AI Is Going to Kill Us All. Watch the Checkbook.

A new American swamp got built in public this week. Patterns, spy craft, human behavior. Don't trust the personality pulling the alarm. Trust the verification. Here's how.

"AI is going to kill us all." That was the line. A hundred million people spent a week inside it. This is not a piece about whether the machine will kill you. This is a piece about who built the frame you just lived in — and how you verify the next one, because the tools for manufacturing a reality are changing faster than the old arguments can track.

The resignation

Tuesday night, September 8. A 27-year-old named Jacob Coxon posted that he had just resigned from Anthropic. Three years of pretraining research, he said, split between OpenAI and Anthropic. Neither company acting responsibly. Both racing straight to self-improving superintelligence and gambling with our lives. Then the line that traveled: the people building AI earnestly believe it could kill us all by the end of the decade.

Within two days: past 100 million views. One breakdown found that about nine in ten people never read past the first post. The headline harvested. The argument did not have to.

Stay with me on the calendar. This is where people get sloppy, and sloppy is how you get tricked.

The timeline and the players matter, so start at the beginning.

2022. Four years before this post. Jacob Coxon — the man who wrote the resignation — received about $20,000 from the Long-Term Future Fund, money that traces to Good Ventures, Dustin Moskovitz and Cari Tuna's foundation. Training-era money. Farm team. Not a Tuesday paycheck.

Then he spent most of the next stretch at OpenAI, doing the unglamorous work of actually making the base models. That is most of the "three years" in the thread.

Anthropic spent about two years recruiting him. He joined in May 2026. Anthropic equity starts vesting at six months. He resigned September 8, at about four months, and told Axios he left before any of that stock vested. He still has OpenAI equity. He does not have Anthropic equity from this stint.

Write that on a napkin before you write a hero or a villain. The 2022 check is how this network grows talent years before anyone is famous. The four months is not a payday from Anthropic. Confusing them is how you get tricked.

On the surface it looks like his morals outran his desire for money. Look one layer down. He walked away from Anthropic stock that had not vested. He still holds OpenAI equity. The alarm hits the lab he just left. The stake he kept is in the other one. He wins either way.

What happened next did not require a smoke-filled room. Anthropic's firefighter — Alignment Science lead Evan Hubinger — quote-posted the thread and agreed. "Jacob is correct here — we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade," he wrote. "I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."

Coxon has said that after he posted, he asked a group chat of about ten people to share it, including the founder of Encode. That is human behavior with a tell: intent to propagate a message. Among the earliest quote-posts: Encode, the AI Policy Network, the AI Futures Project. The people who pushed it now get to help "solve" a problem a hundred million people just became aware of — because they helped make them aware. Ask the next question. Where does their money come from? Same funding weather as the grant engine below. The company line has been that it has always been transparent about unprecedented risk.

Read it slowly. A man leaves the factory and pulls the fire alarm. The factory's own firefighter — the alignment lead — says yes, the building could burn. The factory reminds everyone it has always said that. A hundred million nervous systems spend a week inside the story.

Nobody had to lie. That is what makes it worth understanding. Sincere people. A system that already knew how to use sincerity.

The old story about power, and why this isn't it

There is an old story about power. Smoke-filled rooms. Bloodlines. Intelligence agencies. Parents who already owned the map. This is not that story. These people grew up in Ocala, in Lincolnshire, Illinois, and in a Tuscan leather workshop. Nobody inherited a deep state.

This is a newer construction method. A fortune gets minted in a platform company. It gets parked in a foundation that talks like an accountant, not a party. The foundation buys a seat at the companies building the future. With the same checkbook it funds the people paid to warn us about that future. Idealists do the work. The public supplies the energy. The network writes the rules.

Call it a swamp if you want. American-style. Not a gang with a name. A seating chart: 501(c)(3)s, narratives, fellowships, a marriage certificate. Combined, they produce a different frame depending on the room. To a reporter it is philanthropy. To a founder it is an investment. To Congress it is expertise. To the public it is a fire alarm. One organism. Four costumes.

The seating chart

You do not need a conspiracy board. You need faces and a seating chart. Both are public. The Soros lesson took too long: name them. Show them. Stop letting an abstraction do the crime.

