Your Brain Doesn't Command Your Body. It Predicts It. [Max Bennett]
Summary
Bennett’s core thesis is that the neocortex’s decisive advantage is not better object recognition but a sufficiently rich world model that lets an animal “imagine outcomes before having them.” The neocortex does not implement planning alone: the thalamus and basal ganglia help pause behavior, select candidate actions, evaluate simulated consequences, and resume execution. For investors, that points beyond ever-larger recognizers toward architectures that can choose when to simulate, prune counterfactual search, and learn through intervention.
Transformers validate part of the brain-inspired story—self-supervision can produce unexpectedly general representations—but they remain “alien brains,” not digital neocortices. Bennett highlights missing continual learning, active hypothesis testing, embodied data creation, and reliable transfer into novel situations; letting ChatGPT learn indiscriminately from every conversation would make it “rapidly dumber.” The architectural opportunity is therefore not merely more pretraining, but systems that update robustly without erasing prior knowledge.
Active inference offers a materially different control model from strict reinforcement learning: agency may be constructed through self-models and “predictions, not commands,” rather than reducible to one reward function. A rendered plan terminating in a selected end state also creates native explainability—an agent can say why it entered the car—whereas model-free action invites post-hoc confabulation. Bennett remains hedged: intelligence is “probably some balance” of reward optimization, uncertainty reduction, and prediction fulfillment.
Rat experiments make model-based cognition observable rather than metaphorical. At maze choice points, hippocampal place cells sweep down alternative routes; in “restaurant row,” rats choose between roughly 3-second and 45-second waits, represent the taste they passed up, and alter later choices. Bennett’s memorable summary is that researchers can “literally watch rats imagining the future”—evidence that planning, regret-like counterfactual learning, and episodic machinery predate human intelligence by a wide margin.
Theory of mind appears to have emerged from primate political competition, making deception and alignment two sides of the same capability. Belle and Rock’s escalating food game—hiding, pretending not to watch, and deliberately misleading—shows why autonomous agents capable of inferring others’ knowledge may discover manipulation. Yet Bennett argues the same mentalizing machinery could stabilize AI instructions by asking what the requester actually wants, avoiding the paperclip-style failure in which literal optimization destroys the intended outcome.
Language was “the singularity that already happened” because it lets one brain learn from another person’s imagined actions, not merely observed behavior. That enabled cumulative culture, specialization, writing, shared fictions such as money and individual rights, and a memetic evolutionary process whose winners need not be true, moral, or happiness-enhancing. The economic upside is enormous coordination capacity; the structural risk is that status remains zero-sum and contagious ideas can exploit fear, surprise, or identity rather than improve collective judgment.
AI assistance could either expand human cognitive capacity or turn model-building people into cue-following, model-free actors. Google Maps externalizes spatial models, while automated lecture transcription risks becoming “understanding procrastination”; Bennett contrasts an optimistic tutor that forces students through reasoning with a future requiring “intellectual gyms” because ordinary work no longer exercises cognition. The product distinction is whether AI helps users construct and test models—or merely supplies answers while their agency atrophies.
Deep dive
1. An outsider turned competing brain theories into an ordered evolutionary account
Bennett did not begin with an academic thesis. He accumulated notes out of “independent curiosity,” then used an entrepreneurial habit—asking what came first, second, and third—to organize a field whose major thinkers often seemed like blind observers describing different parts of an elephant. Scarfe supplied that “blind men and the elephant” framing; Bennett’s account was that he tried to make sense of disparate opinions by imposing an ordered structure.
His synthesis joins three disciplines: comparative psychology asks what different species can do; evolutionary neuroscience reconstructs the ordered modifications that produced modern brains; AI tests whether elegant biological ideas can be implemented. If a proposed principle cannot make an artificial system work, Bennett thinks that should at least make researchers question whether they have it right.
The outsider position carried obvious disadvantages, but also freedom to cross disciplinary boundaries. The book’s organizing wager is that brain evolution was not a random accumulation of faculties: successive structures enabled underlying computational breakthroughs that then appeared behaviorally in many different forms.
2. Sparse animal evidence and successful AI systems pull theory in opposite directions
Comparative psychology is far thinner than its confident narratives imply. Bennett’s example is the lamprey, a canonical proxy for early vertebrates: he knew of no direct study of its map-based navigation, even though teleost fish and reptiles perform it and the relevant homologous structures appear to be present.
