Why we studied the industrial process
It started with a deliberately basic exercise: walk through the entire industrial process. How are raw materials extracted, refined and manufactured? How do concrete and electricity get made and moved to where they're needed? It seemed like a detour. It turned out to be the foundation for everything else.
The reason is simple. If an optimizer somehow accumulated ten trillion dollars, that would matter only insofar as it could turn into fabs, data centres, power generation, copper, steel, land, transport, expertise, laboratories and manufacturing capacity. Those things break down into chains of physical transformation. Money, contracts and software sit on top of that substrate. Any end-game optimizer would have to understand how the world works as a whole, not one sector in isolation.
Six physical layers
A useful simplification of industrial civilization:
| Layer | What it does | Examples |
|---|---|---|
| 1. Resource extraction | Get atoms out of nature | Mining, drilling, quarrying, forestry, agriculture |
| 2. Bulk processing | Separate and purify | Ore concentration, smelting, oil refining |
| 3. Primary materials | Make standardized substances | Steel, cement, aluminium, plastics, glass, chemicals |
| 4. Components | Shape materials into standard parts | Bearings, pipes, beams, motors, chips, wire |
| 5. Capital systems | Build productive systems | Factories, power plants, ships, machines |
| 6. Goods and services | Satisfy human wants | Food, housing, transport, computing, medicine |
Then the key feedback loop appears: layer 5 produces the equipment that increases the productivity of layers 1 to 5. A steel mill makes steel. Some steel becomes excavators, which expand mines, which supply more ore, which becomes more steel. Some becomes turbines, which generate electricity, which runs factories, which make better machine tools, which make better machines. That's the physical mechanism behind capital accumulation. The system becomes able to make more and more sophisticated transformations of matter and energy.
Seven levels of optimization
Most AI deployment today sits at the bottom of a ladder that goes much higher. Each rung treats more of the surrounding system as changeable rather than fixed:
- Task: inspect this weld, predict this failure, adjust this furnace. The rest of the system is fixed.
- Process: jointly control a whole process, such as ore in and steel out, choosing operating policies rather than isolated decisions.
- Factory: ask why it's made this way at all. Change layout, tooling and sequence, moving from control to design.
- Supply chain: the factory boundary becomes artificial. Optimize mine → refinery → material → factory → logistics → customer as one system.
- Capital: not "how should I run my assets?" but "which assets should exist?" Allocate productive capacity itself.
- Research: production is limited by available technology, so search for new materials and processes in closed experimental loops. Self-driving laboratories already do pieces of this.
- Optimizer: improving the optimizer improves everything below it. Allocate resources to building a better optimizer.
Give a system the narrow goal "produce electricity as cheaply and reliably as possible" and a sufficiently general version climbs the ladder by itself. Better turbine geometry. Less copper. Predictive maintenance. New generation sites. Better alloys, and finding them. More efficient mining of their ingredients. Building its own mining machinery. A better AI to find all of this faster. "Optimize electricity" has expanded into "optimize a large part of the productive economy." That's the non-cartoon version of the paperclip argument.
Recursive intelligence improvement is recursive industrial improvement
Hypothesis
"Self-improving AI" usually brings to mind code rewriting itself. There's another loop that may matter far more:
The recursion runs through the physical economy. It doesn't need a model literally editing its own weights. Engineered systems like AlphaEvolve already show a bounded version: an evolutionary coding system found improvements to data-centre scheduling and to components used in training the very models it runs on. That links search to its own supporting infrastructure. It isn't a demonstration of an indefinitely accelerating loop.
The loop has two distinct parts that are easy to blur. One is improving the learning and search process itself: better representations, experiments and algorithms. The other is increasing the resources available to it: compute, energy, data, tools. They reinforce each other, but only if the improvements outweigh what they cost, and one agent's share of resources growing is different from the whole system's capacity growing.
Capital is a claim, not the thing
A generalized optimizer might reasonably start in capital markets. There, an information advantage becomes financial return, and capital is a generalized claim on society's productive capacity. Capital can turn into compute, energy, land, labour, machinery, companies and research without first owning a single mine.
But capital isn't physical power. It's an institutionalized claim on physical power. A billion dollars is useful because banks, courts, governments, firms and workers recognize what it entitles you to. Finance rests on institutions, which rest on energy, infrastructure and industry. These don't form a neat vertical stack. They form a graph. Finance shapes industry, industry produces compute, compute improves AI, AI improves financial prediction, governments set financial rules, energy powers everything, and research improves energy. It's a recursive control network.
Several kinds of capacity at once
An optimizer with a civilization-scale horizon would have reason to model several distinct kinds of capacity, because each one limits which future states it can reach:
- Epistemic: data, sensors, models, experiments, scientific knowledge. You can't optimize consequences you can't predict.
- Computational: chips, data centres, electricity, algorithms.
- Financial: a fast routing mechanism for productive capacity. You can move a billion dollars in seconds; you can't move a steel mill.
- Industrial: the ability to transform matter, which decides which plans can actually be carried out.
- Energetic: almost every physical transformation needs energy. Usable energy is close to the capacity to cause physical change.
- Institutional: ownership, contracts, regulation and permission to operate define the available action space. Working inside stable institutions often gives more long-run optionality, at lower cost, than trying to control them.
- Security: the systems societies keep for when voluntary coordination breaks down. They belong in any serious world model because they sharply constrain which states are reachable. Our interest here is purely analytical.
Not exponential forever
A feedback loop can look exponential for a while. The physical world imposes limits: energy, ore grades, land, capital, lead times, heat dissipation, information, coordination costs and diminishing returns. History looks less like one exponential and more like successive S-curves. One paradigm saturates, and a new innovation opens up more room. Economic growth theory, from Romer's endogenous technological change to Kremer's long-run view of population and technology, studies exactly this interplay between ideas and the resources that carry them.
The bottleneck is legibility
Language models improved quickly partly because information was already digital. A PDF or a spreadsheet is easy to hand to a model. "Try a new alloy" is not. It needs material handling, robotics, instruments, safety systems, time, energy and compliance, across equipment from dozens of manufacturers and decades of legacy protocols. The physical world is far less API-compatible than the digital one.
That suggests a reframing. The most important step may not be "build an AI that runs civilization." It may be making parts of the physical economy legible and controllable to machine intelligence, the way shipping containers standardized freight or HTTP standardized information. And once that happens, prices could become the signals that coordinate many specialized machine agents without a central planner. That's an artificial economy, and it's very close to the multi-agent systems Rubedo began with.
An open question we're sitting with
Much current AI attention goes to applications high up the stack, such as summaries, copy and chat. Those can be valuable, but they optimize human information work around an otherwise largely unchanged physical economy. A small improvement far upstream, in steel or energy or materials, spreads through everything built on top of it. So we keep asking:
Where in the physical economy is there an open optimization loop (observe, model, decide, act, measure, learn) that machine intelligence could close?