Distant supervision — is instantiated by → DeepDive

Distant supervision is introduced as a general strategy — trade per-example evaluation for aggregate statistics or noisy labeling rules — and the paper cites three concrete instantiations of it side by side: generalized expectation criteria, which asks users for population-level statistics; DeepDive; and pattern extrapolation from low-recall rules. DeepDive is specifically the rule-based member of that trio: a knowledge-base-construction system in which users supply rules that each generate many weak labels automatically, rather than hand-labeling data directly (concrete-problems, §"5 Scalable Oversight", p. 13). Citing DeepDive matters because, unlike semi-supervised RL, it is a working, previously deployed production system rather than a fresh proposal — evidence that the distant-supervision branch of scalable oversight already has real engineering precedent outside of AI safety research.