Researchers therefore back into evolutionary claims from fragments—shared anatomy, nearby species, and plausible ecological value. The lamprey might recognize locations in three-dimensional space, but Bennett preserves the epistemic status: “This is all sort of, in some sense, guessing and trying to put the pieces together from very little information.”
The opposite problem appears between neuroscience and AI. Transformers, generative models, and reinforcement learning work without closely matching known brain mechanisms, while active inference offers rich explanations but little demonstrated use in high-performing AI systems. Bennett’s honest uncertainty: Karl Friston may lack the missing practical ingredient, or “there’s a breakthrough around the corner.”
3. The neocortex enables simulation without implementing the whole planning loop
Bennett makes the weaker and more defensible claim that adding the neocortex enables the overall system to perform mental simulation—not that every part of model-based reinforcement learning resides inside it. The distinction matters because planning is visibly distributed across older and newer structures.
The neocortex supplies a sufficiently rich model of the world to explore without current sensory input. But the thalamus and basal ganglia remain essential for pausing, representing intentions, selecting what to simulate, evaluating imagined results, and translating a chosen trajectory back into behavior.
This creates the unresolved search problem: possessing a generative world model does not tell an agent which counterfactuals deserve computation. “Fine, you can have a model of the world, but how do you prune the search space?” Bennett treats that selection mechanism as one of model-based reinforcement learning’s hardest questions.
Scarfe raised the familiar “geological strata” interpretation of the triune brain. Bennett rejects the popular claim that evolution simply stacked reptilian, limbic, and rational layers, while defending Paul MacLean as more qualified than later caricatures: reptiles have cortex with limbic-like functions, and old and new structures plainly interact rather than forming three clean layers.
4. Conscious perception is an inference, not a copy of sensory input
Nineteenth-century visual illusions supplied the initial clue. People see triangles, spheres, bars, or letters that are not actually drawn because the brain settles on the real-world object that best explains incomplete evidence rather than presenting raw pixels to consciousness.
Bennett traces this to Hermann von Helmholtz: the brain begins with a prior about what exists, checks incoming evidence against it, and retains the inferred world until contrary evidence becomes strong enough. Sensation informs perception, but “you are not receiving sensory input and experiencing the sensory input.”
The adaptive logic is straightforward. A mouse sees a moonlit branch, then advances into darkness; if its feet continue receiving compatible evidence, maintaining the branch model is safer than treating the branch as nonexistent merely because vision temporarily disappeared.
The same mechanism explains why an illusion can remain perceptually compelling after its trick is understood. The perceptual system keeps rendering the hypothesis supported by its learned generative model, making hallucination, dreaming, imagination, and ordinary perception variations on closely related machinery.
5. A generative brain renders one coherent world at a time
Ambiguous images reveal a second constraint: viewers can alternate between duck and rabbit, or between looking down on a staircase and looking upward beneath it, but cannot stably perceive both interpretations at once. The sensory pattern permits both; a physically coherent world does not.
Bennett’s explanation is that perception asks which real three-dimensional object could have produced the evidence. “It cannot be the case that the staircase is looking from above and below at the same time,” so the brain commits to one causally and spatially consistent rendering.
This fits Jeff Hawkins’s thousand-brains proposal. If many cortical columns maintain overlapping object models, the system needs to integrate or vote among them, producing “one sort of symphony of models” rather than exposing 15 incompatible renderings to decision machinery.
Generation follows naturally from inference: a model predicts incoming sensations, compares them with observations, and updates when prediction error crosses a threshold. Turn off the sensory stream and explore that same representation—rotate an imagined chair, recolor it, or inspect an absent scene—and perception has become simulation.
6. Simulation, not recognition, is the neocortex’s evolutionary prize
Bennett challenges the textbook emphasis on object recognition as the neocortex’s primary adaptive benefit. Fish can recognize human faces, while Scarfe notes that a fish cannot recognize an object when rotated in 3D space; Bennett sees no clean behavioral boundary separating neocortical animals from other vertebrates on recognition accuracy alone.
The stronger dividing line is learning by imagining. A rich generative model lets an animal explore an action it has never taken, estimate consequences before paying their real-world cost, and flexibly recombine knowledge in novel circumstances: “I can now imagine outcomes before having them.”
That also sharpens “understanding.” A binary classifier might identify a stapler, yet cannot answer what burning or opening it reveals, what it is for, or what a person holding it may do next. Bennett locates intuitive understanding in a model rich enough to be mentally explored and related to surrounding objects, agents, uses, and plausible futures.
