Introduction

Core Idea

AI can beat any human at chess but cannot load a dishwasher better than a six-year-old. Bennett argues that we cannot explain gaps like this because we don’t yet understand the thing we are trying to re-create: human intelligence. The human brain is too complex to reverse-engineer by looking at it directly. So the book traces how the brain evolved, starting from the simplest brains and following the human lineage forward. It summarizes that history as five breakthroughs, each built on the ones before, and ties each breakthrough to what AI research has learned.

Frameworks Introduced

• The five breakthroughs (the organizing map of the book; one part per breakthrough):
  1. Steering (first bilaterians, about 600 million years ago). Animals sort stimuli into good and bad and turn toward good things and away from bad things. The first brain appears. (Ch 2–4)
  2. Reinforcing (first vertebrates, about 500 million years ago). Animals learn to repeat behaviors that led to positive valence and suppress those that led to negative valence. The AI parallel is model-free reinforcement learning. (Ch 5–9)
  3. Simulating (first mammals; the part opens at 200 million years ago). The neocortex lets animals simulate actions and stimuli before acting, which supports planning, episodic memory and counterfactual learning. (Ch 10–14)
  4. Mentalizing (first primates; the part opens at 15 million years ago). Animals model their own minds, which lets them model the intent and knowledge of others and learn by watching. (Ch 15–18)
  5. Speaking (first humans; the part opens at 100,000 years ago). Naming and grammar let humans link their inner simulations and pass thoughts across generations. (Ch 19–22)
  • When to use: as a map for placing any brain structure, ability, or AI technique. Ask which breakthrough it belongs to and which earlier breakthroughs it depends on.
  • Why it works: each breakthrough was built on the foundation of earlier ones, so the order explains the complexity. Failure mode: it is a first approximation of the human lineage only. It says nothing about which modern species is smarter.
• Reverse-engineering through evolution:
  • When to use: when a system is too complex to understand by inspecting it in its current form.
  • How: start from the simplest ancestral version, then follow the changes forward and ask which new ability each change made possible. Reconstruct ancestral brains from fossil skulls and by comparing living animals. A feature shared by all descendants of an ancestor was probably present in that ancestor, so each living brain is “a time capsule” of earlier minds.
• The two-way test between AI and the brain:
  • When to use: when judging a claim about how the brain computes something.
  • How: if an algorithm that we think the brain uses fails when built in a machine, that is evidence the brain does not work that way. If an algorithm works well in AI and its properties match properties of animal brains, that is some evidence the brain does work that way.
  • Why it works: building a thing tests whether you understand it. Failure mode: outperforming humans at a task does not show that an AI solves the task the way humans do. A calculator crunches numbers faster than a human, but that does not mean it understands math.
• Breakthroughs come from pressure: each breakthrough emerged when ancestors faced extreme situations or were caught in strong feedback loops (predators, environmental disasters, desperate niches). To understand a brain change, also study the survival problem that the ancestor faced at that time.

Key Concepts

• Rosey: the robot from The Jetsons (1962). The show predicted video calls, cell phones and smartwatches correctly, but nothing like Rosey exists. She stands for human-level intelligence: reasoning, common sense and motor skill, plus empathy and social skill.
• Brain complexity: 86 billion neurons and over 100 trillion connections, each less than 30 nanometers wide, with over one billion connections per cubic millimeter. Hundreds of different chemicals pass across connections, and the wiring constantly changes. A wiring map alone would not explain the brain.
• Similarity across species: a chimpanzee brain differs from ours in little besides size, a rat brain by a handful of modifications, and a fish brain has almost all the same structures.
• The first brain: a collection of neurons in the head of a worm the size of a grain of rice, about 600 million years ago. This worm was the ancestor of all modern animals that have brains.
• Triune Brain Hypothesis (Paul MacLean, 1960s): the claim that the brain has three layers. A reptile brain handles survival instincts, the limbic system on top handles emotion, and the neocortex on top of that handles cognition. Figure 1 draws these as nested shells. Bennett calls it “The Myth of Layers.”
• Scala naturae (“scale of nature”): Aristotle’s ranking of life, with humans above other mammals, then reptiles and fish, then insects, then plants.
• Convergent evolution: lineages that evolve the same solution independently. Wings evolved separately in insects, bats and birds, and eyes probably evolved many times. An ability that the book places in early mammals, such as episodic memory, may also have evolved in other lineages.
• The business-versus-machine-learning gap: business teams propose AI tasks that seem simple only because they are simple for human brains. Machine learning teams then explain why the task is “astronomically difficult.” These meetings, at the author’s company Bluecore, started his interest in the brain.

