There's nothing special about
character.ai in particular that prevents us from using that concept elsewhere. It's just an autoregressive model, anything else in that bucket should work. I've been thinking, couldn't we tune a module for NAI on a series of our established chatlogs + perhaps other character background info? Euterpe is a Fairseq 14B model, it should be perfectly capable of simulating what
character.ai was doing. I made it into a Spanish translator of all things, which it should have had no business being able to do (this is using just the default module, no training or anything). You could set the stop sequence to the newline character and have essentially the same experience. In terms of training data for a module, you could go with the existing chatlogs, wiki articles, and (if you really want to go for it) transcriptions of their streams, though that's probably overdoing it.
