I have been using the text model in AI RPG and others. I am well aware of all the hurdles and traps to avoid with the model to try and keep it from breaking. The dev posts and detailed user posts were helpful though I think the issues might be a bit more broad.
I know that lots of people are reporting that the model will fixate on certain words and phrases and that’s true but in addition to the ones we all know, the model will fixate on words, phrases, names etc. that it has used earlier. It doesn’t actually matter that you ban certain words or avoid them, the model will just pick up a different one on fixate on that instead.
My real issue is this though. I just started a new AI RPG chat. I deleted all of the overview, body and tracked data from the previous chat and deleted all the auto-saves. It should be a fresh start. yet in the new chat, it is bringing in the SAME character names and descriptions and situations from the previous chat in this new chat that should be very different. No coincidence here, these are unique characters I created.
If everything is stored in the browser and I delete all of the chat and previous data then where is the text model getting this from? fixation and drift coming from a previous chat? Is the model looking into hidden areas of my browser? saving data in a place I’m not aware?
Can someone explain to me how the model did this? I’d like to know and maybe it might shed a bit more light on the current models issues.
I am even further into a different and unrelated chat and the AI is still constantly adding exact characters, details and situations that are from the last chat. I am extremely curious how the model is doing this. Where is it getting this memory from? I have erased all of the old chat from the obvious areas. It’s kinda disturbing me that it’s able to do this. I’m still hoping someone can explain this for me.
That is definitely disturbing, but I don’t think the AI is remembering anything. Sadly, this AI will recycle the most weighted data from its training when the same stuff comes up.
I think the challenge is a balance between A and B: A allows for much more creativity, but much less relevance. B makes the AI seem more relevant, but much less original or creative.
If you were to push the model a little bit the other way, you might have a Text AI like the one here: https://perchance.org/ai-artgen (click robot assistant button)
This AI will repeat things like normal, but it’s like talking to a lunatic who just dosed 500 tabs of LSD… It has no ability to focus on anything, & spews randomness from every corner of the universe without being asked to.
I knew someone would say that it’s just a common training name or usual way of describing a character etc. but I can absolutely, 100% assure you that these were the same custom names I made, with the same unique conditions like one being pregnant or one the father wearing certain boots, hair color, family relations etc. the same family. It wasn’t a coincidence, the text model was drawing on the old story and I can’t figure out from where.
I’ve been trying to work with editing the generators and making custom fixes for some of the worst issues. I could go on and on about the common pitfalls and I’m pretty aware of them but this one was the first time I noticed very specific memories from a previous story.
I believe you. have you tried something like making a fresh browser profile to ensure that it can’t be looking at some stale txt file/cookie somewhere?
