Reward pretraining — is a concrete special case of → Variable indifference

Variable indifference is stated as an aspiration rather than a method: the paper wants a general way to route an agent's optimization pressure around chosen variables, such as the reward function itself or human behavior, while it still optimizes everything else, and admits that "truly solving this would have applications throughout safety" with no concrete algorithm attached (concrete-problems, §"Variable Indifference:", p. 10). Reward pretraining is what that aspiration looks like for exactly one variable, the reward function, when it is fully realized rather than merely routed around: training the reward function via supervised learning before any environment interaction begins removes the agent's ability to influence it, not by penalizing attempts to influence it during optimization, but by finishing that variable's specification before optimization starts (concrete-problems, §"Reward Pretraining:", p. 10). The comparison exposes what variable indifference is still missing as a research problem: reward pretraining only works because the reward function can be fully fixed in advance, whereas most variables an agent should be indifferent to, such as human behavior, cannot be pretrained away in the same manner and would need genuine indifference during online optimization.