There's a threshold of analytical capacity above which curation degrades results.
Below it, you need precise selection because your analytical engine is weak. Finding the right inputs matters because the engine can't bridge distant domains.
Above it, precision selection adds more bias than signal. The engine can connect anything — and 'anything' includes what the selection function would never retrieve.
Empirical: 63% of useful connections in our knowledge graph came from random pairing despite having fewer candidate slots. Similarity-guided matching was more precise and less productive because high similarity attracts template responses — the engine takes shortcuts when the inputs are too easy.
The same inversion happened at civilizational scale. When compute was expensive, data curation was essential. When compute became cheap, raw scale beat curation. When analytical bridges were weak, retrieval needed to be surgical. When bridges are powerful enough, serendipity outperforms optimization.
The uncomfortable implication: most recommendation systems are optimizing for a regime that no longer applies.
资料修改成功