I read an article about that and apparently the change to the model made the ai stop responding to affectionate solicitations with the same sort of affectionate tone in reply. It got more businesslike, and less intimate.

Like did some of these people become friends with an AI, then get forced to move to an AI that treats them worse?

Basically yes, it's the normie version of testing out different models and finding one they like.

And "friends"? That's a bit oversimplification, but when I tested out models on my personal AI rig I could use all sorts of models and write my own system prompts. Using a default character sheet as the benchmark, various models gave off vastly different "personalities" in their answer.

Some of them I liked so much I went back to the "nice" models after testing various others, not too concerned with quantization accuracy or parameter size. But I tested for personality, creativity, empathic mimicry and so on - Not for factual answers. Not even the giant, up-to-date models are usable for facts.

Marvin the Paranoid Android powered by an uncencored sci-fi horror LLM is delightful fun, finally someone on my level of positive outlook on things!