After Coxon's resignation post neared 150 million views, Anthropic Alignment Science Lead Evan Hubinger publicly agreed: Coxon is right. Extinction risk over 10% this decade; no clear plan yet for superintelligence alignment.
You have probably already seen yesterday's post.
Anthropic researcher Jacob Coxon announced his resignation. He said he had worked on pretraining at both OpenAI and Anthropic, and that both companies are racing toward self-improving superintelligence—effectively betting humanity's life on the outcome.
As of now, the post has drawn nearly 150 million views, more than 730,000 likes, nearly 150,000 reposts, and 260,000 bookmarks. No video. No jokes. Just long sentences. At this scale, that is already abnormal.
But what turned this from "another AI-safety post" into news was not the view count. It was that he had barely left when people still inside the lab stood up and nodded.
Not a Departing Employee Shouting—People Still on the Job Confirming
Coxon left. His superiors did not distance themselves from him.
Anthropic's Alignment Science Lead, Evan Hubinger, replied directly:
Jacob is correct here — we really do earnestly believe AI could kill all humans!
His personal judgment: the probability of human extinction within the next decade is over 10%.
Then he added:
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.
The company is trying its best, but it does not yet have a plan to solve alignment for superintelligence, and it is not clearly on track.
This is not a former employee venting. This is the person responsible for "making models safe," saying in public that he does not yet have the answer.
Hubinger now leads alignment science at Anthropic; he previously worked at OpenAI and Google. He is not a passerby liking a post. He is a key node on this risk-judgment chain.
In the same comment thread, Anthropic safety researcher Samuel Marks also chimed in in a personal capacity: AI developers do believe the technology they are building could lead to human extinction or something equally bad—and that this could happen "within the next few years."
Outside the Lab, Others Corroborate from Another Angle
Former Google DeepMind researcher Alex Turner (@Turn_Trout), who left in June, also spoke up:
I left Google DeepMind in June. Jacob is right: many researchers believe they are building something that could kill everyone on the planet. It was literally my day job to think about how to stop that.
Worrying about AI ending the world was not a hobby. It was what he was paid to do.
Jonathan Richard Schwarz, who spent seven years at DeepMind, takes a slightly different angle. He stresses excessive concentration of power: frontier labs have pulled capability, compute, and decision-making too tightly together, and that structure itself is unhealthy. He later turned down offers from two other frontier labs.
These people are not reading from one script. Some talk extinction probability. Some talk the missing alignment plan. Some talk concentrated power. But the direction is the same: this is not one person exaggerating. Several nodes on the same information chain lit up at once.
Why This Matters More Than the Resignation Itself
A resignation statement can be read as an emotional exit. Coxon was at Anthropic only a short time—about four months—and left before equity vested. That detail has already been used to question him: Was this media coordination? A narrative war inside the safety community?
Those questions can stand, and they should be on the record. An account with almost no prior content, and overnight view counts past 100 million, is unusual on its face.
But one thing is hard to explain away as "hype":
When a lab's alignment lead says publicly, "we do not have a plan, and we are not clearly on track," the company does not come out to deny it.
The people who understand the risk best are participating in this—and they themselves cannot clearly say how it ends.
Coxon wrote that inside Anthropic people privately use words like "endgame" and "crunch time." In a later Wired interview he put it more bluntly: colleagues' words were that the next year or two is humanity's critical window, and that Anthropic and its competitors will decide humanity's fate in that period. Hubinger's reply, in a sense, simply moved those private words onto the public stage.
One Detail That Must Be Kept Straight
Hubinger later added a line that matters:
He believes the risk from current models is low.
What he worries about is not today's Claude or GPT suddenly turning on people, but superintelligence from recursive self-improvement—and he thinks that process is coming faster than expected.
So this is not "today's AI will kill you." It is: the road to that endpoint is already accelerating, and the brakes are not installed yet.
Coxon himself split the two companies:
- At OpenAI, many people have not yet truly internalized civilization-scale risk;
- At Anthropic, the risk is understood, but they believe others will not act responsibly, so they must get there first themselves.
A classic prisoner's dilemma: whoever stops first cedes the frontier to whoever worries least. So everyone keeps running.
What Do 150 Million Views Actually Mean?
For empty posts, question bait, and meme posts, hundred-million view counts are not rare. Algorithms push low-information content into timelines again and again.
This was not that kind of post.
It has a concrete identity, concrete charges, concrete colleague endorsement, and follow-up interviews from The Wall Street Journal, Wired, and Axios. Views are high because three completely different groups are sharing it:
- Safety advocates forwarding it as evidence;
- Critics of big tech using it as material that "even Anthropic is racing";
- Skeptics citing it as a target—checking how long he stayed, whether the account is new, whether there is a funding network behind it.
Opposing sides both feed the traffic. The numbers will keep rising.
But the number is not the point. The point is: an internal resignation statement, in a little over a day, was picked up at once by a sitting lab lead, a former DeepMind researcher, and mainstream media.
What This Episode Actually Weighs
A 27-year-old researcher resigning is not unusual.
What is unusual: when he left, the people responsible for safety inside the company did not say "he exaggerated." They said "he is not wrong—and I personally put the probability above one in ten."
Add former DeepMind people corroborating from the other side.
How you read this depends on whether you buy the "over 10%" figure. Probability judgments differ by person; plenty of researchers in the field think that number is too high, the timeline too short, and the mechanism unclear.
But one thing is hard to dodge—the people saying this are not shouting from outside. They are working on the inside.
They have not given a verifiable timeline, or reproducible technical evidence. What they have given is a slice of internal consensus: the fear is real, the plan is not there yet, and the race continues.
Record this not to scare people, but to keep "one person left" and "a group still on the job did not deny it" clearly apart.
Earlier: Jacob Coxon resigns: OpenAI GPT-4o core contributor says labs are gambling with our lives