AI Shortcomings

This post was originally published on golifelog.com.

Over the weekend I tried to familiarize myself more with the Kindroid platform. I set up two characters and a few group chats to let them interact with each other. The more I use the platform, the more I can see the boundaries and limits their model has.

For example, if you’ve got a complex scene formed out, like in a pen-and-paper roleplay, the AI regularly forgets which of the characters holds which item or sometimes mixes up the characters, mixing “traits” between them. Not that unexpected, to be honest. On the standard plan I only get 1000 characters to describe the background of a character.

So, I find myself constantly adding in-between-the-line information instead of just chatting back and forth and keeping the story going.

I saw one prebuilt character in their library that sets you in the position of a new commander of a spaceship, with several NPCs (“played” by the AI) that seems to be excellent. I might try that out, but for now, experimenting with my own characters is quite fun.

One thing I should have done from the beginning, though, is copy over every message to a separate document. I’m not sure if I can export the chats later on, but some of those chats could be a spark for a short story. It’s quite fascinating letting the characters “play” on their own, only guiding the general direction, even if it means regularly bringing them back on the rails.

On the other hand, I’m almost surprised by the fact that it’s relatively easy to reference chats from a group chat in a 1:1. Sure, I still need to explicitly reference that group chat and summarize what happened there, but the model picks up the hints. I’m not sure how it’s technically implemented; maybe it’s time to find a good newsletter or blog that digs into such topics.