The asymmetry

Words like enormous and costly have been doing a lot of work so far. Here are the actual numbers, because the shape of the cost is strange.

Training a frontier model. Thousands of specialised chips running without pause for weeks or months, plus the engineers, the data work, and the failed runs that produced nothing. Training GPT-3 was estimated to have used around 1,300 megawatt-hours of electricity, roughly a year of power for a hundred homes, and to have cost somewhere in the millions of dollars. Runs since have been reported in the hundreds of millions. The figures move and the estimates disagree. The order of magnitude is the point.

Producing one answer for one person. Inference, the step from the last slide. A fraction of a second on a handful of chips. Small enough that your answer is often free, or covered by a subscription, or costs a fraction of a penny.

Copying the finished model. Effectively nothing. It is a file. Copying it costs what copying any large file costs, and the copy is not a lesser version. It is the same model, and it can serve millions of people.

That is not the usual shape of expensive things. A bridge costs a fortune and then carries a limited number of cars, so the cost keeps pace with the use. Here the entire cost is one act, performed once, and everything after it is cheap and endlessly repeatable.