I've used Character AI (https://beta.character.ai) and been seriously impressed with the roleplaying experience that's provided there. A UX feature that really helped was the ability to "swipe left" on a reply and "re-roll" it- this can help keep the AI from getting off topic, or retry when it gives "Eliza" type responses, ie
"AI: Are you ready for your karate practice?
User: Definitely, let's get started!
AI: Great, what sort of things do you think we should do at karate practice?"
If you squint a bit and are willing to provide a little guidance in the form of leading questions, you really can have some pretty fun RP experiences, I've spent hours at this point doing little scenes and I've been really surprised at the wealth of different experiences that the AIs are capable of providing.
Other caveat of course is that it's not really suited to "longform" RP, I can't imagine it scaling to a "campaign" that you return to multiple times per month over the course of a year- I think this is a limitation of the tech at this point, as far as I know the LLM basically is always re-reading the entire chat history to generate the next response and presumably eventually this stops being feasible.
"AI: Are you ready for your karate practice? User: Definitely, let's get started! AI: Great, what sort of things do you think we should do at karate practice?"
If you squint a bit and are willing to provide a little guidance in the form of leading questions, you really can have some pretty fun RP experiences, I've spent hours at this point doing little scenes and I've been really surprised at the wealth of different experiences that the AIs are capable of providing.
Other caveat of course is that it's not really suited to "longform" RP, I can't imagine it scaling to a "campaign" that you return to multiple times per month over the course of a year- I think this is a limitation of the tech at this point, as far as I know the LLM basically is always re-reading the entire chat history to generate the next response and presumably eventually this stops being feasible.