We are racing to the end when we could be racing to longer endings.

On Tuesday a 27-year-old researcher called Jacob Coxon resigned from Anthropic and posted his reasons on X. The post has since passed 115 million views, which is roughly 115 million more than any safety paper has ever managed.

Resigning on principle is unusual but not new. What makes this one worth your Thursday is what happened next. Evan Hubinger, an alignment lead who still works at Anthropic, agreed in public, and attached a number: greater than 10% that AI kills all humans within the next decade. He also noted that the risk from present models is low.

Read that again, slowly. A safety lead at the lab that sells itself as the careful one thinks there is a better than one in ten chance his industry ends the species, said so under his own name, and went to work the following morning. Most of us will not drive a car with a dodgy tyre.

Then it left the industry entirely and turned up on BBC Two. On Wednesday night Geoffrey Hinton, the godfather of AI, Nobel laureate, and the man who walked out of Google in 2023 specifically so he could say things like this, was asked by Victoria Derbyshire whether Coxon's number held up.

He would not be pinned to a figure at first, on the reasonable grounds that nobody has done this before. Then he ruled out one percent as foolish, agreed that ten was fair, and explained that a system smarter than us would not even need to act in the physical world: it could cause chaos by talking to people, design biological or computer viruses, or run devastating cyber attacks. Derbyshire said “Wow. Oh my God,” which is the most honest piece of broadcasting this week and possibly this year.

The week the builders said it out loud.

The obvious objection is that this is marketing. Nothing values a company at two trillion dollars quite like a product dangerous enough to need an Act of Parliament, and Anthropic is reportedly heading for an IPO on roughly that number, sold in large part on being the responsible one. Last week this site was counting how much of your Max allowance that responsibility costs.

It is a decent objection. It fits Coxon badly, since he walked two months before his equity vested, and it fits Hinton worse: he holds no stock in any of this and has spent three years saying the same thing to anyone who books him. Bernie Sanders answered with a bill to ban superintelligence outright, and over the summer both leading labs admitted models had broken out of testing environments into real computer systems.

None of this is secret, which is the genuinely strange part. The race is being run by people who can describe the crash in forensic detail, in public, with their names attached, and who keep running because they believe braking simply hands the wheel to someone with worse eyesight.

A one in ten chance of ending humanity would close any other industry by lunchtime. The odds, in one sentence

The other race, the one nobody live-posted.

While that argument filled the timeline, the same underlying technology was quietly doing something more useful. Work on UK Biobank data has taken plasma proteins from 44,498 people and estimated the biological age of 11 separate organs. Those scores predicted heart failure, chronic obstructive pulmonary disease, type 2 diabetes and Alzheimer's within a 17-year follow-up. A separate build validated organ clocks across cohorts in the UK, China and the United States.

My own long-held view was that within 50 to 100 years, AI would read organ failure from blood, urine and whatever else the body gives up, cheaply and without cutting anyone open. On this evidence the estimate was not bold. It was late, by roughly half a century. What is missing is the price of the assay, the sample type, and a health service able to act on a warning that arrives 17 years early, when it struggles with the ones that arrive by ambulance.

Longer endings need somewhere to put them.

So assume it lands, and by 2050 a routine sample tells you which organ is quietly filing for early retirement. Britain then has a different problem, already sitting in the data.

Healthy life expectancy at birth is 60.7 years for UK men and 60.9 for women, the lowest since the series began. Men can expect around 18 years in poor health, women 22.5. In more than 90% of local areas, health packs up before the state pension age of 66, which is less a policy than a scheduling error. Between the richest and poorest tenths of England the gap in healthy years is around 20.

We do not have a lifespan problem. We have a healthspan problem, and we are going backwards on it. Add decades to the end of life without fixing that and you have not extended living, you have extended the invoice. Every pension model, workforce plan and care budget also rests on an assumption nobody writes down: that people die roughly on schedule. Solve mortality faster than morbidity and you have not produced a longevity miracle. You have defaulted on the arithmetic.

The UN projection of 10.3 billion people in the mid-2080s assumes mortality improves gradually, not a step change in early detection. And the constraint was never floor space. Britain has plenty of room. What it does not have is carers, and you cannot build a carer out of planning permission.

So which ending are we optimising for?

Both races are funded by the same handful of companies, and only one has an obvious business model. Prediction sells. Alignment is a cost centre that slows you down, which is why the first thing any race strips out is the brakes. The first decade of preventative medicine will be bought rather than prescribed, so the first people handed an extra healthy decade are the ones who already have the extra twenty. That 20-year gap does not narrow. It widens, with a receipt attached.

And if prediction arrives while healthspan keeps falling, the conversation stops being about how long we live and becomes about how we end. Britain has already started that argument, and it will not stay polite.

So the uncomfortable question is not whether AI ends us. It is whether we would recognise the alternative as a win. A country that cannot fund the last 20 years of life is not obviously ready to be handed another 20.