Economic approach to understanding impact of ML.

Better ML = cheaper predictions

Economies of scale for first two.

Declining returns on scale for accuracy but maybe not declining returns on outcome.

Current limitations:

Break decision process into components.

Complements go up in value as ML tasks get cheaper - leverage.

Judgment is a strong complement (ie objective / reward functions). Probably needs to stay in-house.

Job = collection of tasks, not monolithic. Grouping will shift but human tasks will remain.

High-level implications:

Possible models: