The alignment lead says above 10%. The compute bill says $517 billion

Jacob Coxon resigned from Anthropic and left the AI industry. He is 27 and spent three years on pretraining research, first at OpenAI and then at Anthropic. He posted the reason himself.

He told the Wall Street Journal he no longer believes any single lab can safely build what his employer is racing to build. Under the aggressive scenarios, he said, things could be out of control by the end of next year. CoinDesk carried the account on Wednesday.

He is leaving the industry outright instead of moving to another lab, which rules out the usual reading of a departure as a career step. The Terminator, meanwhile, appears in the headlines rather than in anything he wrote. His post names self-improving superintelligence and a race between two companies; the film belongs to the coverage.

The agreement came from the safety side

Evan Hubinger, the Anthropic alignment lead, answered in public and in plain terms.

“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.”

— Evan Hubinger, CoinDesk, 9 September 2026

Evan Hubinger, Alignment Science Lead at Anthropic. Source: CoinDesk, 9 September 2026

Hubinger wrote that he and colleagues earnestly believe AI could kill all humans, and put his own estimate above 10% over the next decade.

Note who is speaking. Coxon built pretraining, the stage where a foundation model absorbs the bulk of its data, before anyone tunes its behaviour with RLHF. Hubinger runs the team paid to make them safe. Both ends of the building said the same thing on the same day.

Two public numbers, pointing opposite ways

We counted Anthropic's compute commitments last week and reached $517 billion across eleven months of contracts. That is the company's own spending on making the systems larger.

Set it beside the other public figure from the same company: the head of alignment putting the chance of human extinction above 10% over ten years. Both are on the record, from Anthropic, in the same week.

Hubinger's phrasing carries the weight. He said the company has no plan yet for aligning superintelligence, and that it is not clearly on track to get one.

The part that can be checked

Coxon put a date on his warning, which most AI forecasts avoid. He named the end of next year as the point where the aggressive scenarios turn ugly. That is fifteen months out and inside the horizon of anyone reading this.

Compare it with the forecast we covered this morning, where the chief executive of Arm said AI will help cure cancer within our lifetimes. That one cannot be tested for forty years. Coxon's can be tested by the end of 2027.

A resignation proves nothing by itself, and neither does a percentage. What makes this one worth filing away is the deadline attached to it, offered by someone who spent three years doing the work.

This piece is informational, not a recommendation to buy, sell, or hold any asset.