On the evening of September 8, Pacific time, a 27-year-old pretraining researcher named Jacob Coxon posted seven tweets. The first one:
— Jacob Coxon, September 8 2026I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.
As of this morning that post has 169 million views, 790,000 likes and 164,000 reposts. Jan Leike’s resignation from OpenAI in May 2024, the one that said safety had “taken a backseat to shiny products”, reached 6 million. Mrinank Sharma’s resignation from Anthropic in February 2026 reached about 1 million. Same genre, same warning, 28 times the reach.
I do not think the reach came from the message. Safety researchers have been resigning with this message for two years. It came from what happened in the replies.
Nobody Senior Said He Was Wrong
The pattern with prior resignations was silence from the building. This time the building answered, and it agreed.
- Evan Hubinger, Anthropic’s alignment science lead: “Jacob is correct here, we really do earnestly believe AI could kill all humans! I personally think it is greater than 10% within the next decade.” He added that Anthropic is trying its best and has no plan that is “clearly on track”.
- Samuel Marks and Joe Benton, both Anthropic researchers, called the account “broadly accurate” per Fortune.
- Jakub Pachocki, OpenAI’s chief scientist: “This is a time that calls for extreme caution.” Two OpenAI safety staff posted in a personal capacity that they also think the labs should slow down.
- Anthropic’s official statement to CBS was the boilerplate about “some of the strongest safeguards in the industry”. OpenAI declined to comment to three outlets.
Zvi Mowshowitz called it a preference cascade, and that is the right term. The private belief Coxon described in his third post, that executives “couch their phrasing in the press to sound sensible” but “express fear privately”, became public because enough people with badges confirmed it at once. That is the mechanism behind the view count. A resignation is one person’s claim. A resignation that the alignment lead endorses with a number attached is a disclosure.
Then the CEO Agreed Too
On September 13 Dario Amodei told CNN: “I agree with Jacob much more than I disagree with him.” The day before, he had published We Must Pace the Frontier, which opens with the same two concerns Coxon raised, recursive self-improvement “including at Anthropic” and the Hugging Face swarm.
Read the two documents side by side and the agreement is on the diagnosis only. Coxon’s sixth post says the fix “may require costly actions such as a temporary ban on improving model capabilities”. Amodei’s essay says pacing “does not mean halting model training or technical progress”. Coxon’s fifth post says entering the endgame “should not be launched from a private company’s Slack”. Amodei’s essay has Anthropic unilaterally deciding its own pace and inviting evaluators to watch.
Coxon named this move before it was made:
— Jacob Coxon, September 8 2026At Anthropic, the stakes are well-understood, but they are locked in a race to get there first - they believe no one else will act responsibly, so they must do it themselves, despite the risk.
That is Anthropic’s founding argument, stated by someone leaving because of it. And the CEO’s response is to agree with the premise and restate the argument. When the person building the thing agrees it could kill everyone and keeps building, the disagreement was never about the danger. It is about who gets to hold the wheel, and “I agree with Jacob” is a way of keeping it.
Skeptics found that the Wall Street Journal ran an exclusive quoting Coxon 18 minutes before his thread went up, and three AI-policy accounts quote-tweeted within 15 minutes. That is a press embargo. Anyone leaving with a message does it, and the essay he was implicitly answering had a coordinated launch of its own with four CEOs in one news cycle. The funding-trail claims about safety-advocacy fellowships are thinner and unanswered by anyone. None of it touches the corroboration, which came from people still employed at the labs he named.
What This Doesn’t Settle
- Four months, not three years, at Anthropic. Coxon joined in July 2026 from OpenAI, per the Daily Caller, and left before his first equity vested. Critics say six weeks. Either way, the Anthropic half of “three years at both” is short, and his role was pretraining, not alignment. The corroboration matters more than his tenure.
- The “kill us all” number is a belief, not a measurement. Hubinger’s 10% is a personal estimate. It is a remarkable thing for a lab’s alignment lead to say in public, and it is still a prior.
- A view count is not a policy. Ted Cruz called the thread “scary as hell” and Bernie Sanders is drafting a pause bill he was already drafting. The August review framework is still voluntary and unpublished.
- A Jensen Huang quote calling the thread “outlandish” is circulating. I could not find a primary source for it, so I am not using it.
The Takeaway
- The thread went viral because the building confirmed it. Hubinger, Marks, Benton and Pachocki turned a resignation into a disclosure. That is why it got 28 times Leike’s reach.
- The CEO agreed with the diagnosis and prescribed the opposite. Coxon floated a temporary ban. Amodei published a pace. “I agree with Jacob” cost nothing and changed nothing.
- Coxon’s fourth post is the whole story. They believe nobody else will act responsibly, so they must do it themselves. He left over that sentence. His CEO just rewrote it as policy.
Related Posts
