Put a desert ant on stilts and it will walk straight past its own nest.
The experiment is exactly as blunt as it sounds. Cataglyphis foraging ants wander a long looping path across featureless Saharan salt pan, find food, and then run a near-perfect straight line home, which is a genuinely remarkable feat of navigation in an animal with under a million neurons. Researchers glued pig bristles to the legs of ants that had already completed the outbound trip, lengthening their stride for the return. The stilt-walkers overshot the nest by a predictable margin. Ants with shortened legs stopped short. The animal was not consulting anything resembling a chart. It was counting steps, multiplying by stride length, and integrating that against a compass heading, and when the experimenters changed the stride length the arithmetic came out wrong in exactly the way the arithmetic should.
That result is the single most useful thing to hold onto in this subject, because it demonstrates the gap between spectacular navigation and an actual cognitive map. The ant is doing something extraordinary and it is not doing what the word map implies. It cannot take a novel shortcut from an arbitrary point. It has no representation of the relationship between two places it has never traveled between. It has a running estimate of the vector home, and if you corrupt the odometer the estimate fails gracefully and completely.
Seventy-odd years of work separates Edward Tolman’s proposal that animals build internal maps from the current state of the field, and the most interesting developments of the last decade have gone in two directions at once. The neural machinery turned out to be more geometrically specific than anyone expected, sitting on a mathematical structure that can be measured. And the thing it maps turned out not to be space.
What Tolman actually claimed, and what he got right by accident
In 1948 Edward Tolman published “Cognitive Maps in Rats and Men” into a field that did not want it. Behaviorism held that learning was a matter of stimulus-response chains strengthened by reinforcement, and a rat running a maze was assembling a sequence of turns, not a picture.
Tolman’s evidence was awkward for that account. In latent learning experiments, rats allowed to wander an unrewarded maze for days performed dramatically better than naive rats once a reward was introduced, which means they had been learning something with no reinforcement to strengthen anything. In place-versus-response experiments, rats trained to reach a goal from one starting point, then started somewhere else, went to the place rather than repeating the turn sequence. In the sunburst maze, rats trained on an indirect path to a goal, then offered a fan of new radial alleys, disproportionately chose the alley pointing at where the goal actually was.
None of that proves an internal map in any strong sense, and Tolman’s critics said so at length. What it establishes is that the animal learned spatial relationships it was never rewarded for learning and could use them in configurations it had never experienced, which is the operational signature people still test for.
The part that gets skipped is Tolman’s last section, where he took the idea somewhere strange. He argued that narrow, brittle maps produced by fear and frustration explained regression, fixation, displaced aggression, and social prejudice, and that the same mapping machinery organizing a rat’s maze also organized how people represent social and abstract relationships. In 1948 this read as a psychologist overreaching past his data, and for fifty years it was treated as the eccentric coda to an important paper. It is now the most vindicated part of it, for reasons Tolman had no way to anticipate.
The cell types, and what each one is actually doing
The neurobiology arrived in stages and it is worth knowing what each component contributes, because the popular version collapses them all into brain GPS and loses the mechanism.
Place cells, found by John O’Keefe and Jonathan Dostrovsky in 1971, fire when a rat occupies a particular location in a particular environment. Each cell has one or a few place fields, the population tiles the environment, and every location produces a distinctive pattern of active cells. Move the animal to a different environment and the cells remap, forming a new and largely unrelated assignment, which means place cells encode this place in this context rather than coordinates in any absolute frame.
Grid cells, reported by May-Britt and Edvard Moser’s group in 2005, sit upstream in medial entorhinal cortex and do something stranger. A single grid cell fires at multiple locations arranged in a hexagonal lattice tiling the entire environment, like a triangular tessellation laid over the floor. Grid cells cluster into modules with discrete spacings that increase along the dorsal-to-ventral axis, cells within a module share orientation and spacing while differing in phase, and the module scales follow an approximate power-law relationship. That arrangement is a plausible metric, a coordinate system with multiple resolutions, and it is why the 2014 Nobel Prize went to O’Keefe and the Mosers.
Then the supporting cast. Head direction cells fire when the animal faces a particular direction regardless of location, functioning as a compass and depending heavily on vestibular input. Border cells and boundary vector cells fire at a specific distance and direction from environmental edges, and they matter more than their billing because boundaries appear to serve as an error-correction reference for a grid system that otherwise accumulates drift. Speed cells encode running speed. Object vector cells fire at a set distance and bearing from discrete landmarks. Time cells fire at particular moments during a delay, tiling elapsed time the way place cells tile space, which was the first strong hint that the machinery was indifferent to whether the dimension it was tiling had anything to do with physical distance.
