History — 2026 week 29 (Jul 13 – Jul 19, 2026)
Newest first.
2026-07-19 — LTR concept enriched with Grebennikov’s two training lessons (0 new notes)
Enriched Learning to Rank with the two production lessons from Roman Grebennikov - Personalizing Search Results in Real-Time that previously lived only in the video note: a new section “What the Model Optimizes Is a Business Decision” (the Stanley bong story — raw clicks reward curiosity; weighting purchases in the NDCG perfect ranking encodes the business goal, and the same lever optimizes margin) and a new subsection “Feedback Loops: Don’t Train on Data Your Model Produced” under Position Bias in LTR Training (the +8% → +6% degradation from training on the model’s own clicks; fix via an Exploration vs Exploitation ~1% shuffled exploration segment, with IPS as the alternative). Also expanded the video’s blurb in the note’s Articles list to name both lessons.