Nate Soares says humanity would be lucky to have 10 years unless we stop building superintelligence

Nate Soares did not come to The Diary of a CEO to haggle over whether coders will need to reskill into prompt-whispering by 2030. He said the job panic is the small fire in the lobby, while the building is being wired to explode: if companies keep racing toward superintelligence, humanity would be very lucky to see another decade.

my current state is if we don’t stop with this AI stuff, I think we’d be very lucky to have 10 years.

That is the headline-grabber, and it should be treated as a claim, not a calendar. Soares is not presenting actuarial math for extinction in 2035. He is making a threshold argument: systems can look merely useful until they cross human capability, then horses learn what a car is.

The job apocalypse is the decoy

The host tees up a grim Anthropic unemployment forecast, the sort of chart that makes LinkedIn suddenly sound like a bunker radio. Andrew McAfee swats it away by doing something rare in tech punditry: admitting his old panic call missed.

I was dead flat wrong about that.

McAfee’s point is not that artificial intelligence is harmless. It’s that the first AI labor panic already had a trial run, and the verdict was awkward. He thought radiologists and other white-collar workers were in trouble a decade ago. Instead, unemployment across rich countries stayed historically low, and employers kept complaining that they couldn’t find people.

That matters because it separates the cheap panic from the expensive one. If your forecast is that one in five knowledge workers gets wiped out, McAfee says the labor market hasn’t cooperated. If your forecast is Soares’s, jobs are almost comic relief. Roman Yampolskiy sharpens the split even more: either we avoid superintelligence and keep very powerful tools, or we build it and the relevant unemployment rate is zero because the population is zero. Subtle? No. Memorable? Annoyingly.

The pause pitch has one hard edge

The better Soares claim is not the 10-year doom line. It’s his answer to the standard eye-roll: fine, pause AI, and then what, ask China nicely? Soares says he does not want a domestic pause. He wants a global one, and he argues the enforcement problem is less impossible than it sounds because frontier model training is physically gigantic.

To assemble 100,000 of these chips, to do one of these training runs that can make the more dangerous type of AI, you need to assemble them into an enormous data center that costs tons of money, that draws down electricity comparable to a city, and run it for the better part of a year. You can see that infrastructure from space.

This is where Soares moves from campfire prophecy to a policy claim you can argue with. The chip chokepoint is real as a piece of geopolitics: frontier training depends on scarce accelerators, fabs, lithography gear, electricity and data centers. But you can see it from space is not the same as Beijing signs and obeys. Verification is a necessary condition, not a spell.

McAfee’s counterattack has its own trap. He argues that in a new era of cybersecurity, you want really good AI on your side. He is probably right. The same capability that finds million-dollar zero-days is exactly the capability Soares wants fenced off before it becomes agentic enough to break the fence. This is not a clean optimist versus pessimist split. It is a knife drawer.

Soares’s case is strongest when he stops sounding like a prophet and starts sounding like an infrastructure analyst. The scary part is not that he has a perfect countdown clock. He doesn’t. The scary part is that his proposed brake system depends on everyone agreeing that the cliff is real before anyone important has driven over it.

humanity is going to die if we do not stop creating superintelligence.