Today's issue is supported by Chartsy
Chartsy turns your Stripe or Paddle data into answers. Ask "show MRR by plan" or "why did churn spike" in plain English and get the chart instantly - no spreadsheets, no pivot tables. Built for founders who'd rather read their numbers than wrestle them.
Start with the fact that decides how much weight to give this: who was speaking, and from what seat.
Jacob Coxon, 27, Cambridge graduate, spent three years in pretraining research, first at OpenAI, where he was among the contributors to GPT-4o, then at Anthropic. Pretraining is not the ethics team, not policy, not comms. It's the part of the pipeline where raw compute and data get turned into a model's core capability.
People in that seat see the scaling curves and internal capability evaluations before anyone outside the building does.
That's the clinical distinction. When an outside ethicist warns about AI risk, they're reasoning from principle. When the person whose literal job was making models more capable says the capability curve is outrunning the ability to control it, they're reporting from the instrument panel.
Same words, different evidentiary weight. He isn't speculating about what the labs can do. He was watching the numbers.
76M
Views on his resignation thread within a day
3 yrs
Pretraining research across OpenAI and Anthropic
>10%
His estimated chance advanced AI kills all humans this decade
end 2027
When he says things could already be "out of control"
Strip out the drama. Three testable claims remain.
The viral quotes "gambling with our lives," systems that can "hack anything" are the emotional surface. Underneath, Coxon's position rests on three claims an engineer can actually evaluate. That's what makes this worth your time instead of your alarm.
Claim 1 | Recursive self-improvement, AI that improves AI has moved from theory to active engineering goal at the top labs |
Claim 2 | The capability curve is accelerating faster than the control and alignment tools meant to contain it |
Claim 3 | Competition makes safety trade-offs unavoidable, no single lab can slow down without losing, so none do |
Notice what these have in common: none of them require you to believe in killer robots or malice. Claim 3 in particular is just game theory.
Each lab believes its rivals won't act responsibly, so each concludes it must reach the frontier first "to do it safely" and the sum of those individually rational decisions is a race nobody chose and nobody can stop. Coxon's own words for the two cultures he worked in are precise about this.
"At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they're locked in a race to get there first."
His employer's own safety chief agreed with him
Here is the detail that separates this from every other dramatic exit, and the reason a clinical read lands harder than an outraged one.
When a researcher resigns in protest, the standard corporate response is silence or a gentle distancing. Instead, Evan Hubinger, Anthropic's alignment science lead, the person currently running the safety work Coxon left, publicly backed him.
His words: "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. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence."
Read that again as a clinician would. The person still in the chair confirmed the diagnosis of the person who left it. He did add an important qualifier, he considers the risk from today's models low. The fear is about the trajectory, not the model you used this morning.
The honest read - hold both of these
Two things are true and neither cancels the other. One: this is a genuine, credentialed insider warning, corroborated on the record by the safety lead of the company he left not a crank, not a marketing stunt.
Two: probability estimates like ">10%" are subjective judgments, not measurements; smart, informed people at these same labs put the number far lower, and "out of control by end of 2027" is a forecast that could simply be wrong. A clinical read means taking the signal seriously without treating one researcher's probability as a fact.
He is not the first instrument to flash red
Part of reading this clinically is noticing it's a pattern, not an event. The same seat has thrown the same warning before.
Frontier-lab safety researchers who resigned with public warnings
Bar length reflects relative public reach of each departure, not a precise metric. The cluster is the point: the people closest to the models keep leaving over the same concern.
When one person quits, it's a story about that person. When four people whose job was to see the risk first all walk out citing the same thing across two years, it stops being about temperament and starts being about what they can all see from inside that the rest of us can't.
🔮 The Bottom Line
The clinical reading isn't "AI will kill us all" and it isn't "ignore the doomer." It's narrower and more uncomfortable than either: a credentialed insider, backed on the record by his former employer's own safety lead, is telling you the control tools are lagging the capability curve and that the race structure makes slowing down individually irrational.
You don't have to accept his probability to take the signal. The people with the clearest view of the instruments keep resigning rather than keep reading them. That's the data point. What you do with it is yours.
That's this week - a clinical read on the story everyone else covered as a scream. If you want more of these, our breakdowns hit Instagram first at @unseen_ainews. And hit reply with your honest read: signal, or noise? We'll share where readers land next week.
- hiPreneurs
📧 Forward this to 3 entrepreneur friends who need to see this opportunity
