Penicillin, the transistor, vulcanized rubber, kevlar, aspirin, stainless steel…

Every breakthrough you can name waited on one thing: a pair of hands.

A graduate student pipetting at two in the morning. A mad scientist running the same reaction for the ninth time, writing down what happened, setting up a tenth. 

All of it moved at the speed of somebody's Tuesday in the factory or the lab.

The idea was the easy part. 

Running the experiment, ten thousand times—or long enough to have a happy accident—was the wall. 

Pete Florence, co-founder of Generalist, hit that wall a decade ago. 

He was trying to build cheaper solar cells. He realized the only way to go faster was to have a hundred more of himself in the lab, running experiments around the clock.

But life isn’t a cartoon. He couldn't do that. But… what if he could? 

So he quit and went into robotics. 

He now runs a company building general-purpose robots. The thing he's chasing? The hundred clones.

And these clones are real. Every frontier AI lab is taking them seriously. Because they are where the ENTIRE trillion-dollar AI industry is inevitably headed. 

Three words to write in Sharpie on your mirror so you know exactly what’s set to shape the next decade: 

Self-Driving Labs

They’re called “self-driving labs.” 

The name is self-explanatory: they are labs that run themselves. Tip to tail. 

Who reads the literature? The lab. Who chooses which experiment to run? The lab. Who mixes the materials, bakes them, measures the result, and decides what to try next? THE LAB.

No human in the loop. No nights off.

A lab at Berkeley called A-Lab ran itself for seventeen days straight and worked through fifty to a hundred times the samples a person could touch in a day. A sister lab down the hall makes a fresh sample every few seconds.

Point a machine like this at a problem and it does a year of grunt work over a long weekend.

And the problems they could solve are plenty. 

We’re awash in giant, invisible, boring difficulties that come down to hunting a haystack of materials for the one that works.

Consider just a few. 

Trillion-Dollar Haystacks 

Distillation burns 10 to 15% of the world's energy boiling mixtures apart. Even a dent in that is worth hundreds of billions—and enough freed-up energy to power entire nations.

The catalyst that makes fertilizer, and feeds half the planet, is a century old and eats another 1 to 2%. Solve that? You feed the next billion on the energy we already waste.

Batteries wait on the next electrolyte. Jet engines and turbines wait on an alloy that takes more heat. Cement, barely changed since Rome, burns 2 to 3% of the world's energy—and the recipe's been begging for an upgrade for two thousand years. 

Green hydrogen leans on metals too rare to scale.

If self-driving labs can shave even a few points off any one of these, the number is measured in hundreds of billions of dollars.

Meanwhile, the cost for these labs is plummeting.

Cheap Enough for Everyone

A robot arm has shaved off 80% of its cost since the turn of the century. 

The machine that once cost six figures when Clinton was in office is less than $20,000 today and 1,000x more useful—and the price is still falling about 40% a year on the humanoid end.

Florence's bet is that once a general-purpose robot can use its hands like a lab tech, that cost collapses and every lab becomes a self-driving lab.

The Cold Water 

Of course, self-driving labs aren’t without problems. A-Lab's team said their robot made forty-one materials that had never existed. 

A chemist named Robert Palgrave went through the data and concluded it had discovered nothing—the samples were misread. 

But that mistake, though perhaps embarrassing, reveals where the real value sits. The machine can run the experiment faster than any human alive. It can't tell you what it made. 

Genius at doing, fool at judging—and judging was always the hard part.

So the bottleneck will move. 

Judgment Gets Expensive

Software gets cheap, hardware will get cheaper, which makes the only thing left expensive: judgment. 

The bottleneck used to sit at the bench, in the hands. 

Increasingly, it will sit in the eyes—in the few people who can look at what the machine coughs out and tell gold from garbage.

Every lab and drug company will own a firehose of maybes. 

The scarce thing is the judgment to tell which answers are real. The premium goes to whoever turns maybes into breakthroughs.

The robots will run the lab. Somebody still has to be a scientist. The future is weirder than you think.