Bounded rationality and satisficing: why continuous ingestion follows

The stopping rule inside every decision

Economics once assumed the optimising agent: a decision-maker who surveys every option, prices every consequence, and picks the best. Herbert Simon's objection was not moral but arithmetical. Real agents have finite attention, finite memory, finite time. A purchasing manager cannot enumerate every supplier on earth, a firm cannot price every possible contingency before committing to a plan. Search itself costs something, and the cost rises before the returns do. Simon's claim was that this scarcity is not friction around rational choice; it is the shape of choice.

His answer was satisficing, a deliberate portmanteau of "satisfy" and "suffice." An agent sets an aspiration level in advance — good enough, given the stakes and the cost of looking further — and takes the first option that clears it. Search terminates on adequacy, not on proof of optimality. This is not laziness dressed up in jargon. It is a description of a stopping rule, and Simon's deeper point is that the stopping rule usually goes unexamined. People argue about objectives constantly. They rarely interrogate when they stopped looking, or why there.

That asymmetry matters because the stopping rule, not the objective function, is what predicts behaviour. Two firms with identical goals and identical information will act differently if one has a higher aspiration threshold or a cheaper search process. Simon's bounded rationality reframes economic behaviour as procedural: rational given the constraints on computing, not rational in some unconstrained sense. It is a theory of limits treated as load-bearing rather than as noise to be averaged away.

Origin

Simon set this out in Administrative Behavior (1947), built from his study of how administrators actually decided things, and sharpened it into "satisficing" in a 1955 paper in the Quarterly Journal of Economics. The empirical puzzle was concrete: firms visibly did not compute optima — no manager was solving a global maximisation over an unbounded option space before lunch — yet firms were not obviously irrational either. Something coherent was happening that neoclassical theory had no room for. Simon's resolution, treating rationality as bounded by attention and search cost, earned him the 1978 Nobel Memorial Prize and fed directly into behavioural economics, organisation theory, and artificial intelligence, where it shaped early work on heuristic search long before anyone spoke of large models of anything.

The turn

Every generation in the Large Language Model to Large World Model to Large Universe Model lineage can be read as a different stopping rule applied to evidence about the world, and once you see it, the lineage stops looking like a sequence of bigger models and starts looking like a sequence of relaxed satisficing boundaries.

The Large Language Model satisfices architecturally. A corpus is assembled, judged sufficient, and frozen at a training cutoff. Nobody revisits that cutoff during deployment; the aspiration level was set once, at construction time, and the system inherits it as a fact about its own existence rather than a live decision. The Large World Model relaxes the rule in duration rather than in kind: sensing runs while a scene is present — the camera pointed, the episode live — and terminates when the scene ends. There is no standing belief carried between engagements, only a search process that restarts each time perception restarts. The Large Universe Model is defined by removing temporal closure on intake itself. Every stream keeps running. Nothing is filed as finished. Beliefs are held open, revisable, tagged with provenance and subject to decay rather than fixed at either a cutoff or a scene boundary.

Simon's framework says why this matters more than it looks. He argued that the boundary on search was not a defect awaiting engineering, but a direct consequence of scarce attention: looking further cost more than it returned, so agents stopped. That argument holds exactly as long as attention stays expensive. Once attention over intake becomes cheap and continuous — not free, but cheap relative to the stakes — the equilibrium behaviour of any agent facing that boundary should shift, because the calculation that justified stopping no longer produces the same answer. The Large Universe Model is what that shifted equilibrium looks like when the constraint being relaxed is specifically the temporal one on observation.

What does not follow

The common misreading says continuous intake finally builds the optimising agent economics used to assume, closing Simon's gap by brute force. This is wrong on two counts, and worth disowning explicitly. First, Simon's bounds were never only informational. Computation and attention are bounded too, and continuous intake addresses only the temporal cutoff on information, leaving the cost of computing over that information exactly where it was. Second, and more importantly, this misreading inverts Simon's actual lesson. Satisficing is not a failure mode to be eliminated; it is efficient behaviour under real constraints. A system that never closes its evidence window still has to decide, at some point, what to act on — and that decision remains a satisficing decision, made under an aspiration level, terminating on adequacy rather than proof. The intake boundary has moved. The judgement boundary has not vanished; it has relocated, from what the system observes to what it is willing to believe and act on given what it has observed.

Objections, taken straight

Simon's whole point was that unbounded search is not just costly but incoherent — the space of possible observations has no natural boundary, so "everything" was never a coherent target to begin with.

This is correct, and the honest version of the claim concedes it rather than arguing around it. "Everything, continuously" does not mean every possible observation captured; it means no temporal closure on the streams a system is actually connected to. The category claim concerns the absence of a cutoff, not the presence of totality. Which streams to connect, at what resolution, remains a satisficing choice, made and remade. That is precisely why provenance is not decoration here: it is what makes the residual sampling decision inspectable rather than buried inside an architecture nobody revisits.

Gigerenzer's ecological rationality shows that fast heuristics on less information often beat exhaustive search under genuine uncertainty, because added data adds variance faster than it removes bias. Continuous intake could make judgement worse than a well-chosen frozen corpus.

The less-is-more results are real and well replicated, but they are results about which inputs feed an inference rule, not about how stale the evidence base is permitted to become. A frozen corpus does not reduce variance relative to continuous intake; it substitutes a stale estimate for a current one. Gigerenzer's heuristics work because the environment is stable enough that a rule learned once stays valid — an assumption a cutoff can never verify, since a cutoff has no way of noticing the environment has drifted. Continuous intake is what makes drift detectable at all. Choosing few, well-selected cues downstream of that detection remains sound practice; it is a separate decision from whether the window on evidence stays open.

If every agent gains continuous intake, the strategic environment shifts and the gains dissipate — an arms race in observation, as in high-frequency trading, where microsecond advantages consumed enormous infrastructure spend for no aggregate benefit.

This is the strongest objection, and it is right about competitive, zero-sum settings. But most consequential intake is not adversarial in that sense. A grid operator watching phasor measurements, an epidemiologist watching wastewater sequencing, a bridge inspector watching strain gauges — none of these are racing another party for private advantage over the same signal. Their gains are absolute. The arms-race dynamic narrows the claim to a specific class of situations, observational competition between strategic agents, and leaves it largely intact elsewhere. Outside that class the binding cost shifts from latency to trust: verifying what a stream actually shows.

What this establishes, and what it does not

Bounded rationality locates the lineage's terminal rung precisely, on this one axis. There are exactly three positions available for when evidence-gathering stops: once, at a cutoff; for the duration of a scene; or never. The third exhausts the axis — any proposed fourth position is either a subset of continuous intake or a claim about processing evidence rather than observing it, which is a different axis entirely. Simon's contribution is to explain why the third position was historically unreachable rather than merely unattempted: attention was the binding constraint, and search stopped when it did. What Simon does not establish is that removing this particular boundary yields better judgement, or that judgement itself becomes unbounded. It does not. It only relocates where the unavoidable satisficing decision sits — from whether to keep watching to what, among everything still arriving, deserves belief.