DeepDive
Concrete Problems in AI Safety — inherited
A prior knowledge-base-construction system (Shin et al.) that asks users to supply rules generating many weak labels, cited as an example of the distant-supervision approach.
DeepDive is a working system from Stanford that builds knowledge bases from rules its users write, each rule generating many rough labels at once. Concrete Problems in AI Safety (2016) cites it as proof that this style of weak supervision holds up in practice.
One section describes the rule-based system itself; the other explains why the paper treats it as evidence that weak supervision holds up in practice, rather than as a technique to adopt directly.
A rule-based system for building knowledge bases
DeepDive, due to Shin et al., is a knowledge-base-construction system, developed at Stanford, that lets users supply hand-written rules rather than hand-labeled examples: each rule, once specified, automatically generates many weak, noisy labels over a large corpus, which the system then combines using probabilistic inference to produce a coherent knowledge base. It predates Concrete Problems as a working piece of deployed software rather than a research proposal, and the paper cites it as one of three concrete instantiations of Distant supervision alongside generalized expectation criteria and low-recall pattern extrapolation (concrete-problems, §"5 Scalable Oversight", p. 13).
Its role as evidence, not technique
DeepDive does no work inside Concrete Problems beyond being cited; its importance to the paper's argument is that it exists at all. The edge Distant supervision — is instantiated by → DeepDive singles it out from the other two examples as the rule-based member of the trio, and observes that citing it matters because, unlike the semi-supervised RL machinery the paper proposes fresh, DeepDive is proof that the distant-supervision branch of Scalable oversight already has real engineering precedent outside AI safety research. The concept is a child of Distant supervision within the Scaling oversight without a learned reward predictor theme, sits in the Two or more remedies for the identical problem, sorted by where they intervene connective theme, and belongs to the Scalable oversight made real: from 2016 proposals to InstructGPT's RLHF pipeline supertheme — one small, concrete anchor for a much larger and mostly speculative section of the paper.