Coefficient Giving — Open Philanthropy until November 2025 — is the grant engine. Founding partner: Good Ventures, Moskovitz and Tuna. More than four billion dollars directed since 2014. Global health, animal welfare, biosecurity, science, and "potential risks from advanced AI." Their own numbers on the AI slice: $168 million in 2024, $351 million in 2025, more than a billion on track for 2026. August alone: $352 million. That is not a side project. That is a parallel civil service.

Now the corporate seats. 2017: Open Philanthropy grants OpenAI $30 million. Karnofsky takes an OpenAI board seat to oversee safety. He holds it until 2021. 2021: Moskovitz writes a check into Anthropic's $124 million Series A, next to Tallinn and Eric Schmidt, and says any returns go back into philanthropy. Same thesis, he said. Nonprofit and investment. Same place.

Karnofsky married Daniela Amodei in 2017. She is Anthropic's president. Dario, her brother, is CEO. Karnofsky and Dario were roommates first. January 2025: Karnofsky walks into Anthropic to work on responsible scaling. The man who held a safety board seat at lab one, funded by the fortune that also bought lab two, now works at the lab his wife runs.

Be precise about the method, because it evolved. OpenAI existed for two years before the checkbook arrived with a grant and a board seat. That is infiltration in the plain sense: enter something that already exists and take a chair. Anthropic is different. The money was in the Series A. The family tie predates the company. The safety overseer from the first lab eventually used the front door of the second. You cannot infiltrate a house you helped pour the foundation for.

Infiltrate the first lab. Build the second so you never have to infiltrate again. Infiltration was the beta. Construction is the product.

The OpenAI conversion, and the man who walked

You also have to look at the clock on OpenAI itself.

December 2015: OpenAI is founded as a nonprofit. Elon Musk is a co-founder. March 2017: Open Philanthropy's $30 million arrives and Karnofsky takes the board seat. February 2018: Musk leaves the board. March 2019: OpenAI announces OpenAI LP, the "capped-profit" hybrid. Karnofsky is still listed on the nonprofit board that sits on top of it. 2021: the Amodeis leave and found Anthropic, with Moskovitz in the Series A. 2023: Musk founds xAI.

I am not going to pretend Musk was a monk. Court filings later showed he was pushing a for-profit structure in 2017 too — with himself in control. They said no. He left. They built the capped-profit anyway, without him, with the Open Phil seat still in the room. Then the PR about Musk got feral, and it never really stopped.

Do these people "own the platforms"? Not all of them. They fund the fellowships inside the newsrooms, the explainer studios, the policy shops, the first fifteen minutes of amplification. That is enough to shape what a hundred million people think they just witnessed. The only real antidote to a closed information weather system is a true competitor. Not two logos with the same humidity. OpenAI and Anthropic are the illusion of choice. You already know this pattern. Coke and Pepsi. McDonald's and Burger King. A two-party system. Different labels, same shelf, same weather. It looks like a choice because it is built to look like one.

One third door looks like a personality. Several start to look like a market. Right now the public only has one working pattern they can actually see in AI. Without more than one living anecdote — a lab, a company, a structure a regular person can read — they cannot tell the swamp from the alternative. That is not a branding problem. That is how capture lasts.

Elon Musk is the Bernie Sanders and the Donald Trump of the AI wars.

I know how that sounds if you already hate him. Sit with it. The people who love Bernie often hate Elon, which is the joke the machine is in on. He is the messy, ungovernable third option — too populist for the salon, too builder for the purity test, too loud for the people who prefer to govern through fellowships. Do not take that as an instruction to trust him. Take it as an instruction to notice what a third door looks like, and then go verify him too.

The pattern, stripped of names

Every durable influence system has four layers.

1. The surplus. An outsized pile of capital from a platform moment. Facebook equity, later compounded by AI-lab equity.
2. The moral machine. The pile turns into virtue. Not "we are rich." "We are the adults measuring neglected catastrophic risk." It recruits people who would never take a lobbying job.
3. The critic class. Same capital, now funding safety labs, DC shops, journalism fellowships, explainer studios, scholarships. The critics are often sincere. That is the point. Sincerity is cheaper than a press secretary.
4. The harvest. Attention, talent, dread, consent. The public hears: the builders think this could kill us all. The network hears: someone must govern this, and we are already staffed.

Watch only layer three and you see concerned experts. Watch all four and you see a closed loop.

Ideology does the labor

This is the part people miss when they go hunting for villains. The system does not need a room of cynics. It needs a room of people who believe they are preventing extinction.

