If our universe is more likely to be a simulation, a reconstruction of the past by our descendants than the base level of reality; then that reconstruction is likely to be imperfect, based mainly on surviving records (and memories of anyone whose brain survives intact long enough for upload-style scanning). Therefore, if somebody precommits to only leave behind records which correspond to particular events… then it seems plausible (to within the bounds of ‘the brain is a quantum computer’ levels of plausibility) that those events become more likely to be experienced.
How does this cope with simulation dynamics? They may seed the simulation with lots of old real-world data, but once it starts running, there are going to be tons of micro and macro divergences which make data dated after the simulation start less than useful. (eg. imagine the simulation run starts at a point in 2000; by the time it reaches its version of today, it likely will have diverged significantly—it doesn’t matter if you write that you had dinner with your girlfriend because in the simulation you just barely missed meeting her a year ago etc.)
Suppose we do rejection sampling. Then the universe restarts at 2000 every time a datapoint is violated. The subjective experience of entities inside such a simulation would be best described as random death with occasional survival if some arbitrary criteria are met. Writing (long-lasting) believable false reports will alter the criteria to include those reports somehow being written.
MCMC algorithm is more interesting. It introduces a strange contortion of time; we generate an initially random space-time with the known datapoints clamped, and then re-sample the unknown bits for a long time to get a good posterior distribution over the possibilities. Real time is a sometimes-nonsensical meandering through possibility-space. Time as experienced “in the simulation” is relatively normal, but it is interesting/boring to argue about whether entities living in such a simulation have experiences in a meaningful sense. Causality doesn’t work at all as it should, but it will appear to work roughly as it should most of the time. Events will conspire to bring about a random assortment of facts which are the clamped values, but the better-quality samples will make the conspiracies look like true chance.
Metropolis-hastings is an even more advanced technique, but I don’t think there is anything special about the subjective experience if the simulated entities in MH as compared to basic MCMC.
How does this cope with simulation dynamics? They may seed the simulation with lots of old real-world data, but once it starts running, there are going to be tons of micro and macro divergences which make data dated after the simulation start less than useful. (eg. imagine the simulation run starts at a point in 2000; by the time it reaches its version of today, it likely will have diverged significantly—it doesn’t matter if you write that you had dinner with your girlfriend because in the simulation you just barely missed meeting her a year ago etc.)
Suppose we do rejection sampling. Then the universe restarts at 2000 every time a datapoint is violated. The subjective experience of entities inside such a simulation would be best described as random death with occasional survival if some arbitrary criteria are met. Writing (long-lasting) believable false reports will alter the criteria to include those reports somehow being written.
MCMC algorithm is more interesting. It introduces a strange contortion of time; we generate an initially random space-time with the known datapoints clamped, and then re-sample the unknown bits for a long time to get a good posterior distribution over the possibilities. Real time is a sometimes-nonsensical meandering through possibility-space. Time as experienced “in the simulation” is relatively normal, but it is interesting/boring to argue about whether entities living in such a simulation have experiences in a meaningful sense. Causality doesn’t work at all as it should, but it will appear to work roughly as it should most of the time. Events will conspire to bring about a random assortment of facts which are the clamped values, but the better-quality samples will make the conspiracies look like true chance.
Metropolis-hastings is an even more advanced technique, but I don’t think there is anything special about the subjective experience if the simulated entities in MH as compared to basic MCMC.