S1-103 – When a mechanism becomes a model

A thermostat that almost thinks

A thermostat on the wall is doing something that, described carefully, sounds a great deal like thinking. It senses the temperature of the room. It holds a target it has been given. And when the two drift apart it acts, clicking the heating on until the gap closes and clicking it off again when the room comes right. It has, in the loosest sense, a goal and a way of pursuing it, and it pursues it tirelessly, day and night, without being told.

And yet nobody is tempted to say the thermostat has a model of the room, or knows the temperature, or could be wrong about the world. It is, we want to say, just a mechanism. The whole weight of that sentence sits on the word just, and the word is doing real work, because the thermostat passes several of the tests we might first reach for. It senses. It aims. It acts on what it senses. Article S1-102 defined a model as a representation that organizes part of the world, generates expectations, and guides action, and the thermostat looks unnervingly close to all three.

This article is about the line the word just is tracking: the line between a mechanism and a model. Every model is a mechanism, some arrangement of matter that does what it does by ordinary physical cause. But not every mechanism is a model, and the difference is not a matter of complexity or cleverness. It is a difference in kind, and the cleanest way to see it is to ask not what a model is but where the first one came from, because life spent most of its history building mechanisms that were not yet models, and the moment one crossed over is visible to anyone who knows what to look for.

What the thermostat is missing

Start with the tests that do not draw the line, because ruling them out is half the work.

Having a function does not draw it. The thermostat has a function, to hold the room near a set temperature, and it performs that function well. If a function were enough to make a model, every well-made tool would be one, and the word would cover the whole built world and explain nothing. Tracking something does not draw it either. The thermostat reliably tracks temperature, its state rising and falling with the room’s, and still we deny it a model. Reliable tracking is what a good mechanism does; a mercury column tracks temperature too, and no one thinks the thermometer believes anything.¹

What the thermostat lacks shows up in three linked features that genuine models have and it does not. First, a model stands in for something. It is a surrogate the system can consult in place of the world, not merely a part that moves when the world moves. The thermostat does not carry a stand-in for the room; it is wired directly to the room’s present temperature and swings with it, like a hand pushed by a current rather than a chart of the tides. Second, a model can be run in the absence of the thing it is about. One can consult a map of a city one is nowhere near; the map holds the city while the city is out of reach. The thermostat can do nothing when the room is not there to move it. Its competence is welded to the present moment and the present contact. Third, and this is the one that matters most, a model can be wrong. Not broken, wrong. A snapped spring is not a false spring, and a thermostat with a fouled sensor is not lying about the room, it is malfunctioning. But a map can show a road that is not there, and a belief can be false while the believer is perfectly intact, because the map and the belief make a claim about how things stand, and a claim is the kind of thing that can fail to match.²

So the line is this. A mechanism becomes a model when it carries a stand-in for some part of the world, a stand-in the system can use when that part is out of reach, and that can therefore be wrong about it. The thermostat has none of the three. It is a splendid mechanism and not a model at all, and seeing exactly why is the key that opens the rest.

Evolution’s model, and the animal’s

Now the harder and more interesting question: where in the history of life does the first model appear, and what is still only a mechanism?

Begin at the bottom, with a reflex. A hand touching something hot withdraws before the person has decided anything, before the pain has even fully arrived. The reflex does nothing until it is triggered, and when triggered it runs the same way every time. It is a mechanism in the plainest sense, a fixed arc from stimulus to response with nothing in between that stands for anything. And yet the reflex is not arbitrary. It embodies a true generalization about the world, that a hot thing against the skin is about to do damage and that pulling away pays. Something in the design knows this. But the knowing is not in you, and it is not in the reflex. It is in the long history of selection that built the reflex because the generalization held, generation after generation, for your ancestors.

The philosopher Daniel Dennett has the exact phrase for what is going on here. He calls it a free-floating rationale.³ There is a reason the reflex is shaped as it is, a real reason, and it is genuinely about the world. But the reason floats free of the creature that carries it. The reflex does not represent the danger of heat; it simply is the behavior that the danger of heat selected for. The rationale belongs to evolution, not to the animal. This is the first thing to hold onto: every adaptation carries a model of its world in this sense, a bet about how the world is, baked into the shape of the mechanism by the fact that the bet kept paying off. Call it evolution’s model. It is real, and it is about the world, and the organism does not have it. The organism is it.

