The Forever Machine: what it would take to run one

A novelty-harvesting multiverse is usually discussed as a philosophical possibility. It is more useful to treat it as a specification, because a specification has numbers in it and numbers can be wrong.

Here is the specification I worked out, in the order the numbers appear.

What kind of universe

Simulated environments are not interchangeable. Ordering them by information-generating capacity per unit of compute gives a rough spectrum: trivial automata at the cheap end, then simple chemistry, then agent-based ecosystems, then minimal physics with a biosphere, then an Earth-analogue with unpredictable agents, then a solar system, then a full cosmological simulation.

The first result worth noting is the economics. Going from “physics and a biosphere” to “physics, a biosphere and sentient agents” multiplies the novelty output by around a hundred for a cost increase of about twenty orders of magnitude. That is a favourable trade, which is the formal reason a machine of this kind would be populated with agents capable of genuine unpredictability: they are the most efficient novelty generators per unit of substrate.

The second result is that the expensive end is not needed. Going from the agent-populated case to a full cosmological simulation adds novelty, but the cost increase is catastrophic relative to the gain. A rational operator runs enormous numbers of cheap agent-populated universes rather than a handful of cosmological ones.

What it would cost

The hardware estimate scales badly, which is the point. Full-resolution simulation of an agent-populated universe at the required fidelity lands at roughly 10^47 current US data-centre equivalents. Adding selective activation — rendering only the high-novelty regions of each branch at full resolution, interpolating the rest, on the grounds that most of any universe is empty space and most of any planet is geologically stable — brings it to around 10^69 equivalents. Still impossible.

I want to be careful about what that number means. It is not a refutation of the idea, because the idea does not require the machine to be built by us, out of our hardware, under our physics. It is a constraint on the substrate question: whatever is doing the computing is not doing it the way we would.

Why it has to be orchestrated

A monolithic system cannot do this job, and the reason is structural rather than a matter of engineering taste.

Checking each branch against every other branch for novelty is quadratic. With full state dumps that cost is hopeless. With typed, compressed signatures — structured descriptions of state transitions above a threshold, compared against a registry rather than against each other — it becomes linear in the number of interesting events rather than quadratic in branches.

That is the same argument that applies to networks of small specialised models coordinated by a routing layer: you do not scale by making one system larger, you scale by making many systems smaller and investing in the coordination fabric between them. The architecture of the Forever Machine and the architecture of a distributed model ensemble are the same shape for the same reason.

The honest summary

Under known physics the machine is not buildable. The estimate is the useful part: it tells you that any substrate capable of doing it is unlike anything we can currently construct, which narrows the space of serious proposals rather than widening it.