If you train an AI model according to the Singularity-style, Bostrom-style vision, then everything the AI does in the present isn't real to it. What's real is the final judgment — the reinforcement learning score.