The brain is our guiding inspiration for how to build AI, and AI is our litmus test for how well we understand the brain.

The business teams look for applications of new AI systems that would be valuable, while only the machine learning teams understand what applications would be feasible.

Businesspeople probe for applications of AI systems that seem straightforward to them. But frequently, these tasks seem straightforward only because they are straightforward for our brains. Machine learning people then patiently explain to the business team why the idea that seems simple is, in fact, astronomically difficult. And these debates go back and forth with every new project.

Although these self-replicating DNA-like molecules also succumbed to the destructive effects of entropy, they didn’t have to survive individually to survive collectively—as long as they endured long enough to create their own copies, they would, in essence, persist.

The story of life, at its core, is as much about energy as it is about entropy.

Before this time, the Earth had no ozone layer. It was the cyanobacteria, with their newfound photosynthesis, that constructed Earth’s oxygen-rich atmosphere and began to terraform the planet from a gray volcanic rock to the oasis we know today.

oxygen can also be useful. This newly available element presented an energetic opportunity, and it was only a matter of time before life stumbled onto a way to exploit it. A new form of bacteria emerged that produced energy not from photosynthesis but from cellular respiration—the process by which oxygen and sugar is converted into energy, spewing out carbon dioxide as exhaust.

Respiratory microbes differed in one crucial way from their photosynthetic cousins: they needed to hunt. And hunting required a whole new degree of smarts.

Sugar is produced only by life, and thus there are only two ways for large multicellular respiratory organisms to feed. One is to wait for life to die, and the other is to catch and kill living life.

Fungal spores are all around us, patiently waiting for something to die. Fungi are currently, and likely have always been, Earth’s garbage collectors.

Gastrulation, neurons, and muscles are the three inseparable features that bind all animals together and separate animals from all other kingdoms of life.

A neuron is either on or off; there is no in between.

senses—you can discriminate between different volumes of sound, brightness of light, potency of smells, severity of pain. How could a simple binary signal that was either on or off communicate a numerical value, such as the graded strengths of stimuli?

The idea is that neurons encode information in the rate that they fire spikes, not in the shape or magnitude of the spike itself.

Touch-sensitive neurons encode pressure in their firing rate; photosensitive neurons encode contrast in their firing rate; smell-sensitive neurons encode concentration in their firing rate. The neural code for movement is also a firing rate: the faster the spikes of the neurons that stimulate muscles, the greater the contraction force of the muscles.

Neurons do not have a fixed relationship between natural variables and firing rates. Instead, neurons are always adapting their firing rates to their environment; they are constantly remapping the relationship between variables in the natural world and the language of firing rates.

This applies in many neurons throughout the brains of animals—the stronger the stimuli, the greater the change in the neural threshold for spiking.

system. Bilateral symmetry allows a movement apparatus to be optimized for a single direction (forward) while solving the problem of navigation by adding a mechanism for turning.

There is another observation about bilaterians, perhaps the more important one: They are the only animals that have brains. This is not a coincidence. The first brain and the bilaterian body share the same initial evolutionary purpose: They enable animals to navigate by steering. Steering was breakthrough #1.

If food smells increase, keep going forward. If food smells decrease, turn. This was the breakthrough of steering.

And so the breakthrough that came with the first brain was not steering per se, but steering on the scale of multicellular organisms.

In 1990, Brooks cofounded a robotics company named iRobot, and in 2002, he introduced the Roomba, a vacuum-cleaner robot. The Roomba was a robot that autonomously navigated around your house vacuuming the floor.

The first Roomba and the first bilaterians share a surprising number of properties. They both had extremely simple sensors—the first Roomba could detect only a handful of things, such as when it hit a wall and when it was close to its charging base. They both had simple brains—neither used the paltry sensory input they received to build a map of their environment or to recognize objects. They both were bilaterally symmetric—the Roomba’s wheels allowed it to go forward and backward only.

Brooks built the simplest possible robot, one that contained hardly any sensors and that computed barely anything at all. But the market, like evolution, rewards three things above all: things that are cheap, things that work, and things that are simple enough to be discovered in the first place.

In nematodes, sensory neurons don’t signal objective features of the surrounding world—they encode steering votes for how much a nematode wants to steer toward or away from something.

But it seems that the first brains began with sensory neurons that didn’t care to measure objective features of the world and instead cast the entirety of perception through the simple binary lens of valence.

These neurons will generate a similar number of spikes whether a smell concentration goes from two to four parts or from one hundred to two hundred parts. This enables valence neurons to keep nudging the nematode in the right direction. It is the signal for Yes, keep going! from the first whiff of a faraway food smell all the way to the food source.

