Side effects and control as a relationship with other agents — passes on only its demonstration half to → Learning from demonstrations instead of a specified reward

CIRL belongs to both themes because it does two things at once: it is a demonstration-adjacent technique for avoiding hand-specified rewards, and it is a relational framing in which safety depends on modeling another agent's goals and interests, a framing this theme also applies to the shutdown problem and to the reward autoencoder's legibility proposal. InstructGPT's supervised fine-tuning stage, the concrete descendant the other theme traces from demonstration-learning, inherits only the first half. It clones behavior from roughly 13,000 static human demonstrations offline (instructgpt, §"3.1 High-level methodology", p. 6); it does not query a human interactively, infer a latent goal, or carry any analogue of the shutdown problem or reward-autoencoder legibility that CIRL's relational half was meant to address. The genealogy from demonstration-learning to SFT is real but partial: the collaborative, human-modeling character this theme built around CIRL has no realized counterpart in the pipeline the other theme documents.