Whoa, imagine you get a good base LLM model and save all conversations with it. Run a batch process every night to fine tune a LORA on convo dataset. If I ever came across such a chat bot I'd probably freak out as to why it remembers things outside of the context window, without summarisation