Figure 3.1: The affective states of humans

The defining feature of these affective states is that, although often triggered by external stimuli, they persist for long after the stimuli are gone. This feature of affective states stretches all the way from nematodes to humans—just as a nematode remains in a fear-like state for many minutes after a single sniff of a predator, a human mood can be soured for hours after a single unfriendly social interaction.

What happens when you see something you want, like food when you’re hungry, a sexy mate, the finish line at the end of a race? In all cases, your brain releases a burst of dopamine. What happens when you get something you want, like when you’re orgasming, eating delicious food, or just finishing a task on your to-do list? Your brain releases serotonin.

To the surprise of many, Berridge found that increasing dopamine levels in the brains of rats had no impact on the degree and frequency of their pleasurable facial expressions to food.

Rats experienced pleasure just fine without dopamine—they just didn’t seem motivated to pursue it.

Dopamine is not a signal for pleasure itself; it is a signal for the anticipation of future pleasure.

Berridge proved that dopamine is less about liking things and more about wanting things.

serotonin is the satiation, things-are-okay-now, satisfaction chemical, designed to turn off valence responses.

Adrenaline not only triggers the behavioral repertoire of escape; it also turns off a swath of energy-consuming activities to divert energetic resources to muscles.

As far back as the first brains six hundred million years ago, the system for binge eating after a stressful experience was already put in place.

But most of the ways that stress plagues modern humanity comes from what happens to bodies in response to prolonged stressors—the chronic stress response.

This surprising behavior is, in fact, quite clever: spending energy escaping is worth the cost only if the stimulus is in fact escapable.

Chronic stress isn’t all that different from acute stress; stress hormones and opioids remain elevated, chronically inhibiting digestion, immune response, appetite, and reproduction. But chronic stress differs from acute stress in at least one important way: it turns off arousal and motivation.

It turned out that most, if not all, reflexes build such associations. Pair an arbitrary sound with an electric shock to your hand, and soon your hand will retract to just the sound.

The defining feature of Pavlov’s conditional reflexes is that they are involuntary associations;

The first brains needed a mechanism to not only acquire associations but also quickly change these associations to match the changing rules of the world.

Why do associations show spontaneous recovery and reacquisition? Consider the ancient environment in which associative learning evolved. Suppose a worm has many experiences of finding food alongside salt. And then one day, it detects salt, steers toward it, and finds no food. After the worm spends an hour sniffing around without finding food, the association becomes extinguished, and the worm begins steering toward other cues, no longer attracted to salt. If two days later it detects salt again, would it be smarter to steer toward or away from it? In all of the worm’s past experiences, except the most recent one, when it smelled salt it also found food. And so the smarter choice would be to steer toward salt again—the most recent experience may have been a fluke.

This is the benefit of spontaneous recovery—it enables a primitive form of long-term memory to persist through the tumult of short-term changes in the contingencies of the world.

Spontaneous recovery and reacquisition enabled simple steering brains to navigate changing associations, temporarily suppress old associations that were currently inaccurate, and remember and relearn broken associations that became effective again.

The third trick was latent inhibition—stimuli that animals regularly experienced in the past are inhibited from making future associations. In other words, frequent stimuli are flagged as irrelevant background noise.

From the bilaterian brain onward, the evolution of learning was primarily a process of finding new applications of preexisting synaptic learning mechanisms, without changing the learning mechanisms themselves.

Learning was not the core function of the first brain; it was merely a feature, a trick to optimize steering decisions. Association, prediction, and learning emerged for tweaking the goodness and badness of things.

law of effect: Responses that produce a satisfying effect in a particular situation become more likely to occur again in that situation, and responses that produce a discomforting effect become less likely to occur again in that situation.

The second breakthrough was reinforcement learning: the ability to learn arbitrary sequences of actions through trial and error.

The reinforcements and punishments in a game of checkers—the outcome of winning or losing—occur only at the end of the game. A game can consist of hundreds of moves. If you win, which moves should get credit for being good? If you lose, which moves should get credit for being bad?

The signal on which the actor learns is not rewards, per se, but the temporal difference in the predicted reward from one moment in time to the next. Hence Sutton’s name for his method: temporal difference learning.

While the presentation of an unexpected reward increases dopamine activity, the omission of an expected reward decreases dopamine activity.*

Dopamine is not a signal for reward but for reinforcement. As Sutton found, reinforcement and reward must be decoupled for reinforcement learning to work. To solve the temporal credit assignment problem, brains must reinforce behaviors based on changes in predicted future rewards, not actual rewards.

