Clay until someone writes the word.

A golem is a figure of clay. It does nothing at all until someone inscribes the right word into it. Then it stands up and goes to work.

It is the oldest story we tell about a made thing that acts on its own, and the mechanism in it is language. Written text, placed into the thing. Inscribe the word and it moves. Remove a single letter and it stops.

Which makes it a surprisingly literal description of what we now call artificial intelligence. Written language goes in, behaviour comes out, and the quality of the writing determines the quality of the behaviour.

The warning is the useful part

Every version of the story ends the same way. The golem does exactly what it is told and never what was meant. Sent to fetch water, it floods the house, because nobody specified when to stop.

That is the alignment problem, stated in the sixteenth century, and it is why buying the model is not the same as being finished. Somebody has to write the instruction, and then somebody has to think about all the ways a literal reader could satisfy it while ruining your afternoon. That gap between what you said and what you meant is the entire job.

Better company than you’d think

In 1964, Norbert Wiener, the mathematician who coined the word cybernetics, published a short book on machine intelligence and moral responsibility. He called it God & Golem, Inc. It came out the year he died, and it won the National Book Award.

The man who founded the field reached for this metaphor when he needed to explain what it means to build something that acts on its own. We took the name because it is the oldest continuous thread in thinking about intelligent machines, and it is still the most accurate one.

The off switch is one letter

The best-known telling is from Prague. The word on the golem’s forehead is אמת, emet, which means truth. Three letters, and they are the only thing holding the golem up.

Erase that first letter and what is left is מת, met, dead. One character is the difference between a machine that works and a pile of clay. The same thing is true of the systems we build: what breaks one is almost always the instruction it was given, and almost never the model underneath it.