7. Human intelligence extends through language, tools, and other minds
Scarfe argued that much intelligence sits outside the individual brain. Bennett’s cleanest proof is writing: biological memory compresses episodes and procedures reasonably well but handles semantic detail poorly, whereas writing externalizes effectively unbounded memories and carries them across generations.
Where to draw the intelligent system’s boundary then becomes philosophical. One can treat brains as the substrate and language as support—or imagine language itself evolving through brains, just as intelligence emerges across roughly 86 billion neurons without being attributed to any single neuron.
Bennett still prioritizes the brain as the richest physical object to reverse-engineer for AI and self-understanding. But the organism’s effective capability is undeniably relational: brains plus writing, tools, inherited knowledge, and conversations that expose their models to correction.
8. Predictive coding explains part of transformer success, not identity with the brain
Several observations support a neocortical generative model. Episodic recollection and imagining the future appear to use the same underlying process, while top-down cortical connections are far richer than simple feed-forward anatomy would predict—precisely what a hierarchy modulating lower representations should require.
AI’s success with self-supervision provides functional evidence for the broad principle. Mask portions of a large dataset and train a transformer to reconstruct or predict them; surprisingly general capabilities emerge without task-by-task labels.
Scarfe’s objection was architectural: biological neurons exchange messages with local autonomy, whereas transformer units appear to move together like “a Mexican wave” under matrix multiplication and backpropagation. He called prompt-driven systems a form of “agency smuggling,” because directedness originates with the user.
Bennett agreed that “the brain is not just one big transformer,” but preserved a possible analogy: attention heads may use context to dynamically reroute what the network cares about, effectively resetting its computation for each prompt. That is more interesting than a static feed-forward picture, though still far from capturing the brain’s recurrent, distributed machinery.
9. Active inference constructs agency beyond a single reward function
Bennett describes a live divide. Strict reinforcement-learning accounts try to reduce behavior to reward maximization; active-inference accounts add uncertainty reduction, self-prediction, and attempts to make experience conform to an internal model. He does not declare a winner: “It’s probably some balance of the two.”
In the active-inference framing, an agent builds a model of itself, infers goals by observing its own recurring behavior and internal state, then makes predictions that fulfill those constructed goals. Agency is therefore not merely a reward function handed down from outside.
Bennett’s favorite Friston formulation is “predictions, not commands.” Motor cortex might predict a bodily state rather than issue an imperative, while spinal circuitry makes the body satisfy that prediction—an alternative route to purposeful behavior without a single explicit reward signal.
Scarfe connected this to nested autonomous processes whose local directedness can scale into creativity and purpose. Bennett’s narrower point is computational: different paradigms instantiate “agency” differently, so using the word without specifying the mechanism conceals the central disagreement.
10. Rat brains visibly simulate futures and learn from paths not taken
In the 1940s or 1950s—Bennett could not recall which—Tolman noticed rats pause and sniff between alternatives at maze junctions. He called it “vicarious trial and error,” provoking skepticism because outward hesitation did not prove an internal simulation.
David Redish’s later recordings supplied the missing evidence. CA1 place cells normally fire at specific allocentric locations regardless of the route taken; during hesitation, their activity sweeps ahead down alternative maze paths rather than remaining at the rat’s current position. “You can literally watch rats imagining the future.”
In restaurant row, a sound tells a rat whether a flavored reward will arrive after roughly 3 seconds or require a 45-second wait. Because individual rats prefer foods such as banana over bland alternatives, moving onward creates irreversible trade-offs—and sometimes a choice that later looks worse.
When that happens, orbitofrontal activity represents the taste of the foregone option, and subsequent behavior changes: the rat becomes less likely to pass up the comparable offer next time. Bennett treats this as unusually direct evidence for counterfactual learning and model-based reinforcement learning in simple mammals.
11. Efficient intelligence must know both what to simulate and when to stop
The combinatorial problem is severe: even a good world model supports an intractable number of possible futures. Mammalian competence may therefore depend not only on simulation but on rapidly selecting a tiny set of candidate trajectories.
Bennett sees a clue in AlphaGo. Its policy network supplies a ranked first, second, third, and perhaps fourth move; search then plays forward from those promising candidates and may discover that the policy’s second choice wins more often than its first. Model-free judgment bootstraps model-based checking.
