The vast gulf between all of our actual daily experience with LLM improvements plateauing, vs what some researchers claim is going on, is pretty interesting. You can't properly understand AI progress without using the models for real work every day.