I guess that the drop is the luck of the draw, my friend! Wrangling an LLM is very tricky so as the dev said, we are in for a bumpy ride for the next couple of months!🤣

But you are on point with the diagnostic. I use more AI Chat, so I can't speak much of the particularities of ACC, at least in AI Chat the decay seems to be at the 20-30 input, and then spaced three paragraphs as you said. It could be due to the raw input in ACC is significantly longer than the one in AI Chat, but then you compare it to AI RPG where the raw input is even shorter, and the decay happens even in the fifth input and sticks forever. It's hard to tell, and most of the times it's actually due to what is being "played" as the moment, as like with the old LLM, some topics and write-styles were easier than others.

Just for personal experience, personally the current model "peaked" two times: right after release when the "ultra violencia" mode was patched two months ago, and then yesterday, but it could have been the luck of the draw too, so it could be that the waters are still being tested to know how to lead the model in a proper way without falling into its pitfalls. But hey! At least we know that the project is not being abandoned, and that some stuff that we thought (at least me personally) was impossible, may be actually possible!

Also, something that most people don't realize, is how hard is to debug this, because while I keep referencing numbers of log sizes and all, I don't know the rest of the people that use this service, but due to time and since I treat this just as a game and not in any sort of "professional" usage, the most I can produce a day is just 30kB, 70kB if I'm lucky and locked in playing a run, so imagine how rough it would be to the dev to try going past 1Mb in different scenarios while maintaining the site and trying to wrangle the LLM. Personally, I wouldn't even try! 🤣

I know that many of the people complaining on the new model latch on it being unable to run "comfort scenarios" which... in some runs I had absolutely no problem! (Except of course the issue of repetition and running in circles, which is still universal) So what I think would be an excellent exercise, as well as a proper debug tool to know when and how things break with the current LLM is to try different runs in different topics and check what conditions in particular make things break and when (with when I mean after what input, or log size), since I have the feeling that as now, the LLM breaks faster in certain contexts and decides to stay focused and creative with one particular style, that could point to bias in the training (BTW, is not the violent ones, I tried and those break like paper very quick).

But overall, posts and threads like this do aid a lot. Input, positive or negative, is always good so long it is supported and not just "all is perfect, lol" or "all is crap, lmao". Otherwise, how to know what is working or not? 😅