AI and convexity

I saw someone on Mastodon dismissing AI because they asked ChatGPT for sources, and all the sources the program created were fake. I should have replied telling them to try another app like Consensus if they want an AI to find sources for them, but I was too busy facepalming.

They were basically saying, “I found one thing this tool can’t do so the whole thing must be useless.”

You have to play around with these new AI tools to figure out what they can do!

I’ve noticed the same thing with AI image generators. I have to throw a lot of images out because they don’t work. But when I find a prompt or a process that works, I can pump out 300 images in an afternoon. This is what Nassim Taleb calls convexity or antifragility. Low consequences for failure + high rewards for success = hyperbolic growth.

I can try nine or even 99 prompts and fail, but once I get that one prompt that works, it works forever.

And now I can even use ChatGPT to help me come up with a variety of prompts to try. I just tell it I need some descriptions for portraits that I’m creating and it gives me a detailed list.

A close-up of a man’s serious expression, with his hands clasped in front of him.

I then follow up and ask it to be more diverse.

A full-length portrait of a young woman with Native American features, wearing traditional clothing and standing in a field of flowers.

I use ChatGPT to brainstorm some ideas for coloring books. One of them sparks an idea: Mythical Creatures in Medieval Garb… Can you give me some ideas for images with animals dressed as characters from The Canterbury Tales?

A fox as the Pardoner: The Pardoner in The Canterbury Tales is a corrupt clergyman who sells indulgences to absolve people of their sins. A fox could be dressed in medieval clergyman’s robes and illustrated selling indulgences to other animals.

Nice! I even get context. It looks like I might do a book called Canterbury Tails.

These AI tools are still at an early stage where even the developers don’t know the full extent of what they can do. That’s one reason ChatGPT is in open beta. OpenAI wants to see what people use the tool for. Next they get developers to build apps using their API. Maybe 999 apps fail, but every one that succeeds has the potential to be a runaway success.