Progress is continuous, but it doesn’t feel that way. Most days the ground moves an inch. Then something arrives and covers a mile. The transistor was one of those. This is another.
We’re in the adjustment now, and it’s going the way adjustments go.
Some people were sure these tools would take their jobs. Many of them found the opposite: the work got easier, they build more, and they build it faster. Others used the tools to make a job that didn’t exist, one person doing what used to take a team. And some jobs are going away. That’s true of every leap, and pretending otherwise doesn’t help the people it’s happening to.
Some people will use the tools to do harm. That’s also true of every leap.
What the thing is
What surprises me is how few people understand what the thing is. They think there’s a brain in there. Hopes, dreams, emotions, ambitions. They talk about what it wants.
Here is what it is. A large language model is a file. A very big one, full of numbers, sitting on a disk. When nobody is calling it, it does nothing. It isn’t waiting. It isn’t wondering about love or the secrets of the universe. There is no “in there” between one request and the next.
A brain is a different kind of thing. A brain never stops. It exists when it’s doing nothing at all, and it generates its own wants. Leave a person alone in a room and they’ll start something. Leave a model alone and it’s a file.
The hopes and ambitions people hear in the output came from people. The model learned from what we wrote, and we wrote about what we want. The training is human too. People chose what it read, what counted as a right answer, and when it was good enough to ship.
What happens while it runs is an argument worth having. It doesn’t change anything that follows.
Every run begins with a person
The tool needs direction. That doesn’t mean hand-holding. It means someone sets the goal, and someone decides how far the tool gets to go on its own.
You put a robot on the Moon. The signal lag is too long for reflexes, so the robot has to make some decisions itself. You give it a learned model for that. Then you put a conventional deterministic controller over it as the master, with hard limits the model cannot cross. You set the goal from Earth. The model picks the path. The controller says no. A person wrote all three layers. The only brain in the system is a quarter million miles away, and if that robot drives into a crater, nobody will ask the robot why.
Bring it back to Earth and the shape is the same. The goal comes from a person. The limits come from a person. Whatever the model decides in between, it never starts on its own. Every run begins with a person.
Someone will say: put it in a loop, give it memory, and it never stops. Fine. Someone wrote the loop.
Someone turned it on.
That’s where accountability sits. The tool can’t be fired, sued, or jailed. It has nothing to lose, and accountability requires something to lose. So it lands where it always has, on the person who set the goal, granted the access, and pressed go. An employer answers for what’s done in the company’s name. An owner answers for the dog.
That includes the people who built it. Nobody can predict everything a model will do. That’s not an excuse. Shipping something you can’t fully predict is a decision, and the people who made it own it.
“It did that on its own” is going to be said in a lot of courtrooms and boardrooms. It shouldn’t work.
The tools can act. Only people can answer for it.