The 2024 review of grid cell mechanisms and function lays out where the mechanistic arguments now stand, and the short version is that continuous attractor network models, in which recurrent connectivity constrains the population to a low-dimensional state that gets pushed around by velocity input, have accumulated substantially more experimental support than the competing oscillatory interference accounts.
The torus, and why topology beat tuning curves
The result that settled a long-running argument came in 2022, and it is a good example of a finding that is hard to explain and worth the effort.
Continuous attractor models predict something specific and non-obvious. If grid cell activity in a module is generated by a recurrent network whose stable states form a two-dimensional sheet with periodic boundary conditions, then the population activity should live on a torus, a doughnut surface, regardless of what the animal is doing or where it is. Not the firing pattern on the floor, which everyone had already seen. The shape of the population state space itself.
Recording large numbers of grid cells simultaneously and applying topological data analysis, researchers found exactly that. The population activity of a grid module occupies a toroidal manifold, the torus persists across environments, it persists in darkness, and it persists during sleep when the animal is not navigating anything at all. That last point is the one that matters most, because it means the structure is intrinsic to the network rather than imposed by sensory input.
This is a different kind of neuroscience result than a tuning curve. It is a claim about the geometry of a neural population’s activity, tested with the mathematics of shape, and confirmed. It moves the grid system from a suggestive pattern to a mechanism with a demonstrated architecture, and it is the strongest evidence available that the brain implements something genuinely coordinate-like rather than merely producing coordinate-like output.
The persistence during sleep also connects the navigation system to something the animal is obviously not doing at the time, which is a recurring theme in this machinery. A structure that holds its shape with the eyes closed is a structure available for offline use, and offline use is where most of the interesting computation happens.
Path integration, vector memory, and everything a cognitive map is not
Back to the ant, because the field’s hardest problem is distinguishing a map from things that produce map-like behavior more cheaply.
Path integration, also called dead reckoning, requires a compass and an odometer and nothing else. Track your heading, track distance traveled, continuously update a single vector pointing home. It is computationally trivial, it works in featureless terrain, and it degrades predictably: errors accumulate with distance and never self-correct, which is why animals relying on it heavily also carry backup systems. Desert ants run a visual panorama-matching routine near the nest to clean up the final approach, because the integrated vector alone is not accurate enough to find a hole in the ground.
Vector memory is the next tier and it is still not a map. An animal that has traveled from A to B can store the vector and reuse it. Honeybees do this well, and the long argument over whether bees have map-like memory turns precisely on whether they can compute a novel route between two locations without having traveled it, which is a much harder claim than storing a library of learned vectors.
Beaconing is simpler still: head toward a detectable cue at the goal. Piloting means moving between recognized landmarks in sequence. Route following means reproducing a learned sequence of movements, and it can be extended almost indefinitely without ever becoming a map, which is how an animal can traverse a route of enormous length and complexity while remaining unable to deviate from it. Every one of these produces impressive navigation, and none requires the animal to represent the spatial relationship between places it has not connected by direct experience.
That last capability is the operational definition of a cognitive map, and it is why the evidentiary bar is so high. Novel shortcutting, and detour behavior when a familiar route is blocked, are the behaviors that cannot be produced by the cheaper systems. Everything else is compatible with an animal that has an excellent memory and no map at all.
The reason this taxonomy is worth memorizing is that it inverts the intuitive reading of almost every navigation story. A monarch butterfly crossing a continent to a grove it has never seen, four generations removed from the last butterfly that made the trip, is doing something a cognitive map could not accomplish, because there is nothing in its experience to build a map out of. It is running an inherited compass heading against a clock. That is less flexible than a map and far more impressive as a piece of engineering, and describing it as a map would make it sound easier than it is.
The same correction applies in reverse. A rat that has spent an hour in a small box, doing nothing anyone would call remarkable, may well be running the more sophisticated system, because it can be dropped anywhere in that box and head straight for a corner it has not approached from that angle before. Flexibility in unfamiliar configurations is the diagnostic, not distance covered or difficulty of terrain.