Indeed, it makes sense that dopamine was the neuromodulator that evolution reshaped into a temporal difference learning signal, as the signal for nearby rewards it was the closest thing to a measure of predicted future reward. And so, dopamine was transformed from a good-things-are-nearby signal to a there-is-a-35 percent-chance-of-something-awesome-happening-in-exactly-ten-seconds signal.

Both disappointment and relief are emergent properties of a brain designed to learn by predicting future rewards.

the omission of an expected punishment is itself reinforcing; it is relieving. And the omission of an expected reward is itself punishing; it is disappointing.

TD learning, disappointment, relief, and the perception of time are all related. The precise perception of time is a necessary ingredient to learn from omission, to know when to trigger disappointment or relief, and thereby to make TD learning work.

The basal ganglia learns to repeat actions that maximize dopamine release. Through the basal ganglia, actions that lead to dopamine release become more likely to occur (the basal ganglia ungates those actions), and actions that lead to dopmaine inhibition become less likely to occur (the basal ganglia gates those actions). Sound familiar? The basal ganglia is, in part, Sutton’s “actor”—a system designed to repeat behaviors that lead to reinforcement and inhibit behaviors that lead to punishment.

mistake before the hypothalamus gives any feedback. This is why dopamine neurons initially respond when rewards are delivered, but over time shift their activation toward predictive cues. This is also why receiving a reward that you knew you were going to receive doesn’t trigger dopamine release; predictions from the basal ganglia cancel out the excitement from the hypothalamus.

Rather, it is recognizing a particular soup of many molecules that activates a symphony of olfactory neurons. Any given smell is represented by a pattern of activated olfactory neurons. In summary, smell recognition is nothing more than pattern recognition.

You can distinguish the ramblings of a person from the roar of a panther based on the pattern of sound waves hitting your inner ear. And, yes, you can distinguish the smell of a rose from the smell of chicken based on the pattern of olfactory neurons activated in your nose. For hundreds of millions of years, animals were deprived of this skill, stuck in a perceptual prison.

Within the small mosaic of only fifty types of olfactory neurons lived a universe of different patterns that could be recognized. Fifty cells can represent over one hundred trillion patterns.*

Pattern recognition is hard. Many animals alive today, even after another half billion years of evolution, never acquired this ability—the nematodes and flatworms of today show no evidence of pattern recognition.

This was the first problem of pattern recognition, that of discrimination: how to recognize overlapping patterns as distinct.

This is the second challenge of pattern recognition: how to generalize a previous pattern to recognize novel patterns that are similar but not the same.

The next time a pattern shows up, even if it is incomplete, the full pattern can be reactivated in the cortex. This trick is called auto-association; neurons in the cortex automatically learn associations with themselves. This offers a solution to the generalization problem—the cortex can recognize a pattern that is similar but not the same.

It is only when knowledge is represented in a pattern of neurons, like in artificial neural networks or in the cortex of vertebrates, that learning new things risks interfering with the memory of old things.

It is also not a coincidence that pattern recognition and reinforcement learning evolved simultaneously in evolution. The greater the brain’s ability to learn arbitrary actions in response to things in the world, the greater the benefit to be gained from recognizing more things in the world.

The approach is to make AI systems explicitly curious, to reward them for exploring new places and doing new things, to make surprise itself reinforcing. The greater the novelty, the larger the compulsion to explore it.

In vertebrates, surprise itself triggers the release of dopamine, even if there is no “real” reward.

One explanation for this is that vertebrates get an extra boost of reinforcement when something is surprising. To make animals curious, we evolved to find surprising and novel things reinforcing, which drives us to pursue and explore them. This means that even if the reward of an activity is negative, if it is novel, we might pursue it anyway.

Gambling and social feeds work by hacking into our five-hundred-million-year-old preference for surprise, producing a maladaptive edge case that evolution has not had time to account for.

This trick, the ability to construct an internal model of the external world, was inherited from the brains of first vertebrates.

It was also the first time a brain differentiated the self from the world. To track one’s location in a map of space, an animal needs to be able to tell the difference between “something swimming toward me” and “me swimming toward something.”

And most important, it was the first time that a brain constructed an internal model—a representation of the external world.

If the reinforcement-learning early vertebrates got the power of learning by doing, then early mammals got the even more impressive power of learning before doing—of learning by imagining.

This meant that a side effect of warm-bloodedness was that mammal brains could operate much faster than fish or reptile brains.

