Intelligent Economics · Emad Mostaque · 2026
People, markets and machines choose by one law.
Every choice starts from what is expected and leans toward what is valued. How far it leans depends on what leaving the familiar costs. The same law is the exact target of the method used to align today’s AI models. This is the story of that law, and of the economy it describes as machines take on more of the work.
- μwhat is expected
- Vwhat is valued
- τwhat leaving costs
01 · The question
Wages carried the economy home. That link is dissolving.
The labour a person did became, through wages, the consumption that sustained their household. For any service, the comparison is the value of its delivery against its full cost, supervision, reliability and integration included. Where a machine supplies it more cheaply at comparable value, the machine has the advantage. The paper writes the difference as ΔH: the surplus from human delivery less that from its machine substitute.
Which tasks cross that threshold, and how quickly, are empirical questions. The structural one is narrower: if a community relies on wages to provision its members, substitution can weaken that channel even while productive capacity grows.
The wage channel
illustrative · no dataDrag the share up and watch which streams thicken
“Whose flourishing does the system serve when an increasing share of its output need not be produced by the people it is meant to sustain?”Intelligent Economics, §1To answer it, start with how any choice gets made ↓
02 · The law
Three ingredients make every choice.
Start from what is expected.
Before anyone weighs anything there is a background: habit, culture, what the people around you do. Some options feel natural. Some are not on the menu at all.
Each point here is one choice. Their cloud is the reference, μ. The dark strip on the right is what this background excludes.
Value pulls.
Now give every option a value, V. Here it rises steadily to the right, and it is highest in the strip nobody considers.
Leaving the familiar has a price.
Departing from habit costs attention, effort and risk. Call that price τ. When it is high, choices stay close to the reference. As it falls, they lean toward value.
Gold points gained weight from value. Blue points lost it.
ρ* ∝ μ · eV/τ
The share of choices on any option is its weight in the background, multiplied by an exponential of its value, at the going price.
Look at the right edge. However cheap departure gets, nothing flows into what the background excludes.
For the curious
Among all ways of choosing with the same average value, ρ* is the one closest to μ in Kullback–Leibler divergence, measured in nats. τ converts nats into units of value. Equivalently, ρ* maximises expected value minus τ times that divergence.
The theorem of §3, built link by link in §3.2: four links run from the premise to the dynamics. The canonical Langevin relaxation carries a distribution to the law and has it as its unique stationary distribution; positive temperature alone does not specify a dynamics.
03 · The reference
Same values. Different worlds.
Hold every value fixed. Change only the background people start from. The choices change, sometimes completely.
The paper’s one identification is that the shared, social part of that background is a community’s doxa: what goes without saying. Institutions are concentrations of it, kept up through repeated practice, and treating that background as something produced and maintained is the paper’s central proposal.
“Doxa is what nobody argues about because the disagreement would not register.”Intelligent Economics, §4.2
Five options, two towns
invented valuesOrthodoxy is what a community consciously affirms. Doxa is what the affirmation rests on.
Preferences are what people want. Doxa is the background that makes the wanting feel sensible.
Constraints are what is physically or institutionally possible. Doxa is what makes them feel binding rather than arbitrary.
An option with a small weight can still be chosen when its value is high enough. An option the reference excludes never is, at any value. Only changing the background brings it back.
04 · One law, five names
The method used to align AI solves the same equation.
A background, a value, a price.
μ is the reference, whose social part the paper calls doxa. V is the value function. τ is the information price.
Attention has a cost.
In rational inattention, μ is the marginal choice law, V the payoff, and τ the cost of attention. The answer is a weighted logit.
A policy that stays near its habits.
In soft RL and control, μ is a reference policy, V the reward, and τ the entropy weight. The answer is the soft-optimal policy.
A generative prior over what to do.
In active inference, μ is the policy prior, V the extrinsic value, and τ an inverse precision. The answer is the extrinsic policy rule.
Reward the model, charge it for straying.
μ is the reference checkpoint, often a supervised fine-tuned model; V the reward model; τ the KL coefficient β. The unrestricted optimum of this objective is the choice law itself.
So a reference policy’s distribution over outputs is the machine counterpart of a community’s doxa. The paper reads the successive stages of training as layered doxa engineering: what each stage learns, retains and makes available.
For the curious
§3.5 and §4.5: the known solution of the RLHF objective (Rafailov et al., 2023), the policy proportional to its reference times exponentiated reward, is ρ* ∝ μ eV/τ with μ the reference checkpoint, V the reward model and τ the KL coefficient. The paper’s boxed results in §4.5: RLHF gives the reference an operational role, and the reference is an alignment target. §8.4 draws the consequence: alignment is doxic alignment.
05 · Alignment
Same reward. Different behaviour.
A system’s behaviour has two sources: what it values, and the background it starts from. Fix the reward and the price, and the background still shapes what the system notices, retrieves and offers.
So a model can score well on a test of its values and still behave differently from another model with the same score. Checking the reward alone does not show that choices suit a community.