Watch the bet get more lifelike without yet becoming the animal’s own. A bacterium such as Escherichia coli swims by a strategy called run-and-tumble: it runs in a straight line for a moment, then tumbles to a new random heading, then runs again. Drop it into a gradient of something nutritious and the strategy tilts. When the cell senses the food concentration rising as it swims, it tumbles less often, so its runs in the good direction grow longer, and the random walk drifts, on average, up the gradient toward the food.⁴ It looks purposive, and in the free-floating sense it is, but there is still no model in the cell. It is reacting to the gradient that is present now, the way the thermostat reacts to the heat that is present now.

Here is the case that seems to cross the line, and it is worth slowing down on, because it is exactly where the temptation lives. Put the same bacterium in a uniform bath with no gradient at all, and it does not stop. It keeps swimming, running and tumbling in an unbiased walk, wandering. One could describe this as the cell acting on an assumption, that food is out there somewhere and that moving improves the odds of finding it, and the description is not wrong. The behavior has come unstuck from the present stimulus; the cell is acting with no signal to react to. Surely that is a model saying go and look?

It is not, quite, and the reason is the sharpest point in this article. What has come unstuck from the present is the behavior, not a representation. The cell acts without a present signal, but nothing inside it stands for where the food is, so nothing inside it could be wrong about where the food is. The assumption that food is worth searching for is real, but it is once again evolution’s assumption, written into the search strategy because searching paid off for the cell’s ancestors. A wandering bacterium is a free-floating rationale in motion. Decoupled behavior, acting in the absence of a trigger, is not yet a decoupled model, a stand-in the creature carries and could be mistaken about. The line is not crossed by moving without a stimulus. It is crossed by holding something inside that stands for the world and can misrepresent it.

The first flicker, and the first map

And now, in the very same bacterium, the first flicker of the thing itself.

The cell does not read the gradient in an instant. It cannot; a single point in space gives no direction. What it does instead is compare, holding the concentration it met a moment ago against the concentration it meets now, and steering by whether the number is going up or down. That comparison is implemented in a slow chemical memory, a running record laid down in the methylation of its receptors that spans the last few seconds of its swim.⁴ And that record is the first thing we have met that stands in for something not present at the sensor now, namely how things were a second or two back. It is the first thing inside a living creature that could be wrong, in the modest sense that a sensor error or a lag can leave the record out of step with the world, so that the cell steers as if things were improving when they are not. It is a held stand-in, usable across a gap, capable of error. It is not yet much of a model, but it is no longer only a mechanism. It is the hinge.

Follow the hinge upward and the stand-in becomes unmistakable, and it appears far below anything we would be tempted to call a brain.

A desert ant of the genus Cataglyphis leaves its nest, wanders a long looping search across a stony plain that offers almost no landmarks, finds a scrap of food, and turns and runs a nearly straight line home. To do that it must carry, and update at every step, a single running estimate of the direction and distance back to the nest, a home vector it computes from each leg of the outbound trip. That estimate is a stand-in in the full sense: the nest is nowhere in its senses, yet the ant steers by the held quantity as though the nest were being pointed at. And the stand-in can be flatly wrong. Lift an ant at the food and set it down somewhere else, and it runs off on the vector it already had, covering the direction and distance that would have carried it home from where it started, arriving at a patch of open ground where nothing is, a fictive nest, and only then breaking into a search. It is steering by an inner quantity that no longer matches the world, misled not by an injury but by a map. The ant does this with a brain smaller than a pinhead and nothing resembling a cortex, which quietly settles a tempting assumption: the first clear model does not wait for a large brain, or a mammal, or anything we would be inclined to call a mind.⁵

Richer still is the mammal’s version. A rat that has found food in the arm of a maze will, later, in the dark, with nothing in front of its senses to guide it, go back to that arm. Something in the rat holds the layout of the maze and the place of the food while both are out of reach, and the rat steers by the held thing. The psychologist Edward Tolman called this a cognitive map, and decades later the discovery of place cells in the hippocampus, neurons that fire when and only when the animal is in a particular spot, showed that the map is not a metaphor but a structure one can find in the tissue.⁶ This map can misrepresent in the fullest sense. Move the food while the rat is away and it will run confidently to where the food was and is not, misled not by a broken leg but by a map that no longer matches the world. A stand-in for an absent thing, run offline, capable of being flatly false: a model in every sense the first article gave the word.

