Jessica Taylor: the status of a simulated copy changes the choice in Newcomb’s problem
Jessica Taylor: the status of a simulated copy changes the choice in Newcomb’s problem
On October 17, Jessica Taylor published an essay on LessWrong about problems in which memory erasure leaves a person in several situations they cannot distinguish from one another. In her version of Newcomb’s problem, the choice between one box and two depends on whether the simulated copy counts as a conscious observer.
Taylor begins with the Sleeping Beauty problem. A coin toss leads to one awakening if it lands heads and two if it lands tails, with the person’s memory erased between awakenings. On waking, the person cannot tell which of three situations they are in: heads on Monday, tails on Monday, or tails on Tuesday. Taylor treats all three as equally likely. The person’s probability of heads upon waking is therefore one third.
In another example, a person chooses between coffee shops A and B on two occasions, each after their memory has been erased. Coffee shop A is more pleasant in itself, but meeting their other-day self matters more than that difference. If the person expects to choose B on the other day, choosing B now is advantageous because it ensures the meeting. Before the memory erasure, they would choose A for both days: they would still meet their other-day self and enjoy the more pleasant coffee shop.
Taylor applies this distinction between a decision made within a situation and a plan chosen in advance to a variant of Newcomb’s problem. The predictor Omega runs a program that exactly reproduces the agent’s reasoning and is presented with the same situation. If the simulation chooses one box, Omega puts a million dollars in the opaque box; the transparent box always contains a thousand dollars. The original agent is then offered the same two boxes.
If the agent regards the simulation as a conscious observer, they include it among their possible selves. Choosing one box within the simulation then determines the payout Omega will prepare for the original. If the agent regards a running program as a conscious observer only when it runs on a physical substrate they consider suitable, they choose two boxes as the original. Taylor connects this difference to computational functionalism, the view that a functionally exact simulation can be conscious regardless of its physical substrate. This criterion allows the agent to follow the policy chosen before memory erasure.
In Max Hodak’s discussion of digital continuation, the issue concerns what happens after a brain scan: whether the original person’s experience continues in a computer replica. In Taylor’s account, the status of the simulated execution affects an earlier decision: when choosing, the agent must decide whether to count that execution as one of their possible observer perspectives.