For the curious
§8.4: where the smooth score is defined, ∇log ρ = ∇V/τ + ∇log μ, so specifying the value term alone does not specify the choice distribution. An AI can perform well on a chosen value benchmark while its reference shapes which possibilities it notices, retrieves or makes available outside that benchmark. When AI systems maintain more of a community’s background, a training or deployment choice travels well beyond one model’s output: the concern is systemic drift or fragmentation in the expectations through which people act together.
One reward, two starting models
computed live- 0.00
- model A reward
- 0.00
- model B reward
- 0%
- choices in common
06 · The economics we already have
Textbook economics lives at the cold edge.
With a price on departure, choices spread.
At a moderate price, people mostly do what value suggests, and some still do what habit suggests. Demand curves, logit choice and quantal response in games live here.
Lower the price and the law sharpens.
Departure gets cheap. Choice crowds toward the best option the background allows, against the wall of what it excludes.
At zero, the perfect optimiser.
All choice lands on one best option. This is the rational agent of the textbook, and Nash equilibrium appears as the cold limit of quantal response.
The background still decides which options were ever in the running.
For the curious
§3.6 and §5: value maximisation is the zero-temperature boundary of the choice law, and positive temperature with a consequential reference gives behavioural economics, rational inattention and bounded rationality a common language. Table 3 collects eleven classical results. Some are recovered exactly within a specified model, such as quasi-linear demand and Slutsky symmetry; others identify a field’s object with one the structure carries, and the paper’s test for those is transfer: evidence, algorithms and comparative statics must carry across with their assumptions intact.
- Demand and Slutsky symmetry
- Multinomial logit
- Exponential discounting
- Black–Scholes
- Condorcet’s jury theorem
- Aumann agreement
- Clearing and welfare
- Landauer’s bound
- Euler equation
- Expectations and learning
- Nash equilibrium
07 · The market
The market price sits at the bottom of a bowl.
Buyers and sellers draw on one limited stock. In the paper’s market model, too much demand is the downhill slope of a single bowl-shaped function of the price.
So clearing a market is rolling to the bottom. Let the price go and it settles where supply meets demand.
For the curious
§6.1: aggregate excess demand is the negative gradient of Φ(p) = Σ τa log Za(p) + p·ω, which is smooth and convex, its Hessian a sum of covariances. On finite menus or compact supports, with the resources in the interior of what is achievable, a minimiser exists and continuous tâtonnement, ṗ = −∇Φ, converges to it. Supply slopes up by the same convexity that makes demand slope down. The clearing allocation is efficient against the given references, and only against them: “Efficiency inherits the reference.”
The demand and supply drawn here are smooth examples.
Release the price
example curvesOr drag the ball along the bowl
08 · The labour link
When the machine is cheaper and as able, the wage hits a floor.
Work feeds production. Wages feed homes.
The labour a person did became, through wages, the consumption that sustained their household. It was never the only route to provision, but it connected production to the lives of a large part of the community.
Machines join, on two axes.
Viability on a task is capability minus cost. A person can be displaced because the machine is more capable, or undercut because it is cheaper, and the two need not happen in the same order.
Where actuation gets cheap before dexterity is reliable, a machine can be cheaper but less capable at a physical task, and human work keeps a dexterity premium for as long as that lasts.
Nothing in the law reserves a task for people.
The law supplies no coordinate a machine can never occupy, and no guarantee that machines acquire every capability. Relationships, trust, legitimacy and the identity of a producer can be part of what is valued; then they belong in V. Physical limits, scarce capacity and complementarities can preserve human roles too.
The question it locates: which human contributions remain valuable, feasible and paid as machine alternatives improve?
Work may continue. It no longer carries consumption.
A cost-minimising employer buys the human service only at a wage below the machine’s delivered price. When every task open to a worker can be bought from a replicable machine service for less than the worker’s subsistence over the time the task takes, no wage such an employer will pay provisions the worker. Work may continue; it no longer carries consumption.
For the curious
§7.5: the floor is a statement in real terms. Human and machine draw on the same goods; when on every listed task the machine’s input bundle is worth less than the share of the household’s needs that the task’s time requires, a wage that covers the person costs more than the machine it replaces. A household provisioned by other means can still add value at a lower wage; what ends is work as the thing that pays for subsistence.
09 · The provisioning floor
When a machine costs less than a home needs, wages stop provisioning.
A cost-minimising employer will not pay a person more than a machine that does the job as well. If the machine’s full cost per hour falls below the wage a household needs, then no wage both keeps the job and covers the household.
The person can keep working. The work just cannot carry the household.
For the curious
§7.5, the provisioning floor: a cost-minimising employer buys the human service on a task only at a wage below the machine’s delivered price less its other costs, so no schedule of such tasks earns more than the best of those caps. “If that cap falls short of what the household’s bundle costs, no wage a cost-minimising employer will pay on these tasks finances that provision. Work may continue; it no longer carries consumption.”
The right-hand chart applies the paper’s employer’s-odds proposition to one employer choosing between a person and a machine for the same output: with an even reference, the log-odds of hiring the person are the cost saving divided by the employer’s price of departure τ. It is drawn for this page.
One household, one best job
illustrative numbers10 · The two failures
Two failures, and the contest beneath them.
