Potassium 737

We have two internal competing "ideas" for PPv3 and a little disclaimer: nothing here is "final", and PPv3 likely won't just drop out of nowhere, future pp updates will be staged as realms. To keep things simple let's refer to one as PPv3.A and PPv3.B PPv3.A: While I'm the one that's pushing this path forward it was built off the back of concepts written in several "white papers" back in 2020 -- modernized of course, (and now) taking from osu in places, this is what PPv3 was initially and it has since been revamped / continued The main thing holding this path behind was algorithmically solving parity as maps got more complex as well as problems relating to the bpm cap. Both of these issues have been solved. It can handle any bpm you throw at it and whatever hellish parity you can conjure (im looking at u doug) I'm in the tuning phase for it, but I keep getting distracted by making it's visualizer fancier, sorry I am extremely optimistic about this path as it means nothing about PP is a black box and once (if) it's deployed PP can be more "live" and iterated on faster, making reworks more commonplace, so please, if this path wins out, understand it's release is not "final" -- this version of PP would be akin to ScoreSaber 2, defrosting it and returning it into active development, constantly evolving. As for per-map curves or dynamic curves in general with this approach.. it's a consideration and wouldn't be hard technically, but the early results I'm seeing I feel are promising enough not to warrant them. PPv3.B Qwaysx is gonna hate me for this one but I'll do my best (just know some of this could be wrong and his work is in much earlier stages due to being employed and not a basement dwelling chud) Disclaimer just to get this out of the way first: Yes I'm about to talk about "AI" but not "AI" as you likely know it in the current day, pp has had an "AI" backbone since 2021 and was trained locally on consumer hardware back then and will more than likely be trained on local consumer hardware now. If you don't know PPv2.9 (what we're using currently) was created in haste as PPv2 was killing the game. It's evolved over time but to put it very simply (definitely missing some details) it's currently an early build of PPv3 modified to output a tonne of difficulty metrics about a map, we then trained a basic machine learning model back in like 2021 based off of these metrics, this has been re-trained once or twice since then. We can give this model a map and it'll spit out a pp value -- we use this value and other statistics for a maps initial weighting then we run a statistical analysis across the leaderboard a month after the maps been ranked and "refit" the difficulty. To put it very short -- PPv3.B will be another crack at the machine learning approach with alot more data and interesting, new methods, ML has very obviously progressed a tonne since the baby linear regressor we made for PPv2.9. The concepts Qwasyx has talked to me about are extremely promising and novel and I can really see them panning out, his ideas allow for more than just pp aswell but until we have a clearer picture of the end result I'm not gonna promise anything we can't deliver. This approach has many benefits, some of which overlap PPv3.A but probably not in the way you'd expect. I mention "constant iteration, tuning and not being a blackbox" a positive of PPv3.A, these are things PPv3.B can't do as training takes time and we don't really know what's going on under the hood, "tweaking" is just adjusting the data / some params, training, then hoping for the best.... but maybe that's a good thing as it can remove several biases? Idk if you got this far let me know what you think in the comments, subscribe and slap that like button I was gonna write a "cons" list for PPv3.B but I think that last paragraph just about covered them, as for pros... that's still be determined / "secret".. which I guess is an inherent con