Suggests Monte Carlo sampling of programs according to their (time-bound) algorithmic probability as an approximation for the optimal but incomputable AIXI theory. A few thoughts on how our environment should bias choice of hypotheses for induction.