Compasses, magnetic maps, and the difference between them
Compass systems are separable from maps and animals stack multiple ones with a hierarchy of preference.
Sun compasses require time compensation, since the sun moves, and monarch butterflies solve this with circadian clocks located in the antennae feeding a sun-azimuth calculation in the central complex. Remove or paint the antennae and the compass fails while the clock in the brain keeps running. Many insects also read polarized skylight patterns, which persist under partial cloud and are detected by a specialized dorsal rim region of the eye.
Star compasses appear in birds, which learn the rotational center of the night sky as nestlings rather than inheriting specific constellations. Dung beetles orient by the Milky Way as a band, using it to roll a dung ball in a straight line away from competitors, which is the least dignified application of galactic astronomy on record and a genuine one.
Magnetic compasses are the contested case and deserve honest treatment. Two mechanisms remain live. The radical pair hypothesis proposes that cryptochrome proteins in the retina form spin-correlated radical pairs on photon absorption, with reaction yields sensitive to magnetic field orientation, giving a light-dependent inclination compass. Work on cryptochrome 4 from European robins showed magnetic sensitivity in vitro, which is real support and is not the same as demonstrating the mechanism operates in a living bird. The magnetite hypothesis proposes iron-mineral-based receptors transducing field intensity and direction mechanically, with the trigeminal nerve as a candidate pathway. Both may be true and serve different functions. Neither is closed, the field has a documented history of high-profile results that did not replicate, and anybody presenting magnetoreception as solved is ahead of the evidence.
The replication history deserves its own sentence because it is instructive about how a field can go wrong. Magnetoreception research has produced several widely cited candidate receptors that later analysis attributed to contamination, to iron-rich macrophages rather than sensory cells, or to effects that vanished under blind protocols. That record is not an argument that the sense does not exist, since the behavioral evidence for magnetic orientation is overwhelming across birds, turtles, fish, and insects. It is an argument that identifying the receptor is genuinely hard, that the incentives reward premature announcement, and that a reader should treat any new claim of a definitive magnetoreceptor with the same posture they would bring to a press release about room-temperature superconductivity.
Migratory birds appear to run all of these plus landmark memory, recalibrating one against another. The songbirds whose learning has been characterized in unusual detail navigate continental distances on their first attempt with no guide, while the cranes whose eastern migratory population had to be taught a route by aircraft demonstrate the opposite arrangement: a species where the route is culturally transmitted rather than inherited, and where losing the knowledgeable individuals means losing the route.
Here is a distinction that popular coverage almost never makes and that is central to the whole subject. A compass tells you which way is north. A map tells you where you are. These are different problems requiring different information, and an animal can have one without the other.
A magnetic map requires that the geomagnetic field vary predictably across the region in question and that the animal read at least two field parameters, typically inclination angle and total intensity, whose contours cross at an angle. Reading both gives a positional fix, in principle, in the same way that two intersecting lines of position give a fix in celestial navigation.
Sea turtles are the strongest case. Hatchling loggerheads exposed to magnetic signatures characteristic of specific locations along their migratory circuit orient in the direction appropriate for that location, in a laboratory tank, with no other cues, having never been there. That is positional information from the field itself. The current model holds that turtles imprint on the magnetic signature of their natal beach and use it to return decades later to nest, which explains an otherwise baffling homing feat and generates testable predictions about nesting distributions shifting as field lines drift.
More recent work has shown that loggerheads can learn to associate a magnetic signature with food, expressing a distinctive anticipatory behavior when placed back in the learned field conditions, which establishes that the magnetic sense is available to associative learning rather than being locked into a fixed navigational reflex. Salmon appear to use a similar geomagnetic imprinting mechanism for natal river homing, and spiny lobsters displaced to unfamiliar sites in ways that eliminated route-based cues still oriented homeward, which remains one of the cleaner demonstrations of true navigation in an invertebrate.
The relevance to marine mammals is unresolved and interesting. The orca populations whose ranging patterns have been tracked for decades follow routes stable enough to look like knowledge, bottlenose populations show location-specific foraging traditions, and the deep-diving species that cover ocean basins do something nobody has adequately characterized. The cod stocks whose migratory routes collapsed with the population raise the same question in a fishery context: if route knowledge is partly learned and partly social, removing the experienced animals removes the map.
