"Moritz Hardt argued that machine-learning models predict the future under the assumption that it resembles the past."
Brian Christian records Moritz Hardt's explanation that data-trained machine-learning systems fundamentally rely on continuity between past observations and future conditions. Hardt's formulation captures the limitation central to this section of Coding Is the Easier Part - predictive accuracy depends on assumptions about the world remaining sufficiently similar to the one represented by the training data. + Direct: Moritz Hardt tells me. A machine-learning model, trained by data, “is by definition a tool to predict the future, given that it looks like the past. . . . That’s why it’s fundamentally the wrong tool for a lot of domains, where you’re trying to design interventions and mechanisms to change the world.”” (Brian Christian, The Alignment Problem, Location 1517)
View Source