Computing Natural Abstractions: Linear Approximation — LessWrong