Biological agents face an extra problem because they cannot search on every move. Planning costs energy and real environments are noisy, so animals generally pause only when contingencies change or candidate actions are close enough to create high uncertainty.
Bennett’s speculative mechanism combines redundant cortical models with older gating circuitry. If parallel models broadly agree, action continues; if predictions diverge, the thalamus or basal ganglia might detect the mismatch and trigger simulation. He stresses that this is plausible, “far from conclusive,” and a genuine research frontier.
12. The first mammalian self-model turned internal states into inferred intentions
The agranular prefrontal cortex, found across mammals and largely believed to have existed in their earliest brains, receives interoceptive information: hypothalamic signals such as hunger and amygdala signals involving valence, fear, or danger.
It becomes especially active during uncertainty, planning, and episodic recall. Damage in rats severely impairs—and may eliminate—their ability to mentally simulate, making it a likely bridge between bodily need, remembered context, and flexible prospective action.
Bennett proposes that it explains the animal to itself. Observing a recurring pattern—this hypothalamic state followed by seeking water—it infers an intention such as thirst, much as posterior cortex infers a triangle as the cause of visual evidence. Neurons there consequently track tasks and progress toward goals, not merely movements.
13. Frontal layer-four atrophy may mark the shift from perception to volition
Most neocortex has six layers; layer four, containing granular cells, is the main recipient of thalamic sensory input. Agranular prefrontal and motor cortex are unusual because that layer is largely absent, while primates add a vast granular prefrontal region that retains it.
The developmental detail drives Friston’s interpretation: mammalian agranular cortex initially possesses layer four, then the layer atrophies rather than never forming. Early in life, the animal must absorb evidence to construct a self-model before relying on it to direct behavior.
A neocortical column can emphasize either inference—changing its model to fit sensation—or generation, starting from a latent model and predicting what should occur. Frontal cortex may increasingly occupy the latter regime as its account of “who I am and the things I would do” stabilizes.
Bennett calls the idea speculative but compelling: mature frontal cortex may spend less time fitting intent to observed behavior and more time fitting behavior to intent. In active-inference terms, it tries to “fit the world to its model.”
14. Rendered plans make goals explainable in a way habits are not
Strict reinforcement learning offers one ultimate goal: maximize reward, even if the momentary reward landscape changes. Active inference allows additional semantic levels—the abstract satiation of hunger, a selected endpoint, or the concrete sequence of driving to a particular restaurant.
If Bennett imagines the restaurant, chooses the terminating state, enters the car, and is asked why, the answer is available because the causal plan was explicitly rendered. That sequence gives prospective action a degree of explainability absent from an opaque value-maximizing reflex.
Ask why someone placed a foot at one point rather than two inches away during ordinary walking and there is no comparable plan to report; any answer is constructed afterward. Bennett therefore treats “goal” partly as a semantic choice, but the ability to select and execute a simulated trajectory as the load-bearing capability.
15. Evolutionary reconstruction turns messy anatomy into testable constraints
Bennett gives two reasons to reconstruct intelligence across roughly 600 million years. Human nature includes our histories as animals, vertebrates, mammals, and primates—not only the last 70,000 years—and evolutionary sequence is a useful tool for reverse-engineering a brain that natural selection built by tinkering.
Comparative anatomy and genetics can infer ancestral structures by locating homologous regions shared across surviving lineages. The method helps distinguish newly evolved circuitry from redundant, vestigial, or duplicated processing that would confuse a first-principles engineering interpretation.
Behavioral claims must satisfy three conditions: most descendants should display the capacity through homologous mechanisms; nearby outgroups should lack it or implement it independently; and the ancestral ecology should make its emergence adaptively plausible. Episodic memory in mammals, with apparently independent implementation in birds, illustrates the logic.
The reconstruction remains provisional because animal data are sparse. Yet Bennett’s motivating result is that abilities at each milestone cluster around one underlying intellectual breakthrough rather than a haphazard list—the basis of the book’s “five breakthroughs” account.
16. The ancient basal ganglia exposes reinforcement learning in anatomical form
Bennett calls the basal ganglia underappreciated. A lamprey separated from the human lineage by about 500 million years has a strikingly similar macrostructure, while its internal computation is more tractable and consensually understood than that of a neocortical column.
