“LinkedIn discovered that training AI agents on its entire, unfiltered knowledge base was ineffective and produced hallucinations, leading to a strategy of curating high-quality "golden examples" for training data.”