Brain regions (front-matter figure). The outer view shows the neocortex, divided into granular prefrontal areas at the front, then motor areas, somatosensory areas, auditory areas on the side, and visual areas at the back. Below it are the cerebellum and the brainstem. The inner view shows, inside the neocortex, the thalamus, basal ganglia, hypothalamus, hippocampus and amygdala.

Lineage with dates (front-matter figure). The bold line of our lineage runs: first animals (before 640 million years ago), then first bilaterians, first vertebrates, first mammals, first primates, and Homo sapiens. The order in which other groups branch off is anemones and jellyfish first; then nematodes, insects, crustaceans, octopuses and squids, and slugs; then fish and sharks, amphibians, and reptiles and birds; then among mammals kangaroos and koalas, elephants, dolphins and whales, horses and deer, dogs and lions, rodents, monkeys, and chimpanzees. The time axis runs from 640 million years ago (Ediacaran) through the Cambrian (from 540), Devonian (from 420), Permian (from 300), Jurassic (from 200) and Cretaceous (from 150) to 66 million years ago and now.

Mental Models

• Use the evolutionary story when the present-day system is too tangled to read. The human brain may be “the last place to look.”
• Think of AI and neuroscience as checking each other: the brain inspires AI designs, and AI tests our theories of the brain.
• When a task seems easy, ask whether it is easy in general or only easy for a human brain.
• In the eyes of evolution, the hierarchy has only two rungs: species that survived and species that did not.

Anti-patterns

• The myth of layers: you do not have a reptile brain in your head. Reptiles have their own version of a limbic system. Instinct, emotion and cognition come from networks that span all three supposed layers. Evolution did not stack new systems on top of old ones without changing the old ones.
• Building a scala naturae: all living species have been evolving for the same length of time. Most niches are better served by smaller and simpler brains, or no brain at all. Rankings by a single skill depend on the skill. An octopus outdoes humans at multitasking, and pigeons, chipmunks, tuna and iguanas process visual information faster than humans.
• Human chauvinism: telling a human-centered story is fine. Treating it as proof that humans are superior is not. Figure 2 contrasts the two readings: the book traces one path down the tree (“THIS”) and does not line up whale, human, bird, octopus and honeybee for comparison (“NOT THIS”).
• Trusting forecasts of human-level AI: in the 1960s Marvin Minsky predicted human-level AI in three to eight years. In the 1980s BusinessWeek declared “AI: it’s here.” Both were followed by stalls, so treat today’s claims that the “game is over” with caution.

Rules of Thumb (author’s wording)

• “Nothing in biology makes sense except in the light of evolution.” (Theodosius Dobzhansky)
• “What I cannot create, I do not understand.” (Richard Feynman)
• “The brain is our guiding inspiration for how to build AI, and AI is our litmus test for how well we understand the brain.”
• “In evolution, systems start simple, and complexity emerges only over time.”
• “Past innovations enabled future innovations.”
• “All animals alive today have been undergoing evolution for the same amount of time.”

Worked Example

Why the triune brain fails as a framework.

1. MacLean proposed three stacked layers: a reptile brain (survival instincts such as aggression and territoriality), then a limbic system (emotions such as fear, parental attachment, sexual desire and hunger), then the neocortex (language, abstraction, planning and perception).
2. The model predicts that reptiles have only the reptile brain, that rats and rabbits add the limbic system, and that only humans have all three.
3. The anatomy is wrong. Reptile brains also contain their own version of a limbic system.
4. The functional split is wrong. Survival instincts, emotions and cognition come from many systems spread across all three layers.
5. The evolutionary story is wrong. Evolution changed existing systems as it added new ones.
6. Even if it had been right, the three functions are too broad and too vague to give a starting point for reverse-engineering intelligence.
7. Conclusion: a useful evolutionary framework must match current anatomy and also current knowledge of how intelligence works, which means using AI. The five breakthroughs are Bennett’s replacement.

Key Takeaways

1. Gaps between AI and human intelligence come from not understanding human intelligence, so study how it came to be.
2. Brains across species are similar because of shared ancestry, which lets us reconstruct ancestral brains.
3. Human brain evolution can be summarized as five ordered breakthroughs: steering, reinforcing, simulating, mentalizing and speaking.
4. Use AI as a test of brain theories, and use the brain as a source of AI ideas.
5. Reject the triune brain and any ladder of species. The book traces one lineage and does not rank species.

Connects To

• Ch 1: the world before brains and the universal features of neurons.
• Ch 2–4: breakthrough #1, steering, and the first brain.
• Ch 14: “The Secret to Dishwashing Robots” returns to the Rosey and dishwasher problem.
• Ch 22: ChatGPT and large language models, which the introduction raises with GPT-3 and GPT-4.
• Ch 23: the conclusion summarizes all five breakthroughs and looks at a sixth.