Prompting AI

People think that AI is almost magic — and from the outside it basically is. But on the inside it’s just a whole crapton of calculations. There’s an often-repeated statement that a trained AI model is just a copy of, basically, the internet. It’s not. It’s not a compressed version of the internet either, certainly not from the conventional concept of “compression.”

But regardless of how it works, it’s certainly not magic. It’s garbage in, garbage out. Sometimes with an extra helping of garbage.

To make something useful come out the end, you need to be not only a good thinker, but a good communicator. You have to be able to convey what you want to the magic box with great clarity. And that’s still something that only humans, for now at least, really possess (*).

The way I tell people how to talk to an advanced AI is to think of yourself giving a week-long independent project to a competent grad student. You’re not going to be talking to them while they are off doing their thing, so you have to be very clear in terms of how and what you want done.

Maybe we’re coming to the point where some of the liberal arts people finally get a win?

* - Fun fact: you can ask LLMs for good ways to prompt them. For instance you can give a corpus of your own writing and have it summarize your writing style in a way that’s useful for a prompt. Kind of inception, but it’s a fun use case.