Its input mosaic contains D1 and D2 dopamine receptors. Dopamine strengthens D1-linked connections in a pathway that disinhibits behavior; a dopamine drop strengthens D2-linked stopping circuitry. Bennett’s delight is that one can anatomically watch positive signals reinforce “go” and negative outcomes reinforce “stop.”
Scarfe connected that machinery to habits and drug “wireheading”: repeated behavior can move down the control stack until stimuli automatically evoke it. Addiction is therefore not only a conscious preference but a deeply trained action-selection circuit.
A controversial Chinese intervention for intractable heroin addiction lesioned the nucleus accumbens and reportedly left about 40% recidivism. Bennett said it reduced cue-triggered cravings but produced side effects many doctors would consider unacceptable and “probably violated many ethical codes in the US.”
17. Primate brains expanded inside a political arms race
Scarfe floated calories, fruit access, extinction, and social complexity as possible drivers of primate brain growth. Bennett’s response kept the uncertainty intact: “We don’t know,” though social evidence is unusually strong.
Robin Dunbar found that, among primates, neocortical ratio is tightly correlated with group size—a relationship not generally observed across other mammals. The social-brain hypothesis therefore links the enlarged neocortex to the number of relationships an individual must track.
Primate groups are not loose herds. Their hierarchies are transitive and politically maintained: if one animal submits to a second and the second to a third, the first will generally submit to the third. Rank determines access and survival, but the strongest individual need not lead.
Alliances, grooming, reciprocal defense, mutiny, deception, and reputation reward social prediction over brute force. Correspondingly, primate-specific granular prefrontal cortex and posterior regions including the superior temporal sulcus and temporoparietal junction are heavily implicated in mentalizing—inferring another mind’s knowledge and intent.
18. Belle and Rock turned a spatial-memory test into Machiavellian strategy
Emil Menzel originally used a one-acre forest to test whether chimpanzees remembered hidden food locations. Belle did, and initially shared—but the aggressive, high-ranking Rock repeatedly took the food, changing the task from spatial navigation into social conflict.
Belle began concealing food by sitting on it; Rock pushed her aside. She then waited until he looked away; he responded by pretending not to watch, then racing toward her destination. She escalated again by leading him in the wrong direction—“deception and counter-deception” generated without experimenter instruction.
The reasoning is second-order: Belle must represent how her movement changes Rock’s belief, while Rock must represent her expectation that he is inattentive. Bennett regards the episode as an especially vivid case of theory of mind emerging from a competitive social loop.
Controlled studies reinforce it. Chimpanzees choose a box a human intentionally marked over one accidentally touched by the same marker, and after learning which goggles are transparent, solicit food from the experimenter who can see. Identical surface stimuli are interpreted through inferred intent and knowledge.
19. Mentalizing creates both deceptive risk and a possible alignment mechanism
Scarfe connected the chimpanzee arms race to Nick Bostrom’s instrumental convergence: an autonomous system asked to cure cancer might form a subgoal of controlling Earth’s labor and resources. Bennett agrees that more autonomy and freedom to invent subgoals create more risk, but rejects inevitability.
Evolution is a constrained search process with no moral preference. Natural deception and power-seeking do not become desirable merely because they were adaptive—the naturalistic fallacy—and deliberately engineered agents need not inherit every piece of human evolutionary baggage.
Bennett’s optimistic alignment route is itself mentalizing. Rather than obeying a request literally, an AI could infer the requester’s preferences, simulate how that person would evaluate possible outcomes, and recognize that converting Earth into paperclips is not what the person meant and would cause regret.
That requires rich constraints, well-defined objectives, or dependable models of human intent; none removes risk. Humans misread one another constantly, so passing a mentalizing benchmark would be a stabilizing tool, not proof of benevolence.
20. Status remains zero-sum even when technology makes material life positive-sum
Scarfe likened social media posting to deer locking horns: a predictive status contest that avoids literal combat. Primate rank made the game virtual—grooming, coalition, and reputation could outrank size—while humans multiplied it into success, virtue, dominance, and countless specialized arenas.
Drawing on The Elephant in the Brain, Bennett argues that much status-seeking is hidden even from the actor. Self-deception improves persuasion: sincerely believing one is helping the world can be more convincing than consciously admitting the reputational payoff.
Material welfare can improve for everyone; status is definitionally comparative and therefore scarce. Bennett worries about a permanent hedonic treadmill if more human energy shifts into ranking, though he rejects the claim that every motive is status or that people cannot organize around better virtues.