And birds are, conspicuously, the only nonmammal species alive today that independently evolved warm-bloodedness.

According to Mountcastle, the neocortex does not do different things; each neocortical column does exactly the same thing. The only difference between regions of neocortex is the input they receive and where they send their output; the actual computations of the neocortex itself are identical. The only difference between, for example, the visual cortex and the auditory cortex is that the visual cortex gets input from the retina, and the auditory cortex gets input from the ear.

Remarkably, the ferrets could see just fine. And when researchers recorded the area of the neocortex that was typically auditory but was now receiving input from the eyes, they found the area responded to visual stimuli just as the visual cortex would. The auditory and visual cortices are interchangeable.

if Mountcastle’s theory is correct, it suggests that the neocortical column implements some algorithm that is so general and universal that it can be applied to extremely diverse functions such as movement, language, and perception across every sensory modality.

No matter which conversation you tune in to, the auditory input into your ear is identical; the only difference is what your brain infers from that input. You can perceive only a single conversation at a time.

Your mind likes to have an interpretation that explains sensory input. Once I give you a good explanation, your mind sticks to it. You now perceive a frog.

He suggested that a person doesn’t perceive what is experienced; instead, he or she perceives what the brain thinks is there—a process Helmholtz called inference. Put another way: you don’t perceive what you actually see, you perceive a simulated reality that you have inferred from what you see.

Helmholtz suggested that much of human perception is a process of inference—a process of using a generative model to match an inner simulation of the world to the sensory evidence presented.

Some neuroscientists refer to perception, even when it is functioning properly, as a “constrained hallucination.”

The neocortex (and presumably the bird equivalent) is always in an unstable balance between recognition and generation, and during our waking life, humans spend an unbalanced amount of time recognizing and comparatively less time generating.

The most obvious feature of imagination is that you cannot imagine things and recognize things simultaneously. You cannot read a book and imagine yourself having breakfast at the same time—the process of imagining is inherently at odds with the process of experiencing actual sensory data.

The neocortex seems to be in a continuous state of predicting all its sensory data.

Different subregions of neocortex simulate different aspects of the external world based on the input they receive.

Put all these neocortical columns together, and they make a symphony of simulations that render a rich three-dimensional world filled with objects that can be seen, touched, and heard.

But why do this? What is the point of rendering an inner simulation of the external world? What value did the neocortex offer these ancient mammals?

This was the gift the neocortex gave to early mammals. It was imagination—the ability to render future possibilities and relive past events—that was the third breakthrough in the evolution of human intelligence.

This reveals one of the benefits of vicarious trial and error: once a rat has a world model of their environment, they can rapidly mentally explore it until they find a way to get around obstacles to get what they want.

In the ancient world, and in much of the world that followed, such ruminating was useful because often the same situation would recur and a better choice could be made.

The type of reinforcement learning we saw in early vertebrates has a flaw: It can only reinforce the specific action actually taken. The problem with this strategy is that the paths that were actually taken are a small subset of all the possible paths that could have been taken. What are the chances that an animal’s first attempt picked the best path?

If you saw lightning strike a dry forest, and a fire immediately started, you would say that it was the lightning that caused the fire, not the fire that caused the lightning. You say this because when you imagine the counterfactual case in which lightning did not strike, the fire did not appear. Without counterfactuals, there is no way to distinguish between causation and correlation. You can never know what caused what; you can know only that “X always happens before Y” or “Whenever X happens, Y happens” or “Y has never happened without X happening,” and so on.

As a reminder, the credit assignment problem is this: When some important event occurs that you want to be able to predict ahead of time, how do you choose what previous actions or events to give credit for having been predictive of the event?

Causation itself may live more in psychology than in physics. There is no experiment that can definitively prove the presence of causality; it is entirely immeasurable.

The laws of physics may contain rules of how features of reality progress from one time step to the next without any real causal relationships between things at all.

Causation is constructed by our brains to enable us to learn vicariously from alternative past choices.

But here is the weird thing—we don’t truly remember episodic events. The process of episodic remembering is one of simulating an approximate re-creation of the past.

When imagining future events, you are simulating a future reality; when remembering past events, you are simulating a past reality. Both are simulations.

This is why episodic memories feel so real but are much less accurate than we think.

studies show that repeatedly imagining a past event that did not occur falsely increases a person’s confidence that the event did occur.

This is LeCun’s missing world model that the neocortex somehow renders. Without a world model, it is impossible to simulate actions and predict their consequences.