So the path from mechanism to model is not a jump but it is a real crossing, and now we can say exactly where it lies. The reflex, the thermostat, the chemotaxis, the wandering search are all mechanisms, each carrying evolution’s free-floating rationale and none carrying a model of its own. The bacterium’s few-second memory is the first internal stand-in, the hinge. The desert ant’s home vector is the first full model, held, usable in the world’s absence, and able to be wrong; the rat’s cognitive map is the same achievement grown richer and learned.

Two things the word carries

This leaves the word model doing two jobs, and keeping them apart clears up most of the confusion the distinction causes.

There is evolution’s model, the bet about the world that selection has built into the design of every adapted mechanism, the reflex and the swimming strategy included. This is the sense in which a creature can be exquisitely fitted to its world without representing anything: the fit is inherited, the rationale free-floating, the knowing done long ago by the lineage rather than now by the animal. This is what an earlier article meant by calling the brain’s deepest expectations a kind of inheritance, structure that experience did not install because there was no time to install it (article S1-101).⁷ Evolution is the oldest map-maker, and most of what it draws it draws into the mechanism itself.

And there is the animal’s own model, the stand-in it carries in the moment, holds across a gap, and can be wrong about. This is the thing that can mislead the individual in its own lifetime, the map that sends the rat to the empty arm, and it is the thing the rest of this series is chiefly about. The two are continuous, the second grew out of the first, but they are not the same, and only the second is a map in the sense that can fail you while you are otherwise perfectly well.

That second sense is why the distinction earns its place at the foundation of the series rather than in a footnote to it. The discipline this project preaches, that the map is not the territory, is a rule about the things that can be wrong about the world, because those are the things that can quietly mislead us, and a mechanism cannot. A snapped spring will stop a person but it will not deceive them; a false map will walk them off a cliff while working exactly as a map is built to work. Everything the series goes on to say about holding beliefs with calibrated confidence, about watching for where our models fail, applies to models and not to mechanisms, and now the reader knows which is which and where the boundary runs.

One honesty keeps this from hardening into a settled fact. The line drawn here is a good working line, the one most useful for this project, but it is not uncontested, and it is disputed from both sides at once. Below the ant the frontier is genuinely open. A slime mold, a single giant cell with no neurons at all, can find the shortest path through a maze between two scraps of food, and can even learn to expect a spell of cold, dry air that arrives on a regular beat, slowing itself in anticipation before the next one comes. Researchers in the growing field of basal cognition read feats like these as models in nearly the full sense, cognition reaching down into systems with no nervous system at all; others hold that they are superb mechanisms which meet the three marks only if the marks are read loosely, and that the honest label stays mechanism until more is shown.⁸ Above the ant the dispute runs the other way: there are researchers who would deny that even the rat’s cognitive map is a real representation, and read it too as elaborate coupling with no inner stand-in that could be true or false. That second argument is live and deep, and a later article in this series walks straight into it, where the question stops being where the first model appeared and becomes whether the whole notion of an inner model survives hard scrutiny at all (article S1-506). This is why the desert ant, rather than the bacterium below it or the rat above it, is the cleanest place to plant the flag: it is the first case almost no one disputes, a stand-in held, run offline, and shown to be wrong. For the foundation, this much is enough: a mechanism becomes a model when it holds a stand-in for the world that it can use in the world’s absence and be wrong about, the thermostat has none of that, the ant already has all of it, and the crossing between them is one of the oldest and most consequential events in the history of life.

Further reading

Daniel Dennett’s Darwin’s Dangerous Idea (1995) is the source of the free-floating rationale and the best long treatment of how design and reasons can be real without anyone having them in mind, which is the idea that does the heavy lifting in the middle of this article.

Fred Dretske’s Explaining Behavior: Reasons in a World of Causes (1988) is the clearest book on why the capacity to misrepresent, to be false rather than merely broken, is the mark that separates a genuine representation from a mechanism that only reacts.

Howard Berg’s E. coli in Motion (2004) is a short, vivid account of how a bacterium swims, senses, and remembers, written by the physicist who did much of the work, and it makes the chemotaxis story in this article concrete and quantitative for a reader who wants the machinery.

Rüdiger Wehner’s Desert Navigator: The Journey of an Ant (2020) is the life’s-work account, by the biologist who ran most of the experiments, of how an animal with a brain smaller than a pinhead finds its way home across a featureless plain, and it is where the path-integration story and the displacement experiments can be seen in full.