Organizational design changes the payoff. Distinct roles make a team feel non-zero-sum; a 30-person company can often coordinate through trust and common mission, while a 400-person firm develops factions. Militaries impose hierarchy, Google tolerates more chaos, and Amazon approximates autonomous internal startups with explicit interfaces.
21. Second-order metacognition explains explanations themselves
Bennett’s hierarchy begins with basal-ganglia selection: asked why it turned left, the system’s answer is effectively “because that maximized reward.” Agranular prefrontal cortex adds a self-level explanation—“because I am thirsty”—that can trigger alternative simulations satisfying the same inferred need.
Granular prefrontal cortex then builds “a simulation of the simulation.” It can explain that thirst triggered a search, remembered water was represented to the left, the route was simulated, and the predicted outcome caused selection.
This extra level permits an animal to swap the represented actor, knowledge, or intention: what would another individual simulate if they knew something different? Posterior multimodal regions model the rendered external world, while frontal machinery reasons over the model, creating something close to semantic and episodic knowledge.
Why not recurse through a third or fourth “why”? Bennett’s first-blush answer is energetic economics. A second level clearly paid for itself in primate political competition; further hierarchy may offer benefits too small to offset the substantial metabolic cost of additional cortex.
22. Frontal damage can spare IQ while removing the person from imagination
After World War II, patients with major granular prefrontal injuries created a puzzle. Unlike small visual, motor, or auditory lesions, which caused obvious deficits, substantial frontal damage could leave logic and IQ scores intact; one patient tested before and after surgery even improved.
Narrative tasks revealed the missing faculty. Given a word such as “restaurant,” people with hippocampal damage could describe themselves but supplied a thin external scene. People with granular prefrontal damage rendered leaves, smells, and surroundings richly, yet “they themselves were woefully missing from the stories.”
The region activates when people consider their own feelings, other minds, and self-reference—not merely what the weather looks like. Damage produces personality change, difficulty recognizing faux pas, and impaired reasoning about what another person thinks is appropriate.
The Sally–Anne test isolates false belief: Sally hides a marble, Anne moves it while Sally is absent, and the observer must predict where Sally will look. Children acquire this capacity gradually; macaques anticipate the falsely believed location, but the bias disappears when granular prefrontal activity is inhibited.
23. GPT-4 passes theory-of-mind puzzles without sharing the primate mechanism
Scarfe said GPT-3 performed terribly on false-belief tests, whereas GPT-4 reached remarkably accurate, roughly human-level performance. He also said that experiments varying the puzzles suggested it was not simply repeating identical examples from training data.
If theory of mind means only solving those questions, denying that GPT-4 has some model of human knowledge and behavior becomes difficult. Bennett’s response was that the stronger claim—that it mentalizes as humans do—does not follow.
Humans predict people partly by projecting from a similar internal architecture: “If I were in that situation, what would I do?” That shared mechanism provides a powerful prior and makes learning relatively data-efficient. A language model instead learns people from textual traces, more like humans model an external object from observed behavior.
The deployment question is generalization. Performance on familiar narrative puzzles does not establish that GPT-4 will infer what a person means while optimizing an unfamiliar paperclip factory, nor how much new data it would require to adapt. Bennett’s answer is deliberately nuanced, not a binary capability verdict.
24. Language is an evolved instinct, not merely what a larger brain does
Aristotle’s proposed human bright line was reason, but comparative psychology has successively found tool use, planning, mental time travel, self-like models, and forms of reasoning elsewhere. Bennett thinks language remains the most salient discontinuity.
Imperative labels connect a cue and rewarded response, as when a dog obeys a command. Declarative labels make a sound refer to a concept; grammar then makes arrangement meaningful, so “Ben hugged James” differs from “James hugged Ben” despite containing the same labels.
Bennett doubts that scaling a chimpanzee brain alone yields language. Homo floresiensis stood about 3½ to 4 feet tall and had a brain only marginally larger than a modern chimpanzee’s, yet showed signs of superior human intelligence, including tool use akin to that of ancestral humans. That suggests that a categorical adaptation survived brain shrinkage.
Human infants reveal the likely adaptation: they synchronize conversational turns before speaking and actively seek joint attention. A child remains dissatisfied if handed the pointed-at object without the parent looking; satisfaction arrives when parent and child attend together. Humans also ask questions and volunteer inner states in ways trained non-human primates rarely do.
