Chapter 35
Entropy, and the Cost of Staying Level
Chapter 26 gave the learner a goal it can state about itself, and a reading of its own distance from that goal. It did not say what the goal costs, and it could not: the currency had not been built yet.
This part has now built it, and this chapter spends it. Homeostasis is not free, and it is not free for a reason that has been sitting in this part since the conservation chapter: a cost-accounted theory is frictionful by construction, and a learner holding a level against friction is paying a bill in exactly the currency Chapter 31 priced. What follows is short, because the work was done in the chapters above. Its job is to collect it, and to hand the result back to a learner who was given a goal several parts ago and no way to pay for it.
35.1 Dissipation is where information leaves
Chapter 31 established that a cost-accounted theory is frictionful by design and therefore lives in the dissipative regime: every lossy enzyme that fires strictly decreases the free energy, and the virtual token is a Lyapunov function rather than a conserved charge.
Read that result backwards. A frictionless theory is one whose enzymes are exactly value-preserving, and by Theorem 15.1 such a theory has an exact potential, which is to say it has enough structure to say where every unit of value came from. A frictionful theory does not. The dissipation \(\theta\) accumulated around a cycle is precisely the amount of provenance the theory cannot reconstruct: value went somewhere, and the ledger does not record where.
That is an information-theoretic statement wearing economic clothing, and Chapter 31 already turns it into one, by way of the Landauer bound: the erasure a computation performs is bounded below by the holonomy of the energy form. Erasure is forgetting which of several possible predecessors a state had. Dissipation is the price of forgetting. The two thresholds coincide, which is the most satisfying result in that chapter and is also, for present purposes, a signpost.
35.2 What a setpoint costs
Now put the bathtub of Chapter 26 in that setting.
A learner holding \(\vect{A}\) near \(\setpt\) is not sitting still. It is running an assay on its own account, comparing the result to a held formula, and adjusting an outflow it does not fully control against an inflow it does not control at all. Every one of those operations is a metered rewrite, and by Definition 30.6 every metered rewrite is drawn from the same account it is measuring.
Let \(\Lrn\) be an internalizing learner in the sense of Definition 26.1, and let \(\kappa_{\mathrm{reg}}\) be the per-cycle cost of the assay, the comparison, and the corrective rewrite. Then \(\kappa_{\mathrm{reg}}\) is charged to \(\vect{A}\) itself. Consequently the inflow required to hold \(\vect{A} = \setpt\) is strictly greater, by \(\kappa_{\mathrm{reg}}\) per cycle, than the inflow required to sustain a non-internalizing learner with the same rewrite load.
The account is vectorial and single: there is no second account from which regulation could be funded, since by Remark 31.8 an account merely adjacent to a process is ambient authority, which the framework has already refused. So the assay of §26.3 draws on \(\vect{A}\), and the drift it reports is a drift of the quantity that just paid for the report. The stated inequality is the difference of the two flow balances.
Proposition 35.1 is the reason Chapter 26 could not simply assert that internalizing is better. A learner that watches its own reservoir is, by that act, draining it. The watching is worth doing only when what it buys — anticipation, which is to say correction begun before the level has actually fallen — exceeds what it costs, and whether it does is a fact about the environment rather than about the learner.
This is also the sharpest available statement of what suffering is for, and it is unsentimental. Suffering is the readout of a meter the organism is paying to run. An organism that felt nothing would be cheaper, and in a placid world it would win.
35.3 The two erasures, and what is left over
There are in fact two distinct erasures in play, and they are not the same map [108]. One forgets the history: the history monad’s counit, which discards the log and keeps the state. The other forgets the cost: the cost monad’s counit, which discards the meter and keeps the computation. Neither factors through the other, and the failure to factor is graded — there is a lattice of intermediate laxities between them, each corresponding to a theory that remembers some of one and some of the other.
The gap between the two erasures is not a technical annoyance. It is a supply of positions. A theory sitting strictly between them knows something about its own past that it cannot express as a cost and something about its own costs that it cannot express as a history, and a computation with information of that shape is exactly a computation with something to decide.
An internalizing learner sits in that gap by construction, which is worth saying plainly because it was not obvious that anything did.
Its held setpoint is cost information: \(\setpt\) is a level in the account, and \(\phi_{\mathrm{home}}\) is a formula about the meter. Its drift is history information: a drift is a difference taken across time, and reading one requires having kept what the level was. Neither reduces to the other. A learner that kept only the cost would know its level and not know whether it was falling; a learner that kept only the history would know the shape of its trace and not what it could afford. Suffering, as Chapter 26 defines it, is precisely the datum that requires both — which is why it could not be a component of the account, and had to be a formula about one.
That leaves an obligation the Turn does not discharge. If an internalizing learner is a computation with something to decide, then something must do the deciding, and nothing constructed so far says what. The reduction relation offers continuations and is silent about which occurs.
The Prestige takes that silence seriously, and argues that it is where the only genuinely non-Turing thing in this book is located. Readers who want that argument now will find it in Chapter 57. Readers content to let the Turn finish its own business should carry forward only this: the learner of Chapter 26 has been given, for the first time, information of a shape that makes a choice meaningful, and it was homeostasis that gave it.