Kenneth Li argues that LLM chatbots are missing a "goal" for each conversation and proposes fixing this with human-in-the-loop model evaluations.