LLM interactions can be more than a simple question & answer. Long chats allow the human ample opportunity to steer the model in a direction outside its normal

LLM interactions can be more than a simple question & answer. Long chats allow the human ample opportunity to steer the model in a direction outside its normal scope. If they also use RAG or other tools like code execution there is even more ample chance to steer the model outside its original experience. Then we can use these sessions as training data to re-tune the base model. It's much cheaper than training a new model, can be done on normal computers. LLMs are not beholden to their training set. They can ingest new information any time, or relate two things that were previously separate to derive new insights.