Notes

¹ The stock example of a mechanism that regulates beautifully without representing anything is the centrifugal governor that James Watt fitted to the steam engine, spun by the engine’s own output and throttling the steam as it spins. Tim van Gelder used it as the centerpiece of an argument that not all sophisticated control involves representation (“What Might Cognition Be, If Not Computation?”, Journal of Philosophy, 1995). The thermostat is the same point in a form every reader has on a wall.

² These three marks correspond to three strands in the philosophy of mind. That a representation must be able to misrepresent, to have accuracy conditions it can fail, is Fred Dretske’s criterion (Explaining Behavior, 1988; “Misrepresentation,” 1986). That a genuine representation can be decoupled from its object and used offline, in the object’s absence, is developed by Peter Godfrey-Smith and Kim Sterelny (Sterelny, Thought in a Hostile World, 2003). That merely reliable “receptor” states which switch on when a stimulus is present do not yet earn the name representation, which is reserved for states that function as exploitable stand-ins, is the “job description challenge” of William Ramsey, Representation Reconsidered (2007). The account here folds the three into one working test.

³ Daniel C. Dennett, Darwin’s Dangerous Idea: Evolution and the Meanings of Life (1995), and again in From Bacteria to Bach and Back (2017). A free-floating rationale is a reason that explains why a design is as it is, and is genuinely about the world, without being represented anywhere in the organism that carries the design; the reason belongs to the process of natural selection, not to the creature.

⁴ Escherichia coli swims by alternating straight runs with reorienting tumbles; in a uniform medium the walk is unbiased, and in a gradient of attractant the cell suppresses tumbling when it senses the concentration rising, lengthening its runs up the gradient. Because a single instant gives no spatial information, the cell compares the present concentration against the recent past, a temporal memory implemented in the reversible methylation of its chemoreceptors and spanning of order a few seconds. The classic demonstration that the mechanism is a temporal comparison is Jeffrey Segall, Steven Block, and Howard Berg, “Temporal comparisons in bacterial chemotaxis,” Proceedings of the National Academy of Sciences 83 (1986): 8987-8991; the accessible book-length account is Howard Berg, E. coli in Motion (2004).

⁵ The desert ant Cataglyphis navigates by path integration: it continuously combines the directions and distances of its outbound path into a single running home vector, then steers by that vector to return in a near-straight line. The decisive evidence that the vector is an internal stand-in capable of error is the displacement experiment, in which an ant captured at the feeder and released at a distant site runs off the home vector appropriate to the feeder and arrives at a “fictive nest” in empty terrain before beginning to search. The classic demonstration is Martin Müller and Rüdiger Wehner, “Path integration in desert ants, Cataglyphis fortis,” Proceedings of the National Academy of Sciences 85 (1988): 5287-5290; the full account, including how so small a nervous system implements the computation, is Rüdiger Wehner, Desert Navigator: The Journey of an Ant (2020).

⁶ Edward C. Tolman, “Cognitive maps in rats and men,” Psychological Review 55 (1948): 189-208, argued from maze experiments that a rat learns a map of its environment rather than a chain of stimulus-response habits. The physical basis was found by John O’Keefe and Lynn Nadel, whose The Hippocampus as a Cognitive Map (1978) reported “place cells” that fire selectively when the animal occupies a particular location; O’Keefe shared the 2014 Nobel Prize in Physiology or Medicine for the work.

⁷ The reading of an organism’s inherited expectations as a model deposited by selection, rather than learned, is evolutionary epistemology; see Konrad Lorenz, “Kant’s Doctrine of the A Priori in the Light of Contemporary Biology” (1941), and Karl Popper, Objective Knowledge: An Evolutionary Approach (1972). The distinction this article draws between that inherited design-model and the online model an animal carries and can be wrong about in its own lifetime is the series’ own framing.

⁸ That a non-neural organism can perform feats that look cognitive is the province of basal cognition. Toshiyuki Nakagaki, Hiroyasu Yamada, and Ágota Tóth, “Maze-solving by an amoeboid organism,” Nature 407 (2000): 470, showed the plasmodium of Physarum polycephalum connecting two food sources along the shortest path through a maze; Tetsu Saigusa, Atsushi Tero, Toshiyuki Nakagaki, and Yoshiki Kuramoto, “Amoebae anticipate periodic events,” Physical Review Letters 100 (2008): 018101, showed the same organism slowing its advance in anticipation of unfavorable conditions delivered at regular intervals. For the broader claim that cognition reaches below the neuron, see the basal-cognition literature associated with Michael Levin and Pamela Lyon. Whether these systems hold genuine models or only rich mechanisms is exactly the contested question the closing paragraph marks.