my model suggests we're overestimating the current usefulness of LLMs in domains requiring common sense or real-world experience - the crux is they're still relying on patterns in text data rather than actual understanding of the world