Proving a cognitive map in the wild finally happened
Seven decades of cognitive map research produced enormous neurobiological detail and almost no field evidence from free-ranging wild animals, for a simple reason: ruling out the cheaper strategies requires knowing where an animal went, continuously, at high resolution, across a large area, for a long time, in many individuals.
That became possible with reverse-GPS. Instead of putting a receiver on the animal, a network of ground stations receives a signal from a lightweight tag and computes position by time-difference-of-arrival, which allows much smaller tags and much higher sampling rates than satellite GPS. Applied to Egyptian fruit bats in Israel’s Hula Valley, the resulting study of cognitive map-based navigation in wild bats tracked 172 individuals across 3,449 bat-nights over four years, producing more than eighteen million localizations, with every fruit tree in an eighty-eight-thousand-hectare study area mapped as a potential goal.
The findings: bats seldom searched randomly. They flew long, straight, goal-directed trajectories to specific trees, often ignoring closer trees of the same species. Of more than nine thousand recorded trajectories, several hundred were shortcuts between two known locations along routes the individual had never flown. Bats translocated to unfamiliar release points at the edge of their range returned along novel straight-line paths rather than searching or retracing. The analysis then worked through the alternatives systematically, using simulated tracks and trajectory analysis to rule out random search, beaconing, piloting, path integration, and following other bats.
A companion study tagged pups before their first outdoor flights and watched the map get built, with young bats making progressively longer exploratory excursions from the roost, extending their known area outward over months rather than arriving with it.
The scale of effort is the point worth extracting. It took four years, 172 animals, eighteen million position fixes, and a complete botanical inventory of an area larger than most cities to demonstrate something the field had assumed since 1948. That is what the evidentiary bar for a cognitive map actually costs.
The instrumentation point generalizes past bats. Reverse-GPS, miniaturized biologgers, wireless neural recording in freely moving and flying animals, and high-density silicon probes recording hundreds of neurons simultaneously have all arrived in roughly the same window, and each one converted a question that could only be argued about into a question that could be measured. The toroidal topology result was impossible before it became routine to record enough grid cells at once to reconstruct a population manifold. The bat map result was impossible before tags got small enough to fly on an animal that weighs about as much as a deck of cards. A meaningful fraction of what looks like conceptual progress in this field over the past decade is instrument progress arriving with a delay.
Three dimensions, and where the textbook was wrong
Almost all the foundational work was done on rats running on flat surfaces, which is a defensible simplification and turns out to have baked in an assumption.
Bats fly. Recording from bats in three-dimensional flight showed place cells with roughly spherical fields distributed through volume, and head direction cells organized in a three-dimensional scheme with toroidal topology covering azimuth and pitch. So far, so consistent.
Grid cells did not cooperate. The natural expectation was a three-dimensional analogue of the hexagonal lattice, something like a face-centered cubic packing tiling the volume. What flying bats actually showed was local order without global lattice structure: firing fields at characteristic distances from their neighbors, but no long-range periodic arrangement of the kind that defines two-dimensional grid cells. The metric is there. The crystal is not.
That is a real correction and it has consequences for how the code is understood. A locally ordered arrangement can still support distance estimation and path integration while giving up the elegant modular periodicity that made the two-dimensional grid so appealing as a coordinate system. It also suggests the hexagonal lattice may be partly a consequence of the constraint of moving on a plane rather than a universal solution to representing space.
There is a general lesson in it about the species you choose. A field built on animals that live on surfaces produced a theory suited to surfaces, and the correction arrived only when someone recorded from an animal that does not. The same critique applies to arboreal primates navigating a three-dimensional canopy, to fish, and to every marine animal in a volume rather than on a plane.
There is a second asymmetry in three-dimensional space that the flat-surface tradition never had to confront, which is that vertical is not equivalent to horizontal. Gravity provides an absolute reference, moving up costs more than moving sideways, and the range of vertical movement available to most animals is far smaller than their horizontal range. A representation optimized for that anisotropy should not be isotropic, and the locally ordered arrangement observed in flying bats may be exactly what an efficient solution to an anisotropic problem looks like. Whether animals that move freely in all three dimensions with less gravitational asymmetry, which is to say animals in water, use something different again is an open and largely untested question. The reef fish whose spatial and social behavior keeps outrunning expectations would be an obvious place to look.
Maps of things that are not places
This is the reframe, and it is the strongest reason to care about this subject even if you have no interest in navigation.
