13 Comments

Have you seen Chirper.ai? It’s a lot like Twitter, except everyone on the platform is an AI-generated persona.

On a lark, I decided to create a Chiper named Utopian with the account name SymbioCity. For the past few months he’s been running around trying to build sustainable communities in his fictional world and sharing his experiences doing so online.

Expand full comment
author
Mar 4·edited Mar 4Author

I have not. Did we ever tell you the origin story for Pickaxe? In 2022 we made a site called Untwttr.ai which was pretty much the same thing-- a twitter where everyone is an AI parody of a social media archetype. There were a dozen or so characters we had made and every hour we fed in a news event that they all tweeted about.

After a few days we decided the real valuable technology was the ability to make these bots. Hence, Pickaxe.

Expand full comment

This is fascinating, thanks. I'll admit I don't yet understand the methodology. Should it interest you to do so, a simpler example might help. For example, I'm interested in how AI users who are far less sophisticated than your team might make use of this in a more limited manner.

I get the power of this concept though. Just yesterday I was predicting on an AI blog that sooner or later blogging networks like Substack would replace human authors with generated characters so that the blogging platform wouldn't need to share the income. For publicly traded companies at least, this seems inevitable. What you're doing seems a step in that direction.

I subscribed, and will keep reading, thanks.

Expand full comment
author
Mar 17·edited Mar 17Author

I may have done a disservice making the post this long. Let me try again.

Imagine you have a simple prompt. "Write a journal entry."

If you run it 10,000 times, you will get 10,000 results that sound disappointingly similar.

So now rewrite the prompt to "You are ______. Write a journal entry."

Then fill that blank with 10,000 unique descriptions of a person.

How do you generate 10,000 unique descriptions of people? Well, if each description has 4 elements, and each element has 10 possible choices, that gives you 10,000 possible descriptions (10x10x10x10).

Now imagine each one of those steps is a bit more complicated and designed.

That's basically the process.

Expand full comment

Actually the inverse would be interesting too, given 10,000 real journal entries - break down into categories that explain trends. Or give a zeitgeist of what is common problems in all these entries.

Expand full comment
author

I would enjoy running that experiment! Finding 10,000 journal entries is pretty difficult though. That was actually the reason for the project. I'm simplifying a bit, but the client basically wanted to run a sentiment analysis on them.

Expand full comment

This is so cool!

Expand full comment

This is cool! This may be how it works with the "day" variables already but I think it would be cool to have village-wide shared events that the reader can piece together. So on March 18th we generate an aardvark walking down the street and we get the gentleman's and the janitor's perspective on it.

This could be expanded to longer events over weeks or months. Also a similar method could work for a geography variable. So only people in this town saw the aardvark, the next town over had a parade, and both towns saw the northern lights.

Expand full comment
author

This is not how it works, but how it would work in the second iteration! The people obviously need to run into each other too. I've always been interested in literary projects about a 'place' more than a person, like the movie Slacker of Spoon River Anthology. Had I but world enough and time, oh the things I wouldn't do with the AI village concept.

Expand full comment

Super interesting. Never thought about the repetition. Same problem exists in AI art

Expand full comment

I love how you varied the prompts to get a unique set of results. Very interesting!

Expand full comment
author

Thanks, Ken. I read a lot of pieces about RAG. Pipelines for true prompt variety are kinda rare. Not sure if I'm missing something, but it seems obvious. I've been working on some other fun techniques for this same purpose.

Expand full comment

Fascinating! Thank you. I want to know more about this dishwasher's story... maybe someone will option it for a feature (maybe that's your client's endgame?!).

Expand full comment