By skyforbes For a new application I'm considering, while I've got a good idea of the core features that would be necessary and its inherent nature is well defined, I can envisage this usefully evolving over time on a per-user and unpredictable basis. e.g. for some reason the application decides to ask a question of the user such as "would it be useful if …", user says, "sure, that would be good", and this leads to a need to capture arbitrary data that the application decides is relevant for whatever it was talking about. This essentially becomes a new feature that the application has for that user, until perhaps the user decides that it's not needed or needs to be different. While something similar may have been thought of by the app for another user, the precise nature and KV data might be different. This could be generalised with the concept of "abilities", some of which are system defined, while others are LLM defined over time; in the latter case, the LLM decides on a schema for itself and how / when to use the new ability that it thought up.
Wondering if anyone has experience to share of developing an LLM based system that evolves in some way by itself over time based on user interaction, such as whether it actually worked, pitfalls, successes, things to watch out for etc.