told Time, in the same stretch of days that OpenAI spent roughly ten thousand agents and millions of dollars in tokens to crack a ninety-year-old math problem and called it “one of the most amazing moments in OpenAI history.” A columnist read his slowdown talk as cover to make OpenAI’s numbers, an 18% revenue quarter against a $12.3 billion operating loss, look more palatable ahead of the confidential S-1 it has already filed for a 2027 debut. Maybe. But even taken at face value, “slow down” from the man flooring the accelerator is a hedge, not a disclosure.

In markets, you learn to read the expensive signal, not the cheap one. Anyone can talk their book. The move that costs the speaker something, that cuts against his own interest, is the one you weight. By that rule, Dario’s silence is the most expensive sentence in the industry right now, and that is exactly why it’s the truest thing anyone has said all week. A man with a comment ready and two trillion reasons to give it, choosing to swallow it, tells you more than a junior with nothing to lose shouting into a megaphone.

Nobody can pull the lever

Why can’t they just stop? Coxon answered that too, and it’s the most important sentence in his whole thread. At OpenAI, he said, many people haven’t internalized the stakes. At Anthropic, they understand the stakes completely, and they keep building anyway because they believe that if they stop, someone less careful wins the race. Read that again. That is not recklessness. That is a prisoner’s dilemma, confessed by a man who was inside it. And a prisoner’s dilemma has a dominant strategy, and the dominant strategy is to defect. Everyone would prefer that everyone else slow down. No one can afford to move first.

A good essay went around this week called “Can Artificial Intelligence Be Paused,” and it makes the structural case cold. A pause is desirable and probably impossible, because “humanity” is not a political actor. There is no humanity that can decide anything. There are eight billion people organized into rival states and companies and labs, each with its own reason to keep going and its own fear of being the one who stopped. The nuclear-arms-control analogy that everyone reaches for actually breaks in the most important place: you can confiscate enriched uranium, but you cannot confiscate an algorithm. Nuclear weapons need rare materials and conspicuous factories. AI, in the end, is math, and math has never once submitted to an international ban. The physical chokepoints that make control seem possible, the giant clusters and the fabs, get weaker every year as the models get cheaper and the knowledge spreads. Stanford’s 2026 index says capability is still accelerating, not leveling off, and that the gap between American and Chinese models has nearly closed. There is no seat at the table with the authority to call the machine back, because there is no table.

The machine

Which is the whole point of the frame. The Doomsday Machine in Strangelove is terrifying for one specific reason, and it is not that the Russians are evil or that the American general is insane. It is that the device was built to remove human control on purpose, because a deterrent that a human can call off isn’t a deterrent. Once it exists, it runs on its own logic, and every rational actor in the film behaves rationally, and the world ends anyway. That is the part that maps. We are not watching a story about bad people who want to end the world. We are watching a structure, build or be beaten, that produces the outcome no individual inside it would choose, and that no individual inside it can stop. Intent has been amputated from the process. You can watch it happening in real time this week, in the gap between what these men say privately and what they’ll say with their name on it.

And if you want to see the amputation in a single artifact, look at what Anthropic actually shipped the same week. Not just the 10% remark, but a polished interactive tool, the Econ Scenario Explorer, that models how AI reshapes the American economy by 2030. Its own worst case has cognitive-worker unemployment at 17.9%, the labor share of income falling from 60% to 45%, and a required government transfer the size of Social Security and Medicare combined just to keep white-collar workers whole. And that model, by its own admission, leaves out catastrophic risk entirely. So here is the same company, in the same week, publishing an elaborate calculator for how the machine wrecks your economy, and telling you plainly that it does not model the part where the machine kills you. The doom lab built a doom model and left the doom out. That is not a contradiction. That is the amputation, drawn to scale.

The honest part

Let me argue against myself for a paragraph, because a piece like this curdles into a doomsday sermon if you let it, and I don’t believe the sermon.

A pause has a cost of its own, and it isn’t small. The same acceleration that produced the 10% number also produced, this exact week, Google DeepMind (NASDAQ: GOOGL) mapping all nine billion possible mutations in human DNA and a research team using it to crack a rare-disease case that used to take years. The thing that can’t be stopped is also the thing curing the disease that runs in your family. Slowing it down is not free, and anyone who tells you the risk is all on one side is selling you something.

And the cynical read, the “it’s just regulatory capture” read, is half right and half wrong in a useful way. It is right that Anthropic would benefit from being regulated into a protected class. It is wrong about the timing, and the timing is the tell. Capture is a game you play after you’re public, when the moat is worth more than the growth. You do not, if you’re running the capture play, let your alignment lead put a 10% extinction number in the Wall Street Journal during the quiet period before the biggest listing in history. No CFO alive allows that if he can stop it. Which loops us right back to the only fact that survives every reading: the thing that leaked is the thing the strategy couldn’t suppress, and the men who could suppress it have gone quiet.

So the answer isn’t “they’re lying” and it isn’t “they’re reckless.” It’s worse and simpler. Everyone is rational. Everyone is in character. And the machine runs anyway.

The closing bell

Here is the part that turns this from a shiver into a trade, and it’s the reason I wanted to walk into this room at all.

The candor you are hearing this week has a shelf life, and the expiration date is the IPO. Right now, still private, these companies will let a researcher walk out the front door and say the true thing. The day they are public and liquid, that 10% becomes a single line in a risk-factors section written by outside counsel, the alignment leads stop giving interviews, and the most honest information the market will ever get about this technology goes dark. You are listening to the last of it. The quiet you can already feel settling over the principals is not the danger passing. It is the soundproofing going up.

So do three things, and none of them require you to believe in the apocalypse. First, go through your own plan and find every place it quietly assumes a brake, the trade that only works if regulation caps the labs, the forecast that pencils in a US-China deal, the strategy that counts on the labs policing themselves. Cut it or reprice it, because the men inside the race just told you that brake is not coming. Second, harvest the candor now. Pull the actual words, not the headlines, into whatever your organization uses to think about AI risk, while the words still exist. Third, when the talk goes quiet, when the doom chatter suddenly gets disciplined and on-message, don’t read it as reassurance. Read it as the S-1 getting close.

We have been telling you all year to get off the paper that’s priced for a soft landing and own the things that compound whether or not anyone’s in control. This is the same call, arriving through a different door. The soft landing has a name now. It’s the pause, and the people building the plane just told you, by what they said and by what they wouldn’t, that it isn’t coming.

The loudest thing in the room is the silence of the man who used to warn you.

The full Signal/Noise, with the theater sorted from the signal and the three moves spelled out, is what you just read. The email edition is in your inbox. Analysis, not investment advice, as always.