Np, hope it helps (and feel free to let me know how it goes, I’m curious to know if it works for anyone else). And yeah, I remember trying CharacterAI briefly in the past for language learning myself; main difference I find with this type of setup, is 1) The corrections are “on the side”, not organically part of the conversation itself, so you have the main conversation which the AI is focused on and then you have suggestions/corrections you get if you say something its evaluation thinks has incorrectness for that language. 2) Both apps have forms of conversation where they can be more unguided in terms of what you talk about and have “roleplay” type of scenarios that are a little more structured for practicing specific kinds of things. For example, I was just doing a little talking with a “roleplay” on Tutor Lily called “explaining symptoms to a doctor.”

So if I compare it to trying to learn through just any LLM, point 1 seems to be the most significant difference. Since technically you could already get an LLM to roleplay most things, albeit with more effort than with these apps’ scenarios. But getting corrections on the side while talking to an LLM seems to be a more specific engineering/design thing that goes beyond LLMs alone. I might try to ask the creators and see if I can get an answer, but I’m unclear still on whether these apps are powering the corrections purely through an LLM itself or some other kind of AI evaluation with it.