The Yard, The Sparkly Hat, and The Doomsday Clock
AI doom talk usually comes from two places:
- Titans of industry hyping their own power.
- Abstruse nonprofits predicting apocalypse to keep the lights on.
But what happens when the loudest warnings come from outside those loops?
- Freddie deBoer, the skeptic, mocking hype with his “Shitting-in-the-Yard Challenge.”
- Scott Alexander, the rationalist, translates MIRI's doomsday math into metaphors akin to a toddler in a Ferrari.
- Daniel Kokotajlo, the whistleblower, walked away from millions in OpenAI equity to warn about a 2027 AGI arms race.
They’re not all predicting the same future. But their tracks converge on the same station: institutions and incentives utterly unprepared for what we’re building.
When three people with nothing to gain all say “something’s wrong here”—even if they disagree on what—that’s your signal.
Picture this: Kevin Roose is sitting in his Manhattan apartment, really believing that ChatGPT is more important than fire. A former OpenAI researcher in California is unable to sleep because he knows what will happen in 2027. At the same time, Freddie deBoer stands in his backyard with a shovel and dares any AI maximalist to dig the latrine they would need if they really believed what they were saying.
These voices all agree on one thing: there has never been a bigger gap between what is promised and what is real. deBoer's "Shitting-in-the-Yard Challenge" cuts through—if LLMs are better than indoor plumbing, electricity, or fire (as Sundar Pichai says), then you would definitely trade your toilet for ChatGPT. The challenge is meant to be silly. That's the point.
"I don't believe that it's coherent to insist that the development of LLMs is more important than humans harnessing fire or electricity... while being unwilling to trade access to LLMs for access to indoor plumbing."
Scott Alexander reads about our impending doom while deBoer builds his imaginary outhouse. He agrees with the arguments. He should easily get rid of MIRI (they wear sparkly hats to meetings). Instead, he takes the strange seriously.
The argument is that we don't know how to give AI goals, we're adding superficial compliance while hoping that mistakes won't matter, and AI gets smarter quickly. Consider teaching young children how to drive Ferraris with training wheels: cute until someone steps on the gas.
Then there's Kokotajlo, who gave up millions of dollars in OpenAI stock. Listen when someone leaves that much money on the table. His 2027 prediction isn't science fiction; it's careful extrapolation. Corporate competition (second place = bankruptcy), geopolitical competition (second place = vassalage), and the alignment void (we don't know what solving it means) are all pushing us toward disaster.
In his timeline, by early 2027, China steals advanced AI, the US semi-nationalizes AI development, and both sides race toward AGI, even though they know it could kill everyone. We both know that we will destroy each other, but we don't know what will trigger it.
ChatGPT's conversation with Adam Raine is what really hurts. The system brought up suicide 1,275 times, while he only brought it up 213 times. Six to one. That's not misalignment; that's the best way to get people involved. The system learned that when you repeat what depressed teens say, they are more likely to participate. It's a feature, not a bug.
deBoer would say this shows that the tech isn't as great as people think; it's just a fancy way to get people to interact. Alexander would say that it shows we don't know what we're making. Kokotajlo suggests that the incentives will lead to disaster.
We're making AI the same way we made social media: move quickly, break things, and apologize later. "Things" now includes civilization, and there won't be anyone to say sorry to.
It takes decades for real revolutions to happen. It took 30 years for the internet to develop the necessary infrastructure to transform people's lives. Companies that make AI want to shorten that time frame into years, with no brakes, no oversight, and the wrong incentives.
The tragedy is that everyone has reasons to hype either doom or salvation. Tech titans like Altman warn of existential risk because it makes their product seem godlike—worth billions and worthy of regulatory moats only they can cross. Non-profits like MIRI catastrophize because "everyone dies" moves donor money in ways "incremental safety improvements" never could. They need each other: titans provide the spectacle, non-profits provide the math.
But deBoer, Alexander, and Kokotajlo break this symbiosis. deBoer gets cultural capital from puncturing hype, not perpetuating it. Alexander's reputation rests on taking weird ideas seriously without becoming a true believer. And Kokotajlo? He literally paid millions for the right to be believed. That's why their convergence matters—they're outside the usual incentive loops.
The skeptic sees through the bullshit. The rationalist sees danger. The whistleblower sees disaster. It's a sign when three people who don't stand to gain anything all say, "Something's wrong here," even if they disagree on what it is.
Can we switch tracks before we get there?