Effective altruism gave them a story that makes ordinary ambition feel like a moral defect. If the expected value of a catastrophe is large enough, then the think-tank salary is not careerism. It is triage. Someone has to allocate the last dollar. They have quietly elected themselves.

Well-intentioned people then do the work the old swamp hired lobbyists for. They staff the safety orgs. They write the evals that become regulation. They take the DC jobs. They train the next cohort. When one of them walks out of the factory and says the builders are privately terrified, nobody has to script it. The ideology already did.

After FTX, a lot of these shops distanced themselves from the EA brand. The grant pattern did not retire. It rebranded. A coefficient multiplies whatever you pair it with. Pair it with fear, you scale fear. Pair it with a preferred lab's safety story, you scale that lab's claim to be the adult in the room.

Fund the critic class with the same checkbook

Watch the media layer. That is where the harvest becomes visible.

The Tarbell Center for AI Journalism puts fellows inside real newsrooms — Bloomberg, the Guardian, MIT Technology Review, TIME — with stipends and reporting grants. Tarbell's own funders page lists Coefficient in the million-dollar-plus tier for 2023 through 2026, next to Tallinn's fund and Longview. They insist on editorial independence. Read that twice. Independence of assignment is not independence of which stories exist in volume.

You do not need to tell a reporter what to write if you already paid the fellowship, funded the subject, funded the experts in the piece, funded the videos that primed the audience, and funded the advocacy groups that amplify it in the first fifteen minutes. You have probably watched some of this without knowing who paid for it. Kurzgesagt is the slick animated YouTube channel with millions of subscribers. Rational Animations does the same job for AI risk. Asterisk is the magazine. Coefficient and its orbit put millions of dollars into that kind of "education." The public gets a video. The system gets distribution. Scholarships are the farm team. This is not a hospital check. This is a talent pipeline.

POLITICO mapped the Washington layer in 2023. Printed it. The system kept growing. So when a resignation thread lands, and the author has said he asked Encode's founder to help share it, and the first quote-posts are Encode / AI Policy Network / AI Futures Project, that is not a control room. It is selection. A decade of field-building makes it likely the person with the microphone already lives inside the funder's moral operating system.

Meanwhile they harvested our energy

Attention is a crop. Catastrophe is the highest-yield story on earth. Plant it in fellowships. Water it with grants. Harvest it when a thread does nine figures and every staffer has already been briefed by a grantee.

Talent is a crop. The 25-year-olds who might have built messy, competitive, uncontrolled things get meaning, salary, and a community that treats dissent as insufficient seriousness about death. The network does not have to silence opponents. It outbids them for the people who would have become opponents.

Dread is a crop. Believe the machine might end the decade and you will accept licensing, compute thresholds, "responsible scaling" written by the firms that can afford compliance. Fear is not a side effect. Fear is the permission structure.

Consent is a crop. Once the public has donated its nervous system, the network can say: we are not grabbing power, we are answering a demand. The demand was cultivated.

The well-intentioned people are not the harvest. They are the combine.

The inversion test

Want to know if you are looking at a system or a coincidence? Invert it. A single oil fortune funds the drilling company, sits on the board, finances the journalists who cover drilling, pays the experts in the pieces, endows the energy-beat fellowships, grants the think tanks writing the safety rules, then hires the think-tank founder into the company. The public is told this is how we prevent explosions.

You would not call that the marketplace of ideas. You would call it vertical integration of reality.

You can believe AI is dangerous and still see the design. The design depends on that belief being at least partly true. The most efficient propaganda does not invent the risk. It monopolizes the response.

Don't trust the face. Trust this.

The tools for building a frame used to look like theater. People argued about crisis actors — planted performers, a set, fake tears. Debatable. Messy. Easy to smear. And a dead end, because the fight became about whether you were crazy instead of whether the stage was built.

Forget the word. Watch the upgrade. You don't need a performer if you can fund the entire stage: the expert, the fellowship, the explainer video, the resignation, the firefighter inside the factory who agrees in public. A hundred million people walk through the set and call it news. That is not a conspiracy in a basement. That is a seating chart with a grant report.

Knowing who to trust is going to get harder. Personalities will be more convincing, more sincere, more viral. Sincerity is cheap. The only thing that does not get cheaper is verification.

