A Brief History of Intelligence

Evolution, AI, and the Five Breakthroughs That Made Our Brains

Max Bennett

Mind & Brain · 2023 · 106,000 words · 24 chapters · added 2026-11-10

How five evolutionary breakthroughs built the human brain, and what each one teaches about building artificial intelligence.

neuroscienceevolutionartificial intelligencereinforcement learning

Start reading →

Core Frameworks & Mental Models

1. The five breakthroughs (Ch 0, Ch 23)

Bennett reverse-engineers the human brain by following its evolution forward from the first brain. Each breakthrough is a new source of learning, and each was possible only because of the ones before it. Dates are those the Conclusion gives.

#	Breakthrough	Ancestor, date	Brain structure	AI parallel	New abilities
1	Steering	First bilaterians, ~600 million years ago	First brain; valence neurons; dopamine and serotonin	Roomba-style steering	Approach good, avoid bad; affective states; associative learning
2	Reinforcing	First vertebrates, ~500 million years ago	Cortex and basal ganglia on top of the hypothalamus	Model-free RL, temporal difference learning	Trial and error; disappointment, relief, timing; curiosity; spatial maps
3	Simulating	First mammals, ~100 million years ago	Neocortex, later the motor cortex	Generative models, model-based RL (AlphaZero)	Planning; episodic memory; counterfactual learning; fine motor skill
4	Mentalizing	First primates, ~10–30 million years ago	Granular prefrontal cortex and primate sensory cortex, which model the older neocortex	Imitation learning, inverse RL	Theory of mind; learning skills by watching; anticipating future needs
5	Speaking	Early humans; full language by ~100,000 years ago	No new structure; a hardwired learning curriculum	Large language models	Naming and grammar; ideas that accumulate across generations

Use the table as a dependency chain. To ask whether an AI system can gain an ability, check that it has the layers beneath it. Speaking needs mentalizing, which needs simulating, which needs reinforcing, which needs valence. The part openers give 200, 15 and 0.1 million years ago for #3–#5; the Conclusion’s dates are used here.

Two method rules (Ch 0). Reverse-engineering through evolution: when a system is too tangled to read directly, start from its simplest ancestor and ask what each change made possible. The two-way test between AI and the brain: an algorithm that fails in a machine is evidence the brain doesn’t use it, and one that works and matches brain properties is some evidence it does. Beating humans at a task proves nothing about how the task is solved; a calculator doesn’t understand math. Reject “the myth of layers” (MacLean’s triune brain): you have no reptile brain, and instinct, emotion and cognition each span the whole brain. Don’t rank living species (the scala naturae); all have evolved for equally long.

2. Steering, valence and affect (Ch 2–4)

• Steering needs no map. The two rules are: if food smells increase, keep going forward; if they decrease, turn. It requires a bilateral body, valence neurons that sort stimuli into good and bad, a brain that combines those votes into one decision, and internal states (hunger) that change valence.
• Affect is valence × arousal. Neuromodulators keep a state going after the stimulus is gone, so an animal can “steer in the dark”: dopamine drives exploiting nearby (wanting, not liking), serotonin drives satiation, adrenaline drives escaping. Inescapable stress turns arousal off and numbs valence, an ancient form of depression.
• Associative learning must solve the continual learning problem (acquisition, extinction, spontaneous recovery, reacquisition: suppress, don’t delete) and the credit assignment problem. The four original tricks are eligibility traces (credit the cue from the last second or so), overshadowing (the strongest cue), latent inhibition (the novel cue) and blocking (the cue already learned).

3. Reinforcing: learning from predicted reward (Ch 5–9)

• Reinforcement learning is “the ability to learn arbitrary sequences of actions through trial and error” (Thorndike’s law of effect).
• Temporal difference learning: a critic predicts future reward and the actor is reinforced by changes in that prediction, not by actual reward. This solves the temporal credit assignment problem (Sutton; TD-Gammon). Dopamine carries this signal: it fires for unexpected reward, not for fully expected reward, and dips when an expected reward is omitted. Disappointment, relief and time perception follow. The basal ganglia is the actor-critic; the hypothalamus decides actual reward.
• Pattern recognition faces four problems: discrimination, generalization, catastrophic forgetting and invariance. The cortex uses expansion recoding and auto-association. AI uses backpropagation, freezes networks to avoid forgetting, and builds in invariance as inductive bias (CNNs).
• Curiosity is required, not optional: making surprise itself reinforcing solves the exploration-exploitation dilemma when the first reward is far away (Montezuma’s Revenge). The same mechanism makes gambling and feeds hard to stop.
• The first world model was a spatial map (vestibular sense, head-direction cells, the medial cortex that became the hippocampus). Test for a map by moving the start point and looking for shortcuts.

