It was a battle of wits, and clearly 47f came unarmed.

In May 2026, an anonymous account on X paid Eliezer Yudkowsky $10,000 to debate him. The account calls itself “47f.” Its handle is a random string of letters and numbers. It claimed to belong to “a Director and Founder of an AI lab in stealth mode.”

Stealth mode means a startup is working in secret and has not announced itself yet.

Yudkowsky is one of the best-known voices warning that AI could wipe out humanity. His 2025 book with Nate Soares is called If Anyone Builds It, Everyone Dies. The debate aired on Liron Shapira’s YouTube channel on May 5. It runs just under an hour.

I watched it. Then I pulled every one of 47f’s posts that the Internet Archive saved, which is 21,226 of them. That mattered, because his profile now shows only about ten. This post is about what the debate was, and what 47f’s own words say it was.

The opening: “I forfeit”

47f did not open with an argument. Less than two minutes into his opening statement, he quit.

“The point is I lose, okay? I forfeit. I lose. The debate has concluded. Yud the stud has won.”

Then he turned the payment into a threat. Because Yudkowsky took his money, he said, they were now “counterparties to a financial transaction.” So any future criticism of the AI industry could be “potentially actionable trade libel.”

Trade libel is a false statement of fact that hurts a specific business’s products and causes it real money losses.

That legal theory does not hold up. Paying someone to show up for a debate does not buy their silence. Trade libel also protects a specific business from specific false facts, not a whole industry from opinions. Yudkowsky’s answer was short: “I hereby disparage the AI industry. It is worthy of being disparaged.”

47f went further. He called the $10,000 “a kind of protection payment” to “a group of people who are legitimately dangerous.” He said that if anything happened to him, he hoped Yudkowsky would be investigated.

The one real point, buried

There was a real argument in there, and it deserves a fair hearing.

In April 2026, someone threw a firebomb at OpenAI CEO Sam Altman’s home. Two days later, police said someone fired a gun at it. 47f’s claim was simple. When a famous person says “everyone dies,” some unstable listener might decide to act. He wanted Yudkowsky to choose his words more carefully. He was explicit that he did not want him silenced.

That is a fair question about speech and violence. Yudkowsky’s reply was also fair. A civilization cannot hand control of what gets said to the most unstable person who might hear it. If you believe something is true and deadly, you have a duty to say it. He also said, more than once, that no one should go after AI lab workers, and that the only fix he supports is international treaties.

The question deserved a better messenger. 47f buried it under insults (“annoying piece of,” “good boy,” “you’re a joke”), a legal threat that does not hold up, and the talk about protection payments.

The arguments he did make

When 47f did argue, he made three main claims.

“If anyone builds it, everyone dies” is a tautology. A tautology is a statement that is true by definition, like “all bachelors are unmarried.” 47f said the claim boils down to this: “if we create an AI that desires to kill us and has the ability to kill us, it’s going to kill us.” But that is not Yudkowsky’s argument. His argument is that we don’t know how to control what a very smart AI ends up wanting. Most goals it could end up with don’t include keeping humans around. You can disagree with that. It is not circular.

AI “desire” is just a metaphor. 47f compared an AI to a computer program with a random number generator wired to nuclear missiles. Fair enough: “wants” is loose language for software. But a system that steers the world toward some outcome is the danger, whatever word you use for it.

“We very much know what’s going on in an LLM.” An LLM, or large language model, is the kind of AI behind ChatGPT and Claude. 47f said the experts in his lab understand these models. That claim is the one his own posts take apart.

What his own posts say

The archive covers June 2024 to September 2026. Here is what it shows.

The lab was a month or two old. On September 1, he wrote that he had been working long hours “since ~March when I started my lab.” In March and early April he called it “my project” and “the startup I’m self funding.” The first time he called it a “lab” was April 18, eight days after the firebomb at Sam Altman’s home: “I am directing a lab of AI researchers.” Five days later he made the debate offer.

The money story kept moving. On April 2 he claimed “a funding round done.” In May he said VCs were “banging on my door.”

VCs, or venture capitalists, are investors who fund young companies.

On June 1 he wrote: “I still do not have VC funding.” He said he had a first client but never named one. The lab never posted a name, a paper, or a product that anyone could check. It promised to “publicly announce in October.”

He told us he can’t do the technical work. On May 12, in a long post he said he wrote while “VERY VERY stoned,” he wrote: “I suck at math. I cannot code.” The same day, he described researchers “in my lab (which totes does exist fr fr).” Five days later he called himself “a newcomer to this industry.” Five days after the debate, he was asking strangers online to help his lab get access to base models.

A base model is an AI model before the safety and chat training that companies add. The big AI labs mostly keep theirs private.

He planned the circus before it started. Before the debate, someone asked how he would win. He answered: “I won’t win because I’m not going to ‘debate’ him in the conventional sense. I’m going to do something worse.” When someone told him he was being rude to Yudkowsky for no reason, he said “the impoliteness is part of the point.” He said he was spending the money “because I think it’s funny.” He predicted “Yud will ragequit in 5 minutes,” meaning quit in anger. Yudkowsky stayed for the whole hour.

After it was over, the explanations piled up. First it was protection: “my people are safer now.” When people called the debate a mess, he said: “It was supposed to be a shitshow.” In July he said “my goal in the debate was to make Yudkowski look as ridiculous as possible.” Later that month it was a moral act: he had paid to “pull the fucking lever” on a runaway trolley. By August 31 he was taunting Yudkowsky’s co-author: “best 10k I ever fucking spent.”

Then, on September 11, he pinned this post to the top of his profile: “None of my tweets are worth reading.” Almost all of them are gone now.

Fair is fair

Yudkowsky was not perfect either. Asked to say no one should kill 47f, he added a careful exception about police and self-defense laws. That was accurate. It also sounded terrible. He interrupted more than once. At one point he told 47f, “These are deep matters. You may not understand.” And for the hardest step in his argument, the idea that most goals a superintelligence might end up with leave no room for us, he offered a lottery analogy more than a proof.

None of that changes the result. One person came to talk about AI risk. The other came to perform. He paid $10,000 for the stage, forfeited the debate in his opening statement, and spent the next four months explaining why that was the plan.

It was a battle of wits, and clearly 47f came unarmed.

The archive

The Internet Archive saved 21,226 of 47f’s posts before the account deleted them. I’ve collected them, along with a transcript of the debate, so you can check every quote in this post yourself.

Download the 47f post archive (7z, 3.1 MB)

Sources

Written on October 5, 2026