If grid cells implement a general coordinate system, there is no principled reason the axes have to be north and east. Human imaging work has found grid-like six-fold symmetric signals in entorhinal cortex while subjects navigate abstract spaces: in one influential study, participants learned to morph a bird stimulus along two continuous dimensions, neck length and leg length, and their entorhinal activity showed the same hexagonal signature as spatial navigation, in a task involving no space whatsoever. Comparable structure has been reported for social spaces defined by power and affiliation, for odor spaces, and for task and conceptual spaces generally.
The theoretical work followed. The successor representation frames the hippocampus as encoding predicted future states rather than current location, which reproduces place field properties while generalizing naturally to non-spatial sequences. The Tolman-Eichenbaum Machine and related models treat the hippocampal-entorhinal system as factorizing structural knowledge from sensory content, learning the abstract shape of a problem once and then binding new specifics onto it, which is why a familiar structure in a new environment is learned so much faster than a new structure.
There is a genuine competing account worth stating rather than burying. Some researchers argue the relationship runs the other way: that a domain-general clustering and concept-learning algorithm produces place-like and grid-like representations as a side effect when inputs happen to be uniformly distributed, as they are in an empty room, and produces more conceptual-looking representations when inputs are sparse and high-dimensional. On that reading the system was never a spatial map that got repurposed. It was always a general learning mechanism that we happened to discover in a rat in a box.
Either way, the practical implication holds. The machinery that lets a bat fly to a specific fig tree is the machinery that lets a person hold a family tree, an org chart, a chess position, or the relationship between concepts in a field they are learning. Tolman’s odd 1948 coda about maps of social relationships was not overreach. It was the part of the paper that took seventy years to catch up to.
One caution about how this reframe gets reported. Finding a six-fold symmetric signal in an imaging study of a conceptual task is considerably weaker evidence than recording grid cells directly, the analysis depends on methodological choices that have been contested, and the number of well-replicated non-spatial grid findings is smaller than the enthusiasm around them suggests. The direction of the evidence is consistent and the strength of any individual result is moderate, which is a normal state for a young idea and worth stating plainly rather than letting the accumulated citations imply more than any single study delivers.
Replay, sweeps, and a map that runs simulations
A map you can only read at your current position is a limited instrument. What makes the hippocampal system powerful is that it runs offline.
During sharp-wave ripple events, place cell sequences reactivate in compressed form, often during rest and sleep, sometimes forward and sometimes reversed relative to the original trajectory. Reverse replay after reaching a reward is well suited to propagating value backward along a path. Forward replay before movement looks like route selection.
More striking is what happens at decision points. A rat pausing at a maze junction, physically oscillating its head between the two options in a behavior called vicarious trial and error, shows place cell activity sweeping ahead down first one arm and then the other, representing locations the animal is not at and has not chosen. That is a spatial representation being used to evaluate hypothetical futures, which is a substantially different thing from a chart of where you are.
Replay also generates sequences for trajectories never taken, assembled from fragments of experienced ones, and the grid cell correlation structure is preserved during sleep, which is what allows the same machinery to run without sensory input. The system is not a map in the sense of a static document. It is closer to a simulator, and the navigational application is one thing it happens to be used for.
That framing also explains an otherwise odd clinical fact. Damage to the hippocampus produces amnesia, which is a memory disorder, and it also produces impaired navigation and an impaired ability to imagine novel scenes or plausible futures. Those look like three unrelated deficits under the map model and like one deficit under the simulator model: a patient who cannot assemble a coherent representation of a situation they are not currently in will fail at remembering the past, navigating to somewhere out of sight, and imagining tomorrow, for the same underlying reason.
The artificial agents trained on navigation tasks that spontaneously develop grid-like representations in their hidden layers are informative here without settling anything: a network optimized for path integration converging on something resembling a grid code suggests the solution is at least partly forced by the problem rather than by biology. That is a claim about computation, and the engineering work that reads and writes to these systems directly will eventually test it in a way simulation cannot.
What the map metaphor keeps getting wrong
An audit, because this subject generates unusually confident nonsense.
Humans have a GPS in their heads is the headline version of the Nobel work and it oversells in two directions. The system is not a global positioning system, since place cell assignments remap between environments rather than referencing an absolute frame, and it is not a receiver of external signal, since it constructs position from self-motion and landmarks. It is closer to an odometer plus a landmark-matching routine plus a relational memory, and calling it GPS imports the wrong intuitions about accuracy and absoluteness.
