Bottle in the sea
Pool : 1f r1, 4zg, 4zj, 4zk, 4zi. I try all cydonia models updated on weekly basis (10K prototyping sessions, 50K-80K convo goals, I build assistants and RPs).
4zi have some good emergence but the adherence is drastically lowering compared to 4.1 models, and none in the quoted models are sharing the same problematics at sysprompting and structuring the context. Making "cydonia presets" no longer a sub-arch equation but an individual one. I understand that to set a proper 4.2 evolution, it was rodeo. But the crude fact have to be shared without any filter : Snowpiercer, i get more from this dusty model than from the last cydonia iterations. Roccinante too, but it's another style.
The trend is not maintaining my optimistic discover of the model in 4.1, i just share a bottle in the sea for more pragmatism and rationnalism in the strategies. It's really a model i'm appreciating, it's not a rant.
You're right. This model really has a problem wanting to drift into "Polite Assistant" mode. It blows through system prompts and sometimes even author's notes. Yes it's smarter and writes better than 4.1. But likes to get "safe" a LOT. Especially in extreme scenario's. I'll go OOC and it will admit it's not following anymore or it's "Having a hard time".