jake-rg comments on How would a language model become goal-directed?