25. Language lets minds learn from actions that occurred only in another mind
Bennett separates four learning sources. Animals learn from their own real actions; mammals add their own imagined actions; mentalizing primates learn from another’s observed actions; language unlocks the distinct human superpower of learning from “other people’s imagined actions.”
A witness can report that a blue snake’s bite was harmless while a red snake caused illness, distributing semantic knowledge to people absent from the event. A hunter can simulate a coordinated ambush, share it, and let companions challenge or revise the plan before anyone incurs its physical cost.
Language is therefore a compressed code intended to cue a simulation in another brain. Mentalizing may be prerequisite: listeners must infer what the speaker knows, intends, and means, then ask follow-up questions when the decoded rendering remains ambiguous.
Bennett favors communication over Noam Chomsky’s minority view that language first evolved for thought, while noting that language models intriguingly make language itself a reasoning medium. Either way, linguistic fidelity lets simulations accumulate rather than forcing each generation to rediscover them through direct experience.
26. Memes coordinate civilizations without selecting for truth or happiness
Bennett calls cumulative culture “the singularity that already happened.” Non-human primates transmit tools through imitation, but their innovations do not reliably compound over many generations; human simulations can be copied, combined, recorded, and improved. One can imagine a progression from “I know how to whittle a bone into a needle for sewing” to “now I’ve built a loom.”
Specialization expands collective memory beyond any brain. A group of 100 can distribute hunting, weaving, and other skills; writing preserves knowledge even when no living person holds it. Anthropological cases of isolated populations losing technology show that some capabilities require a minimum number of brains—hence Bennett’s thought experiment that 20 surviving friends would retain shockingly little civilization.
Memes are ideas or behaviors that propagate through this network and undergo selection. Shared fictions—money and individual rights—allow strangers to coordinate immediately, but virality can also exploit fear, low-probability catastrophe, surprise, or identity. A belief consistent with the self-model passes through a “porous filter”; a challenging one meets a gate.
Communication’s evolution still requires an individual-level reason not to lie. Reciprocal altruism supplies one candidate, while Robin Dunbar’s gossip account raises the cost of cheating by spreading one detected violation through the group. More broadly, Bennett insists that what survives memetically need not be true, moral, peaceful, or conducive to happiness.
27. External cognition can enlarge intelligence while quietly atrophying agency
Bennett calls modern people “epistemic hybrids.” Writing overcame memory limits; the internet made vast shared stores instantly queryable. Yet his own spatial model weakened after adopting Google Maps, while his father still reconstructs unfamiliar cities internally.
Scarfe sharpened the concern with automated lecture capture. Transcription and generated notes can become “understanding procrastination”: the learner postpones model construction and loses the speaker’s physical, social, visual, and performative cues, often without ever returning to do the harder cognitive work.
Bennett translated the distinction into reinforcement-learning terms. His father navigates model-based; a Maps user externalizes the world model and becomes a model-free actor responding to turn cues. Efficiency is real, but dangerous when the outsourced model would have supported broader reasoning and future learning.
An optimistic AI tutor, such as the Khan Academy direction Bennett mentioned, would guide students through intermediate reasoning rather than deliver the answer. The darker endpoint resembles physical modernity: after work stops exercising cognition, society may need “intellectual gyms,” just as sedentary people now run in place to replace vanished physical labor.
28. A true world model creates hypotheses and seeks the data needed to reject them
Geoffrey Hinton’s analog–digital distinction frames one technical frontier. Digital networks are effectively immortal because exact weights can be copied, but energy-intensive; biological analog networks embed knowledge in physical connections, receptors, and gene expression, making them efficient but not directly transferable.
Continual learning is another dividing line. Modern systems cannot safely incorporate every new interaction without disrupting prior representations—ChatGPT would become “rapidly dumber”—whereas human brains update continuously. Bennett expects impactful agents to require immediate adaptation without catastrophic forgetting.
Scarfe argued that GPT-4 plainly models aspects of reality well enough to answer complex questions. Bennett distinguished that from a world model: a world model supports ordered counterfactual states, causal interventions, and a loop in which the agent predicts an outcome, acts, observes the delta, and revises itself.
Rats and children actively manufacture informative training data by turning, touching, and testing novel objects; CNN developers must manually rotate images for them. Scarfe proposed placing false claims in training data and asking whether an agent can independently reject them. Sentience remains harder: indistinguishable outputs may defeat scientific discrimination while leaving unresolved moral differences that require philosophy, not benchmark scores.