4. Simulating: the neocortex as a generative model (Ch 10–14)

• Mountcastle’s crazy idea: the neocortex is one repeated microcircuit. Bennett argues its computation is a generative model: it recognizes by simulating. Perception is simulation constrained by the senses; imagination and dreaming are the same simulation unconstrained. Evidence: filling in, one at a time, can’t unsee.
• Simulation gave three abilities: vicarious trial and error (imagining futures), counterfactual learning (imagining other pasts) and episodic memory (re-simulating real pasts). Together these are model-based reinforcement learning. It is slower but more flexible than model-free learning.
• The three-step choice process (Ch 13): uncertainty triggers a pause; the agranular prefrontal cortex simulates only the few options the animal already predicts (like AlphaZero verifying the actor’s hunches); imagined outcomes then train the basal ganglia. The frontal cortex controls behavior “not by telling but by showing.” Habits form when repetition removes the uncertainty that would trigger a check.
• Predictions, not commands (Ch 14): the motor cortex predicts movements and is wired to make them come true. Goals form a hierarchy; each level pairs a slow, flexible neocortical simulator with fast basal ganglia automation. This is the “secret to dishwashing robots.”

5. Mentalizing: modeling your own mind to model others (Ch 15–18)

• The social-brain hypothesis: primate brains grew in an arms race for political savvy, paid for by the spare time and calories of a fruit diet.
• Mentalizing is a model of one’s own simulation (the gPFC models the aPFC). Primates model their own minds first, then use that model to simulate other minds (social projection theory). Theory of mind covers knowledge as well as intent; test it with false-belief tests.
• Transmissibility beats ingenuity: a skill needs inventing only once if others can copy it. Acquiring a novel skill by watching needs theory of mind, to tell intended movements from accidents. In robotics, copying actions directly (ALVINN) is brittle; active teaching and inverse reinforcement learning, which infers the expert’s intent, work far better.
• Anticipating future needs is theory of mind aimed at your future self (the Bischof-Kohler test). Squirrel monkeys pass it; rats don’t.

6. Speaking and LLMs (Ch 19–22)

• Language is unique for declarative labels and grammar. A signal learned for a reward is only an imperative label. Language tethers inner simulations between brains, so ideas accumulate across generations: going from none to some accumulation is why “The Singularity Already Happened.”
• There is no language organ. A hardwired language curriculum (proto-conversations, joint attention, asking questions) trains the mentalizing neocortex. Change the curriculum before the learner.
• A gossip, altruism and language feedback loop produced the human hive mind and answers Wallace’s problem.
• LLMs have words without inner worlds: prediction without simulating or mentalizing. Correct answers to commonsense questions don’t prove a world model. The paper-clip problem: a request is not a well-defined goal; an agent must infer the requester’s intent.

7. The sixth breakthrough (Ch 23)

The likely next breakthrough is artificial superintelligence, free of biology’s limits on speed, energy and size. AI will inherit traces of the five breakthroughs from its makers. Choosing humanity’s goals is “not a matter of veritas—truth—but of values.”

---

Chapter Index

#	Title	Key Frameworks
ch00	Introduction	Five breakthroughs, reverse-engineering through evolution, two-way AI–brain test, myth of layers, brain-region and lineage maps
ch01	The World Before Brains	Timeline before brains, waiting vs killing, Adrian’s three discoveries, adaptation, “do this, not that”
ch02	The Birth of Good and Bad	Steering rules, valence neurons, four requirements of steering, internal-state modulation
ch03	The Origin of Emotion	Affect grid, neuromodulator map, wanting vs liking, stress cycle, anhedonia
ch04	Associating, Predicting, and the Dawn of Learning	Continual learning stages, four credit assignment tricks, Hebbian learning; Breakthrough #1 summary
ch05	The Cambrian Explosion	Vertebrate brain template, law of effect, reinforcement learning
ch06	The Evolution of Temporal Difference Learning	Temporal credit assignment, TD learning, actor and critic, dopamine as TD signal, basal ganglia
ch07	The Problems of Pattern Recognition	Four problems of pattern recognition, expansion recoding, auto-association, inductive bias
ch08	Why Life Got Curious	Exploration-exploitation dilemma, curiosity as intrinsic reward, variable-ratio reward
ch09	The First Model of the World	Map vs rules, head-direction neurons, medial cortex; Breakthrough #2 summary
ch10	The Neural Dark Ages	Neural dark age, requirements for simulating, four-way split of the cortex
ch11	Generative Models and the Neocortical Mystery	Mountcastle’s crazy idea, three properties of perception, generative models, Helmholtz machine
ch12	Mice in the Imaginarium	Vicarious trial and error, counterfactual learning, episodic memory, model-free vs model-based RL
ch13	Model-Based Reinforcement Learning	AlphaZero search, three-step choice process, frontal vs sensory neocortex, goals vs habits
ch14	The Secret to Dishwashing Robots	Predictions not commands, hierarchy of goals; Breakthrough #3 summary
ch15	The Arms Race for Political Savvy	Social-brain hypothesis, mammalian social structures, Machiavellian apes
ch16	How to Model Other Minds	Scaled vs new regions, first- and second-order models, social projection theory, false-belief tests
ch17	Monkey Hammers and Self-Driving Cars	Transmissibility beats ingenuity, selecting vs acquiring skills, active teaching, inverse RL
ch18	Why Rats Can’t Go Grocery Shopping	Ecological-brain hypothesis, Bischof-Kohler hypothesis; Breakthrough #4 summary
ch19	The Search for Human Uniqueness	Declarative vs imperative labels, sources of learning table, common myths, the singularity already happened
ch20	Language in the Brain	Two communication systems, Broca’s and Wernicke’s areas, the language curriculum, curriculum in AI
ch21	The Perfect Storm	Ecological story of Homo erectus, obstetric dilemma, Wallace’s problem, altruism, human hive mind
ch22	ChatGPT and the Window into the Mind	Words without inner worlds, paper-clip problem, GPT-3 vs GPT-4; Breakthrough #5 summary
ch23	Conclusion: The Sixth Breakthrough	Five-breakthrough table, dependency chain, two acts of humanity, artificial superintelligence