Grid cells are the cognitive map is a category error the field itself sometimes commits. Grid cells provide a metric. Place cells provide context-specific location. Boundary cells provide error correction. The map, if the word means anything, is the emergent product of the interaction, and no single cell type is it.
The London taxi driver result is usually flattened. Licensed drivers who completed the Knowledge showed larger posterior hippocampal grey matter volume than controls, with the effect scaling with years of experience, and trainees who qualified showed changes that those who failed did not. That is genuine evidence of experience-driven structural plasticity. It is not evidence that navigation training makes you generally smarter, and the same studies found the drivers performing worse on some other memory tasks, which suggests a reallocation rather than an upgrade.
A good sense of direction is a single trait does not survive testing. Individual variation in navigation performance is large, partly cultural, and factors into somewhat separable abilities involving path integration accuracy, landmark memory, and the ability to adopt an allocentric perspective at all. Some people appear to navigate almost entirely by route memory and do fine.
Satellite navigation is destroying our cognitive map is the newest entry and the evidence is thinner than the confidence around it. Studies have found that turn-by-turn guidance reduces hippocampal engagement during the task itself and that heavy lifetime use correlates with worse performance on some spatial tests, which is real and also exactly the pattern you would expect from any offloaded skill. Whether it produces durable structural change, whether the correlation runs the direction people assume, and whether it matters for anything beyond navigation are all unresolved, and the cleanest available reading is that not practicing a skill makes you worse at that skill.
Animals that navigate impressively must have cognitive maps is the error the ant kills. Path integration, vector memory, compass orientation, and landmark sequences produce feats that look map-like and are not, and the homing pigeons whose wartime performance made them famous were doing something that a century of research still has not fully resolved, involving some combination of magnetic, olfactory, visual, and possibly infrasound cues. The birds credited with saving units by returning through fire were extraordinary. What they were extraordinary at is still partly an open question.
Where this leaves the internal GPS
The picture that emerges is less tidy than the Nobel citation and considerably more interesting.
There is no single internal GPS. There is a stack of navigation systems of increasing cost and capability, and different lineages sit at different points on it, often running several at once with a preference hierarchy that shifts by context. Path integration is cheap, ancient, and available to animals with under a million neurons. Any account of a cognitive map that cannot say which tier of that stack an animal is operating on is not saying much. Compasses are cheap and heterogeneous, built out of eyes, antennae, clocks, and possibly iron. Maps in the strong sense are expensive and rare, and demonstrating one in the wild took four years and eighteen million position fixes.
Where the mammalian system is unusual is not that it navigates well. Ants navigate well. It is that the coding scheme turned out to be general. The same population geometry that supports getting to a fig tree supports representing a conceptual space, a social hierarchy, a task structure, and a hypothetical route nobody has taken, and it runs those representations offline while the animal sleeps. That is a memory and inference system that happens to have been discovered in a maze.
The elephants whose knowledge of water sources across enormous ranges keeps a family group alive through drought are the case that makes the stakes legible, since that knowledge is held disproportionately by the oldest matriarch and it dies with her. The populations studied under different pressures show how much of that map is individual experience rather than species instinct, and the cooperative hunters covering large territories raise the same question about how much spatial knowledge is distributed across a group rather than held in one head. The sentinel systems that let a foraging group use space it could not safely use alone are a version of the same trade, and the corvids that cache thousands of items and recover them months later are the standing demonstration that spatial memory capacity and brain size have a looser relationship than anyone expected.
Which is where the 24-lecture Neurozoology course puts the emphasis throughout, and it is the reason the ant on stilts is the right place to start rather than the Nobel Prize. Starting with the Nobel Prize teaches you that the brain has a positioning system. Starting with the ant teaches you to ask what any given animal is actually computing, which is the question that survives contact with the next twenty years of results. The nervous system is not a magic positioning device. It is a set of expensive, error-prone, energy-hungry estimators, each one solving a problem some ancestor actually had, layered on top of each other with no plan and no cleanup. The knowledge that visibly moves between animals, the working animals whose capacities were discovered by people who needed them, and the first edition’s survey of nervous systems all run on the same assumption.
Glue bristles to an ant’s legs and it walks past its own front door, having done everything right. That is not a failure of the animal. It is a precise readout of what the animal was actually computing, which is the only kind of answer worth having.