How long until AI can run without humans? Cotra and Lee disagree by decades

Asterisk ran a structured disagreement between Ajeya Cotra and Timothy B. Lee on a precise question: when could AI sustain itself entirely, running the mines, power, factories, and chip fabs it depends on without human labor? Cotra thinks it is more likely than not within ten years, possibly sooner. Lee puts it below 10% within twenty years, with a median closer to fifty, and gives real weight to it never happening at all.

The split comes down to the bottleneck. Cotra sees cognition as the hard part: once models are smart enough to operate the robots we already build, and humanoid production scales into the tens of thousands of units a year, the rest follows. Lee argues the physical world is stubborn in ways software people underestimate. His example is sidewalk delivery robots, which worked fine six years ago yet never scaled, held back not by any missing breakthrough but by the slow grind of making hardware reliable, durable, and cheap enough to run at a margin. They agree on what to watch over the next few years: humanoid production numbers and falling costs, measurable gains in robotic-hand dexterity, and a detailed mapping of how many humans it still takes to keep the AI production stack running.

Why it matters

If your read on AI risk or economic disruption depends on full automation, this hands you concrete near-term signals instead of dueling intuitions. Watch robotic dexterity benchmarks and humanoid unit costs over the next two or three years; they are the measurements that will show which forecast is holding up.