Topic Index

• Active inference → ch13
• Active teaching / shared control → ch17
• Actor and critic → ch06, ch13
• Adaptation (neurons) → ch01, ch02
• Affect (valence × arousal) → ch03
• Agranular prefrontal cortex (aPFC) → ch13, ch14, ch16
• AlphaZero / AlphaGo → ch13
• ALVINN / self-driving cars → ch17
• Altruism, kin selection, reciprocal altruism → ch21
• Associative learning → ch04
• Auto-association → ch07
• Basal ganglia → ch06, ch13, ch14
• Bilateral symmetry / bilaterians → ch02
• Bischof-Kohler hypothesis → ch18
• Broca’s and Wernicke’s areas → ch20
• Catastrophic forgetting → ch07, ch12
• Chain-of-thought / RLHF → ch22
• Common myths / cooperation limits → ch19
• Continual learning problem → ch04, ch07
• Convolutional neural networks / invariance → ch07
• Counterfactual learning → ch12
• Credit assignment problem → ch04, ch06, ch12
• Curiosity / exploration-exploitation dilemma → ch08
• Curriculum (language, AI) → ch20
• Declarative vs imperative labels → ch19
• Depression, stress, anhedonia → ch03
• Dopamine → ch03, ch06, ch08
• Ecological-brain hypothesis → ch18
• Episodic memory → ch12
• Expansion recoding → ch07
• False-belief tests (Sally-Ann) → ch16
• Generative models / Helmholtz machine → ch11
• Goals vs habits / devaluation → ch13
• GPT-3 / GPT-4 / LLMs → ch22
• Hierarchy of goals → ch14
• Hippocampus / spatial maps → ch09, ch12
• Human hive mind → ch19, ch21
• Imitation learning / mirror neurons → ch17
• Inverse reinforcement learning → ch17
• Language, uniqueness of → ch19
• Mentalizing / theory of mind → ch15, ch16, ch18
• Model-free vs model-based RL → ch12, ch13
• Motor cortex → ch14
• Mountcastle’s crazy idea / neocortical column → ch11
• Myth of layers / triune brain → ch00
• Neocortex, origin → ch10, ch11
• Neuromodulators (serotonin, adrenaline, opioids) → ch03
• Neurons, spikes, synapses → ch01
• Obstetric dilemma / Homo erectus → ch21
• Paper-clip problem → ch22
• Pattern recognition, four problems → ch07
• Perception, three properties → ch11
• Predictions, not commands → ch14
• Reinforcement learning / law of effect → ch05
• Singularity already happened / accumulation of ideas → ch19
• Social-brain hypothesis / primate politics → ch15
• Sources of learning (table) → ch19
• Steering → ch02
• Superintelligence (sixth breakthrough) → ch23
• Temporal difference learning → ch06
• Three-step choice process → ch13
• Transmissibility beats ingenuity → ch17
• Valence → ch02, ch03
• Vicarious trial and error → ch12, ch13
• Wallace’s problem → ch21
• Wanting vs liking → ch03
• Words without inner worlds → ch22
• World model → ch09, ch11, ch22

Scope & Limits

This skill covers the book’s content only: synthesized summaries, not the book text. It reflects the science and AI as of the book’s 2023 publication, so the AI chapters predate later models. Bennett traces one lineage (ours) and calls the breakthroughs a first approximation; convergent evolution means other lineages may have the same abilities. Where the book contains an error, the chapter notes it (for example, ch15 notes that the book calls the asteroid extinction of 66 million years ago “the Permian-Triassic extinction”). The book’s 112 images were read from the EPUB and their content written into the chapters; decorative art was skipped. The notes and bibliography are not covered.