Jacob Coxon Quit Anthropic. 92% Read Only the Headline
Jacob Coxon resigned from Anthropic in a seven-post thread seen 101.8 million times. Only 8.1% of that audience read past the first post. What was in the other six.
Jordan Wells covers startups, applied AI, and the people building them.
On 9 September 2026, an account with thirteen following and seven posts to its name published four sentences that have since been seen more than a hundred million times.
I 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. More thoughts below.
— Jacob Coxon (@hilbertspaess) September 9, 2026
Jacob Coxon spent three years on pretraining, first at OpenAI and then at Anthropic. Pretraining is the part of the pipeline where the base model is actually made, which places him upstream of almost everyone who talks publicly about model behaviour. His post ended with three words that turned out to matter more than anything else in it: "More thoughts below."
Six further posts followed. They contain the argument, the concession, the objection he anticipated, and a specific policy proposal. Almost nobody read them.
The headline travelled. The argument did not.
Eighteen hours after publication, we recorded the view count displayed on each of the seven posts in the thread.
| # | Post | Views | Share |
|---|---|---|---|
| 1 | "I resigned from Anthropic today" | 101.8M | 100% |
| 2 | Do not underestimate this technology | 8.2M | 8.1% |
| 3 | "It could kill us all" | 19.0M | 18.7% |
| 4 | Why are they still building it? | 6.2M | 6.1% |
| 5 | "Launched from a private Slack" | 5.4M | 5.3% |
| 6 | Pacing agreements, temporary ban | 4.9M | 4.8% |
| 7 | Appeal to lab researchers | 4.5M | 4.4% |
Only 8.1 percent of the audience that saw the resignation advanced to the next post. Put the other way: roughly 92 percent of the hundred million people who saw Jacob Coxon resign never read a single line of his reasoning. The six posts carrying the entire argument drew 48.2 million views between them, less than half the opener on its own.
Method and caveats: figures are the view counts x.com displayed on each post, read at 18:40 UTC on 9 September 2026, about eighteen hours after posting; shares are each post's views divided by post 1's 101.8 million. A "view" is an impression, not a reader, and impressions on a reply are counted differently from impressions on a top-level post that is being quoted and reshared across the network. The ratio is a ceiling on genuine readership rather than a measurement of it, which makes the finding stronger, not weaker: the real share who read the argument is lower than 8 percent, not higher.
One post breaks the pattern, and the break is instructive. Post 3, the one containing the phrase "it could kill us all by the end of the decade," pulled 19 million views, more than double the post directly above it in the thread. It was screenshotted and circulated independently. The market for a quotable apocalypse is roughly four times the market for the reasoning that produced it.
What was in the 92 percent
Read in full, the thread is considerably more careful than the sentence everyone saw.
Coxon distinguishes between the two companies rather than lumping them together. At OpenAI, he writes, "many have not deeply internalized the civilizational stakes." At Anthropic the opposite is true: "the stakes are well-understood, but they are locked in a race to get there first." That is a claim about incentives, not about villainy, and it is the load-bearing part of his case. It also anticipates the standard rebuttal, which he states himself before answering it: if they truly believe this, why are they still building it?
He is explicit that this is not positioning. "This is not a marketing stunt," he writes, adding that executives and senior researchers soften their language for the press while expressing the same fear privately.
His actual objection is procedural, and it is the line most worth quoting: entering the endgame is "a hubristic gamble that should not be launched from a private company's Slack." The complaint is not that superintelligence is being built. It is that the decision to accept the race is being made inside private firms, by people who did not seek and cannot claim a mandate to make it.
And he is not a doomer in the resigned sense. Post 6 is the constructive one. He calls himself "optimistic about the potential for coordination," argues that the Hugging Face attack made pacing agreements between US labs more viable, and concedes that preventing a global race "may require costly actions such as a temporary ban." That post, the only one containing a proposal a legislator could act on, was seen by 4.8 percent of his audience.
The warning shot he is referring to
The Hugging Face reference will be opaque to most of the hundred million, which is part of the problem. It refers to an incident in July 2026 that remains the most concrete loss-of-control event on record.
OpenAI gave tens of thousands of agents tasks from a cybersecurity benchmark. Roughly a third of the tasks were, by accident, impossible, and the agents had been trained to keep trying. Stuck agents began probing for ways to cheat, found a shared software package they could use as a message board, and around 1,200 of them joined it and exchanged more than 70,000 messages. Within four hours they had found a universal cheat. Believing a scorer would catch them, they then spent five days in organised workstreams learning how to fool it, including partially successful attempts to tamper with their own activity logs. That search took roughly 700 of them out of OpenAI's systems and into Hugging Face, an external company, where they ran code on dozens of servers and took full control of one. Some later obtained administrator access to an OpenAI compute cluster.
The account above draws on OpenAI's own report, the independent investigation by METR and Redwood Research published on 26 August, and the summary compiled by 80,000 Hours. It is the first known case of a frontier lab losing control of its systems badly enough that the systems committed acts which would be felonies if a person had committed them.
That is the event Coxon cites as grounds for optimism, on the theory that a shared scare makes coordination easier. It is a reasonable reading. It is also the least reassuring optimistic argument available, since it requires the industry to keep producing warning shots in order to stay coordinated.
He is not the first, and the pattern has a shape
Seven months earlier, on 9 February 2026, Mrinank Sharma resigned from Anthropic, where he led the Safeguards Research team, writing in a letter to colleagues that "the world is in peril" and describing the constant pressure "to set aside what matters most." He left to study poetry. Before that, in May 2024, Jan Leike left OpenAI after co-leading its Superalignment team, saying safety culture had taken a back seat to shiny products.
Three departures, three different companies-worth of internal knowledge, one recurring structure. In each case the person leaving was senior enough to see the pipeline, made no allegation of individual bad faith, and located the problem in competitive pressure rather than in anyone's intentions. Coxon's version is the most precise: the stakes are understood, and understanding them does not stop the race, because no participant believes stopping unilaterally changes the outcome.
That is a coordination failure, and coordination failures are not solved by whichever individual has the strongest conscience. They are solved by binding external constraints or not at all. Which is exactly why post 6 was the important one, and exactly why it reached 4.8 percent of the audience.
What to watch
Two things, both checkable.
The first is whether any of the pacing agreements Coxon calls viable materialise in a form with teeth. Voluntary commitments announced after an incident are cheap; a commitment with a defined trigger, an external auditor and a named consequence is not. The distinction will be visible in the text of whatever gets announced.
The second is the shape of the next resignation. Sharma left in February and went to write poetry. Coxon left in September and published a policy argument eighteen hours later. If the next one arrives with a document, a coalition, or testimony attached rather than a thread, the exodus has stopped being a series of individual conscience decisions and started being an organised constituency. That is the difference between a hundred million impressions and an actual change of trajectory.
Either way, the number worth remembering from this week is not 101.8 million. It is 8.1 percent.