This last part carries a lot of weight; a simulacrum, when dormant in the superposition from which it can be sampled, is nonexistent. A simulacrum only exists during the discrete processing event which correlates with its sampling.
There seems to me to be a sensible view on which a simulacrum exists to the extent that computations relevant to making decisions on its behalf are carried out, regardless of what the token sampler chooses. This would suggest that there could conceivably be vast numbers of different simulacra instantiated even in a single forward pass.
One odd upshot of requiring the token sampler is that in contexts in which no tokens get sampled (prefill, training) you can get all of the same model computations but have no simulacra at all.
There seems to me to be a sensible view on which a simulacrum exists to the extent that computations relevant to making decisions on its behalf are carried out, regardless of what the token sampler chooses. This would suggest that there could conceivably be vast numbers of different simulacra instantiated even in a single forward pass.
One odd upshot of requiring the token sampler is that in contexts in which no tokens get sampled (prefill, training) you can get all of the same model computations but have no simulacra at all.