Identical universes are worthless

If the multiverse is real and the Forever Machine wants to harvest discovery from it, the first engineering question is: which branches are worth running? The harvesting theorem answers it from the wrong direction, by proving that the default answer, all of them, is worthless. This post walks the short proof, because its conclusion drives every architectural choice in the programme.

The claim

A homogeneous multiverse returns only order log N bits of total novelty, no matter how many branches N you run. Heterogeneity is not a preference. It is mandatory: without it, the machine’s output does not scale with its cost.

The proof

The argument is compression, applied to branches instead of data. Suppose the branches are identical: same laws, same initial conditions, same rendering. Then the collection of all N branches has one description: the description of one branch, plus the statement run it N times. The statement costs log N bits. That is the entire novelty content of N identical universes: log N, not N times anything.

Now suppose the branches differ but their differences are noisy, unstructured: each branch has independent random deviations. Randomness does not compress, so the total description is roughly N times the per-branch cost. The collection is expensive to describe, but expense is not novelty: the per-branch deviations are incompressible noise, and by the novelty-is-compression definition, noise has zero drops. Expensive and worthless.

The valuable middle: branches whose differences are structured, different constants, different symmetry realisations, different low-energy physics, but generated from a shared substrate. The collection’s description is: one substrate, plus the branch parameters, plus the renderings’ deviations. The branch parameters are a set of structured, compressible data, and each branch’s rendering, run locally, is a genuine new domain whose plateau, by the plateau theorem, has its own finite budget of unmined structure. Novelty scales with the number of distinct regimes, weighted by how different their physics is, and the theorem’s formal statement puts a logarithmic lower bound on the homogeneous case and a linear-in-branches bound on the structured case.

The simulation in the paper makes it concrete: 64 identical branches, fed through the same compression pipeline as real data, compress to the information content of about one branch, ratio 0.016, while 64 branches with distinct physical laws compress to 0.903 of their naive size, because their differences carry structured content. The two ratios are the theorem in decimal form.

Why this kills the obvious designs

The worthless case is not a strawman; it is the default. A multiverse simulator’s natural instinct is fidelity: make the branches as much like our universe as possible. The theorem says fidelity across many branches is the one design that returns nothing, because agreement is compression’s cheapest raw material. The designs that survive are the ones that maximise structured difference: vary the constants along a principled axis, vary the symmetry-breaking patterns, vary the low-energy emergent physics, and keep the substrate common so the differences stay compressible. Identical universes are worthless. The machine’s entire value lives in the distance between branches, and the next posts price that distance.