How do the gains reach the community?
The chain from labour to consumption is an institution, a set of maintained expectations. When productive capability shifts to machines, the institution still covers the people supplying labour; it does not by itself distribute the gains from capital to everyone whose welfare matters. Markets can pass those gains on through lower prices and broad ownership, but guarantee neither.
A basic income repairs one part of the link. It can sustain consumption without reproducing the other things work supplied: identity, community, recognition and structure. The paper calls for reconstruction of the institutional layer, to link provision to membership directly.
Who maintains the shared background?
Journalists, teachers, religious figures, neighbours and families have kept a community’s doxa, unevenly and with room for correction. As language models mediate more of that work, a small set of pipelines can come to dominate it while the people affected have little ability to inspect, challenge or change what those pipelines reproduce. The background drifts toward what the dominant system encodes, with no effective route to correction.
Or the background dissolves.
The opposite risk: references diverge far enough that participants can no longer sustain the expectations a coordination problem requires. Different references need not do this; communication, translation and institutions can keep coordination without identical priors. What is lost is those connections.
For the curious
§8.2: the two risks are concentration without effective correction and fragmentation without effective coordination, risks to the flow, openness and resilience criteria of §4.4. A mixture of references insures its users at the price of a component’s weight, KL(P‖μ̄) ≤ KL(P‖μk) − log wk, and reweighting a fixed reference cannot restore an excluded option. The practical challenge, in the paper’s words, is to “keep enough shared structure to coordinate without removing the evidence or authority needed to revise it.”
One fitness dominates, one reference takes all.
Maintainers’ shares grow with the fitness their references earn. Each point is one mix of three shares. When fitness is a reference’s own predictive accuracy and one reference is uniquely closest to the world, every mix flows to its corner: that reference takes the whole population.
Reward what each adds, and a spread survives.
Under fixed fitnesses only exact ties keep a mix. What holds a spread is rewarding each reference for its marginal contribution to the mixture’s accuracy. Then the mixture converges toward the world, and where the world is itself a mixture every component keeps a positive share: the contest ties the fitnesses itself.
The paper’s point: neither dominance nor a spread settles capture or collapse on its own. The reward rule decides whether the contest rewards the coverage and correction a community needs.
For the curious
§8.3: shares follow replicator dynamics, ẇk = wk(fk − f̄). With fixed fitnesses the solution wk(t) ∝ wk(0) efkt traces the exponential tilt of Link 2, with time as the inverse price. The motion drawn here illustrates the two regimes; the point the spread settles on is chosen for the drawing.
The last economy
The constitutional question stays with the community that owns it.
Provisioning, alignment and the aggregation of beliefs share a governance question without being the same problem: what produces and sustains the shared reference, how its use changes other activities, and who can govern it. The law cannot say who counts as a member or what flourishing is. It lets those questions be stated exactly.
The paper reads the sequence of economic ages, land, labour, capital, intelligence, as a sequence of changing constraints, and proposes where pressure moves next: as cognition and action get cheap, the returns shift toward specifying worthwhile ends and maintaining the references needed to pursue them. Doxa becomes a contested factor of production, an expectational commons that can be built, enclosed or depleted. The economy whose central task is tending it is the last economy.
Intelligent Economics: A foundation for the last economy. Emad Mostaque, May 2026, 47 pages.
Figures are computed in your browser from the paper’s formulas, on an example reference and value drawn for this page. Scenes marked illustrative carry no data.
Appendix · for the economists in the room
Go deeper. Check the working.
The story above is the main path. These two pieces are for anyone who wants to test it: the reason the law must be exponential, and the paper’s boxed results with how strongly it holds each.
A · Why this curve
Unrelated choices must stay unrelated. One curve allows it.
Say value reshapes the background through some response f. Take two unrelated decisions, what to eat and which way to walk home. Their values add. If they really are unrelated, their probabilities should multiply.
That demands f(a + b) = f(a) × f(b) for every pair. Among positive, increasing responses only the exponential passes. Its one free setting is the scale, and that scale is the price τ.
For the curious
This is Cauchy’s exponential functional equation. With the monotonicity the paper requires, its only solutions are f(v) = ev/τ. The derivation is Link 2 of the chain in §3.2. The step is Luce’s choice axiom, the independence condition beneath McFadden’s logit.
Test three responses
computed live| Response f(v) | f(a + b) | f(a) × f(b) |
|---|
Black bar: f(a + b). Gold bar: f(a) × f(b). They match for every a and b only on one row.
B · What is claimed
One premise, one identification, and what follows.
The paper marks which claims are which: identities that are exact for the choice law itself, contextual results that hold given a stated model or geometry, interpretations, and one hypothesis. It is motivated by one premise, that persistence favours systems that predict accurately and cheaply, and the stated consistency requirements give each result its scope. The doxa reading is an interpretation. Where a classical model is recovered exactly, its evidence carries across; the identities and bounds are reproduced numerically by the script verify.py released with the paper.
The paper’s boxed results, by section. Choose a part to bring it forward.
Intelligent Economics: A foundation for the last economy. Emad Mostaque, May 2026, 47 pages.
Read the paper →