There’s no reason this couldn’t be implemented now. The main barriers are inference speed and cost, since to implement this would require continuously running the LLM on all newly available text from the user and choosing quickly when to interject, and the difficulty of programming complex behaviour.