Why a bounded civilization would build a multiverse
If your own discovery space is finite and you need novelty, you have two options: get a better theory, or get data from somewhere with different laws. The second option is what a novelty-harvesting multiverse is for, and the argument for building one is economic rather than philosophical.
The reasoning runs like this. A civilization on the flat part of its discovery curve still has problems it cannot solve, and the solutions to those problems exist in principle. They are just unreachable with the tools and concepts that civilization has, because reaching them requires a conceptual move nobody can find. Novelty, in this framing, is a resource, and resource-limited systems that need more of a resource go looking for a new source.
- Also see: The Cosmic Web: Mapping the Universe’s Invisible Architecture
- Also see: Why Build a Multiverse?
How this differs from the ancestor simulation idea
Bostrom’s simulation argument asks whether we might be running on someone else’s hardware, and the simulations it imagines are reconstructions — ancestor simulations, history replayed. That framing has an odd feature: if the point is to preserve history, the simulator is not after anything it does not already have.
The version that interests me inverts the purpose. A simulation whose output is the point would not be a re-enactment; it would be an instrument. You would run it to produce discoveries you could not produce yourself, and you would care about the novelty coming out of it, not about fidelity to some original.
Under that reading, the interesting questions stop being metaphysical and become engineering ones: what kind of environment generates useful novelty, how much does it cost, and what happens to the machine’s productivity over time.
The formal version
There is a theorem in this programme that makes the distinction precise, and it is unkind to lazy versions of the idea.
Suppose a base civilization harvests from a set of branches. The knowledge it can extract is the joint description length of what it collects. If every branch runs the same physics, then adding branches is nearly worthless: N copies of one output compress to roughly one output plus a few bits for indexing, so a homogeneous multiverse returns order log N bits no matter how many branches you run. Parallelism buys latency, not knowledge.
If the branches differ in their laws, the returns are additive. Each distinct physics is a distinct structure that has to be described separately, and the total harvest scales with the number of distinguishable physics rather than with the branch count.
So the machine has to be heterogeneous, and that is a conclusion rather than a design preference. The corollary is sharper still: a harvesting architecture that varies only initial conditions has provably zero asymptotic return, and any reported success from such an architecture is an accounting artefact.
Why “novelty” and not “intelligence”
I keep coming back to this because it is the part most easily mistaken for science fiction. Nothing here requires the simulated beings to be conscious, or to consent, or to enjoy it. What the architecture needs is unpredictable state transitions — genuine divergence from what the base already knows. Creativity is an efficient way to produce that. It is not the only conceivable way, and the framework does not actually depend on sentience being special.
That is a colder reading of the idea than most versions, and I think it is the honest one.