Comment by SwellJoe

Because an LLM does not get smarter and it does not learn. It is fully cooked when it launches and only gets smarter with more training of some sort, a long, expensive, process that doesn't get phrased as "we thought the model would get smarter".