10. The Last Economy

The fall of ΠH was introduced as a cost-curve observation. Its possibility and its terminus are consequences of the structure—once you separate the information cost the kinematics carries from the thermodynamic cost of running an agent on a substrate.

Two layers

The choice object ρ ∝ µ e^{V/τ} is substrate-independent—same form in neurons or silicon. That is why the identity with machine learning is contentful. What it does not contain is the energy bill of instantiating the agent.

Temperature τ in the kinematics is an information price (value per nat of departure from the reference)—not joules per computation. The bridge is an effective temperature combining both, to first order:

τ_eff ≈ τ_info + p × c_sub

where c_sub is substrate energy cost per information and p converts energy into value units.

For a human, c_sub is metabolic and biologically floored (a brain ~20 W; sustained muscle work higher). No diet pushes joules-per-operation to the Landauer limit. For a machine, τ_info is tunable and c_sub falls toward physical floors with hardware generations.

Metabolic Rift (book name): labour enters whose cost has no metabolic floor.

Landauer in one breath

Erasing one bit against a fair two-state reference raises KL by ln 2 nats. Complexity cost charges τ per nat → cost τ ln 2. When computation is physical at temperature T and the reference is the physical Gibbs state, τ becomes k_B T and you recover Landauer’s bound: k_B T ln 2. The thermodynamic floor on machine instantiation is the same complexity cost the chain already forced—read at the temperature where information and energy currencies meet.

Cost and capability

A cheap agent that cannot do the task does not compete. Capability is how close realised value gets to available value. Surplus on a task:

Π(agent, task) = capability − (fixed costs + τ_eff × information used)

ΠH < 0 is this going negative for the human. Displacement can happen because the machine is more capable or merely cheaper—politically and economically different stories.

Robotics and the share of value

Value is defined over configurations of the world, and most economic output is physical. Digital intelligence collapses cognition cost but cannot move matter by itself; it captures value mainly where action is already informational. Embodied intelligence bids for the whole state space.

Split kinetic cost:

K = K_compute + K_actuation

This wave collapses K_compute first (lower engineering floor), leaving actuation as the lagging human reservoir—inverting prior automation waves that mechanised muscle and left cognition. Moravec’s paradox in these variables: the human–machine gap is large in actuation quality η_act and small in cognitive energy cost.

Humans and humanoids in one equation

Because both are the same object on the same physical state space, write them as parameter vectors:

Coordinate	Human	Humanoid
Info price τ_info	Neural bounds	Tunable
Cognitive energy cost	Metabolic, fixed	Falling toward Landauer
Actuation energy cost	Metabolic, fixed	Falling
Actuation quality	Evolved, high	Rising slowly
Throughput / fatigue	Fixed	Tunable
Cognitive capability	High, relatively fixed	Rising, can pass human

The human is a fixed point in parameter space. The humanoid is a moving point sweeping across it. Tasks flip when the humanoid becomes at least as capable and at least as cheap. Order: routine cognition first; high-dexterity physical work last—held on capability premium even after cost undercutting.

No fixed point for the human

It is tempting to invent a coordinate machines can never occupy. A first-principles derivation cannot supply one without smuggling comfort. Every cost and capability coordinate is one the humanoid can approach or pass. The framework posits no term that keeps humans in the production story because of species identity.

Open content question left explicit: is some value genuinely agent-indexed—constituted by who produces it—so substitution changes the value, not only the cost? Magnitude may be large, small, or zero. If a residual human economy exists, it lives there; the structure does not guarantee its size.

Comparative advantage does not save negative surplus: it allocates among agents still worth employing. Wage adjustment hits a metabolic floor; once machine full cost falls below subsistence-linked wages, market-clearing employment that also provisions the worker can fail. Work may continue for meaning or transfers; it may stop carrying consumption.

Task-based labour economics (displacement vs reinstatement) is recovered as the sweep; the framework adds when reinstatement fails: the reservoir of uncrossed coordinates is consumed in order (dexterity, then whatever agent-indexed value communities insist on).

The factor sequence ends at doxa

Historical ages—land, then labour, then capital, then intelligence—read as successive binding constraints. With τ_eff and K split, the equation names remaining reservoirs as terms peel:

1. Land, labour, capital — constraints on means of action
2. Intelligence — collapse of K_compute
3. Robotics — collapse of K_actuation
4. Free energy / power / cooling — Landauer margin when action is cheap
5. V — the objective no amount of cheap cognition generates for itself
6. µ — the reference that makes any V legible and coordinable

µ is the outermost term of the structure. Within this object there is no further external factor for a new inversion to move to. Contest turns inward: governance of the reference—the expectational commons that can be enclosed, captured, or depleted. Whoever tends doxa when cognition, action, and energy are abundant holds the structural position landowners and capital once held—except a commons is governed more than owned.

That economy is last not as apocalypse branding, but because peeling stops at the reference; past it is only the reference’s own politics.

Next: the two ways the last inversion fails—broken provisioning chains, and capture or collapse of doxa.