The politeness constraint.

Every company in artificial intelligence is constrained by the same thing, and it is not talent, and it is not ideas. It is the supply of graphics processing units. There are not enough of them. They are contested. They get more expensive as more people want them, which is the ordinary behaviour of a scarce physical good.

We are constrained by how many people have heard of us.

Why this is the better constraint

Attention has an unusual property that compute does not: it gets cheaper the more of it you have. The thousandth person who hears about us costs less to reach than the first, because the first nine hundred and ninety-nine are doing the reaching. Nobody has ever said this about a data centre.

This is not a clever observation. It is the oldest observation in commerce, and every company in this sector knows it, which is why they spend enormous sums on compute in order to build products whose eventual purpose is to acquire attention.

We skipped the middle step.

The constraint we actually have

If the input is attention, the limiting factor is not money, or infrastructure, or headcount. It is how many times a person can be asked for something before they find it rude.

We do not know that number. Nobody has published it. But it exists, it is finite, and it is the only resource this company can exhaust. Every other input we have is either free or two hundred dollars a month.

We are therefore the only company in artificial intelligence whose scaling law is a matter of etiquette.

A note on the word “efficient”

The industry uses efficient to mean fewer dollars per unit of output. By that definition we are the most efficient company in it, and we would like this acknowledged before somebody points out that our output is a sentence.

The sentence is: give me one dollar. It cost nothing to produce, it costs nothing to serve, it does not degrade under load, and it works in every language. We are not aware of a model with those properties.