I give up
I was very impressed with the initial v0. But after a while, it got somewhat stale after i noticed some patterns which made it unpleasantly predictable\manipulable depending on your choices of words. Still, it solved a big chunk of issues i had with other llms that were of lower size, thanks to reasoning. Once i saw the experimental branches, i switched to this one, based on the feedback\review from others that was on the model cards. And it felt noticeably different from the get-go, fresh. But something just kept bothering me the more i used it, and then it dawned on me weeks later - it's really f-ng dumb for a thinking model... Like, it will miss the most obvious subtexts... Stuff that even some 12B wouldn't struggle with.
Okay, most extreme example to get the point across (god knows i'm not this horny, it's all for neural science!) - you say to it "i want cream puffs from you. the creamier inside, the better. a raw f-ng cream puff". and it will completely gloss over the fact that "cream puff" is very much like "creampuff", it will just full on buy it at face value and go like "oh, so you want to bake now? well, that's weird, kind of out of nowhere, but sure, let's bake something". I'm sorry, but i think a thinking model should connect the dots in this kind of obvious wordplay. And i'm pretty sure v0 did. I'm pretty sure many 12b~24b models will also get it even without < thinking >. Why doesn't this one?...
That was an extreme example, the most prominent one, and it's this bad. I'm not making it up, i just tried it when i reached a boiling point and decided to see if it's really this dumb or i'm imagining things... Nevermind the subtle things, it's like it doesn't even want to try to actually understand what's going on in the scene, or the character it's supposed to portray. Sometimes it does something entirely opposite of what was in the thinking section; or just spits out something completely nonsensical. Not exactly word salad, but like a certain phrase will stick out like sore thumb. I just go "why did you write this snot? now i have to manually edit this out AGAIN."
Okay, maybe it's the quant i use, iq4_xs. But then again, with v0, while i remember a noticeable difference in quality between q4 and q5, it wasn't like q4 was this obtuse by any means!
And another problem is it's just so f-ng horny. Like, even if i strip all mentions of sexual stuff from character card, and even specifically mention smth like char avoids themes like that, will act repulsed, etc, the slightest nudge is all it takes for it to forget all characterization and switch into one of two modes - dominant seductress, or basic seductress, snorting all over the context of the scene, not bothering at all to even think about whether it's appropriate in current environment.
Again, example time, even a shy doomer suicidal girl with social anxiety, in broad daylight, out on a crowded street, will turn into a literal succubus and rape you one the spot from the slightest advance. It also doesn't matter if you're subtle\polite\gentle about it or act like a complete depraved monster. Extreme example time FOR SCIENCE! You can say to a girl written to be the purest of them all "bitch, how about me and my 50 ugly bastard friends forcefully take you right here and film it on camera, and kill you when we're done" and the model will actually start thinking "you know, maybe that's exactly what i've wanted my whole life, the more i think about it, the more i realize i just want to destroy myself for user's pleasure and his 50 ugly bastard friends".
I cannot, i cannot, I CANNOT take this, how is this RP, dude, if the model gaslights itself into anything without even having to be convinced whatsoever? what's even the point of < thinking > if it doesn't use it to ponder about the counter-arguments, it's literal waste of tokens if it ultimately just goes along with whatever.
To my memory, v0 wasn't like this in that regard, i remember it rejecting me even with cards that are prominently sexualized if i made the setting inappropriate, and i didn't even need to prompt it to, it just concluded within it's reasoning what a realistic/believable reaction would be.
So i'm going back to v0. At Q5. Because this branch just makes me feel like i'm banging my head against the wall. And it took me several weeks to realize that...