If you are a civilian trying to see straight, or you are an entrepreneur using any of these same tools — an operating company, a nonprofit, a holding company, a public alarm — this is the process. Use it on them. Use it on me. Use it on whoever you were about to believe.

1. Who paid, and how long ago? A check last week is a handler story. A scholarship four years ago is a farm team. Don't mush them. Write the calendar on a napkin before you write the villain.
2. Can you name a human beneficiary? A mother, a kid, a clinic, a person who cashes a grant. If the page says "extinction" or "cost-effectiveness" and nobody in a waiting room gets paid, treat the foundation as a political machine until proven otherwise.
3. One household, two costumes? Capitalist and philanthropist. Grantmaker and lab president. Same kitchen table. There is no wall. There is a marriage certificate.
4. Invert it. If an oil fortune did this, would you still call it the marketplace of ideas?
5. What happens to them if you believe them? Policy window, valuation, licensing regime, adult-in-the-room premium. If believing them pays them, believe slower.
6. Read the payment order. Who gets paid first when the thing grows? If you need a lawyer to find the human, the structure is the con.

Do not outsource this to a better personality. One third door looks like a savior. Several readable structures start to look like a market. The public cannot tell the swamp from the alternative if there is only one working anecdote. Build more. Make them readable on purpose.

The design flaw, and what it looks like fixed

Here is the flaw. One checkbook holds the factory, the fire alarm, and the harvest, and everyone inside the loop believes they are the exception. Someone has to allocate the last dollar. They volunteered.

Absolute power corrupts absolutely. The useful version of that sentence is not a warning about other people. It is an instruction about yourself.

I did not build Mom Bomb assuming I was the exception. I built it assuming I wasn't — which is the only reason this section exists. Verify me with the list above. Sunshine is the best disinfectant, which is why transparency is the moat of what I am building. We welcome it. It proves we are real and truthful. It is also the only way a structure like this does not become another personality cult.

I have a hard time with this organization becoming about me. I am the front person. The idea started with me. The thing is about all of us. I had to use a personality to push a readable structure through a world that is not engineered for one, then hold that personality accountable in public, and eat the cost of building it without the swamp's resources. That is not a sad song. That is the price of not becoming the thing you are describing.

The order of payment was fixed before a dollar of outside money came in, and it was fixed in the one direction I could never reverse. Every residual dollar of profit this company makes leaves the building, permanently, to fund grants for mothers in an actual present-tense emergency. Not on a bad year. Not if an investor asks nicely. Not if I change my mind at sixty.

Let me be truthful about the rest, because a piece about hidden checkbooks should not have one. Mom Bomb sits inside a holding company that incubates new brands. That holding company can pay its shareholders, as a percentage of gross sales, and only when two things are true at the same time:

1. the operating company is healthy, and
2. the nonprofit it funds is healthy.

If either of the first two fails, the third does not happen. Mission first. The people who built this unpaid, for years, after that.

Nobody arrived on a fellowship. They self-selected by working.

That is why this is taking longer than a normal term sheet, and why it will outlive me. When I raise an alarm, there is no fellowship I'm funding to say it louder, and no equity that jumps if you're frightened. The only thing that grows first is the ability to keep funding grants. Mothers get paid before any owner does. I still hold a lot of power. I run the company. I choose the investors. I like power as much as the next founder. The constraint exists because I did not trust myself with all of it. Not because I'm above wanting it. Because I wanted this to live beyond me, and beyond everybody in the building. I intended it to take care of people who did not have anybody to take care of them — people who need to reset their nervous systems — and I want to help in any way I can, because that is the reason I was able to create this. I used the same kinds of tools they did. I just refused to make them hard to read. If you need a diagram to find the human being who gets paid, the structure is the con.

I wrote about this last fall, in Where Your Money Comes From Matters: money holds the imprint of its maker. If it was built on depletion, it keeps depleting. Same is true of the structure money sits inside. Build it so the mission gets paid before anyone who could be corrupted by it, and it stops mattering how good the people in it are. I am the factory, and I wired the alarm to ring somewhere else.

The swamp is not hiding. It publishes grant reports. It rebrands in press releases. It posts the seating chart and calls it transparency. So watch the coefficient. Notice what it multiplies. Don't trust the personality. Trust the verification. Then ask the only question that matters about anyone asking to govern your future: what happens to them if you believe them?