Lately I have been trying out a few AI girlfriend and chat platforms like CrushOnAI, JanitorAI, and CandyAI. Using their different models, it is impressive how lifelike and emotionally responsive the interactions feel, almost like GPT or DeepSeek applied really well.
I am curious though, do these platforms mostly rely on integrating existing models or do they also train their own? How do they achieve that level of realism? Would love to hear anyone’s thoughts or insights.
Not sure, but I doubt anything outside “safety” is allowed without custom models. So unless you wanted to talk about weather with your virtual girlfriend, then forget gpt and such.
Yeah, I feel the same… what you said makes sense.
One path is to break the model’s original limitations through prompt words, and another is to fine-tune open-source models through “toxin datasets.”
The first approach seems pretty common. The second one sounds really detailed. I wish we could learn more about how it actually works in theory.
you feed bazillion emotional conversations into an llm → it feeds you the same
The large-parameter model is not incapable of playing the wife role; it simply hides this ability in certain alignments. By fine-tuning on the Explicit Conversation QA dataset, the model can regain this ability.
Bro, this is a programming forum.
I get it, but I thought I’d ask here since there are a lot of tech-savvy people and professionals who might have a deeper insight.