It wasn't just a model switch; it was a complete fuckup.

What on earth has happened here? No one knows.

Slang? Gone. Cursing? Let's keep it PG-13, while wondering about the logistics of what possessed the dev to implement this godawful model.

No, but on a more serious note, I actually liked this model at first... not anymore though. Seriously, fucking c.ai provides a better experience than this now. I've practically abandoned the platform now and moved to this instead: solania.ai. It's limited, but it uses the old LLM, so that immediately makes it better than Perchance atm.

Until the dev removes this model, Perchance is dead to me.

The long-standing issue with Perchance is its inconsistency. If it performs well for a month, you can be sure the next month will be a downgrade. Even throughout the same day, it can be fine during certain hours and then randomly deteriorate out of nowhere.

And then there's the whole censorship thing. I remember two years ago everyone was complaining about how the model was so overly wholesome the classic "Let's not get ahead of ourselves" era. Now I realize that was probably on the Dev. Then they replaced it with another model; it was good at first, but after those stupid safety measures, the breaths started hitching. It feels like the model falls into a loop just to evade something.

And did the model change again recently? Something happened about two months ago where it started off great again, but now here we are. It’s honestly terrible. And we have no way of knowing why because we’re completely in the dark on these things we basically have to stalk the Dev like creeps just to get any info.

I genuinely love the Dev, and they deserve all the praise for this project which we all give them anyway. But on the critical side, credit where it's due: successfully wrecking every single model, whether on purpose or not, is honestly quite an accomplishment as well.

I spent a month specifically on this problem, doing all the research I could about NLG, pipeline output, negative instructions causing priming, etc. Not when I got home "a month," compiling while I'm at my 10 hour shift a day a month. The answer is no. The system bypasses all instructions for raw output, the best you can do is tell it to recreate it again without using what was said, thus priming the system to do it again because it can't retain the memory of it being relevant in the raw output.

I tried using the AI helper to adjust the code to stop the sanitizing and it would adjust it and then run a generation itself and it would still fail. It gave up.

It's SO FUCKING INFURIATING. You type "She was angry" continued into "She was angry, but not really, she was actually teasing and laughing." Worst chat model perchance has ever had.

To be honest, the old text gen model was way better in my opinion. I want the old model back.

Well, they also gave us src ability. Along with Kv storagr,rag and seperating modules instead of hooking, allows for allot more personal pref alterations as opposed to the model making most of the calls in a bogged down index and main. I agree though but there is light at the end of the tunnel I hope lol