Measuring few-shot extrapolation and program induction