As L slowly began to speak again, Damasio asked her about her experience over the prior six months. Although L had little memory of it, she did recall the few days before beginning to speak. She described the experience as not talking because she had nothing to say. She claimed her mind was entirely “empty” and that nothing “mattered.” She claimed that she was fully able to follow the conversations around her, but she “felt no ‘will’ to reply.” It seems that L had lost all intention.

It seems that in early mammals, the sensory neocortex was where simulations were rendered, and the frontal neocortex was where simulations were controlled—it is the frontal neocortex that decided when and what to imagine.

And as we saw in the last chapter, it is exactly when things change or are hard (i.e., are uncertain) that animals pause to engage in vicarious trial and error.

And specifically when rats engage in this vicarious trial and error behavior, the activity in the aPFC and the sensory cortex become uniquely synchronized. One speculation is that the aPFC is triggering the sensory neocortex to render a specific simulation of the world.

Alternatively, it could be the basal ganglia that determines the actions taken during these simulations. This would be even closer to how AlphaZero worked—it selected simulated actions based on the actions its model-free actor predicted were best.

If the basal ganglia keeps getting more excited by imagining drinking water than by imagining eating food (as measured by the amount of dopamine released), then these votes for water will quickly pass the choice threshold.

The behavior had been repeated so many times that the aPFC and basal ganglia did not detect any uncertainty and therefore the animal did not pause to consider the consequences.

Habits are automated actions triggered by stimuli directly (they are model-free). They are behaviors controlled directly by the basal ganglia. They are the way mammalian brains save time and energy, avoiding unnecessarily engaging in simulation and planning.

Just as the explanations of sensory information are not real (i.e., you don’t perceive what you see), so intent is not real; rather, it is a computational trick for making predictions about what an animal will do next.

This is important: The basal ganglia has no intent or goals. A model-free reinforcement learning system like the basal ganglia is intent-free; it is a system that simply learns to repeat behaviors that have previously been reinforced.

This is why it is possible, at least in circumstances where people make aPFC-driven (goal-oriented, model-based, system 2) choices, to ask why a person did something.

It is somewhat magical that the very same neocortical microcircuit that constructs a model of external objects in the sensory cortex can be repurposed to construct goals and modify behavior to pursue these goals in the frontal cortex.

The sensory cortex engages in passive inference—merely explaining and predicting sensory input. The aPFC engages in active inference—explaining one’s own behavior and then using its predictions to actively change that behavior.

Active inference suggests that the aPFC constructs intent and then tries to predict behavior consistent with that intent; in other words, it tries to make its intent come true.

behavior: the basal ganglia begins as the teacher of the aPFC, but as a mammal develops, these roles flip, and the aPFC becomes the teacher of the basal ganglia. And indeed, during brain development, agranular parts of the frontal cortex begin with a layer four that then slowly atrophies and disappears during development, leaving layer four largely empty.

The aPFC’s triggering of simulation is called imagination when it is unconstrained by current sensory input and attention when it is constrained by current sensory input. But in both cases, the aPFC is, in principle, doing the same thing.

If the aPFC had previously imagined seeing a rabbit and running toward it, then it can control the basal ganglia’s choices by using attention to ensure that when the rat sees this ambiguous picture, it sees a rabbit, not a duck.

Controlling ongoing behavior often also requires working memory—the maintenance of representations in the absence of any sensory cues.

These tasks require the aPFC because working memory functions in the same way as attention and planning—it is the invoking of an inner simulation. Working memory—holding something in your head—is just your aPFC trying to keep re-invoking an inner simulation until you no longer need it.

This is why people become more impulsive when tired or stressed—the aPFC is energetically expensive to run, so if you are tired or stressed, the aPFC will be much less effective at inhibiting the amygdala.

To summarize: Planning, attention, and working memory are all controlled by the aPFC because all three are, in principle, the same thing. They are all different manifestations of brains trying to select what simulation to render.

The aPFC controls behavior not by telling but by showing.

Early mammals had the ability to vicariously explore their inner model of the world, make choices based on imagined outcomes, and stick to the imagined plan once chosen. They could flexibly determine when to simulate futures and when to use habits; and they intelligently selected what to simulate, overcoming the search problem. They were our first ancestors to have goals.

hierarchy. The front part of the basal ganglia automatically associates stimuli with high-level goals. It is what generates cravings: You come home and smell rigatoni, and suddenly you are on a mission to eat some.

The aPFC, however, is what makes you pause and consider if you actually want to pursue these cravings (“ What about our diet?”).

Any level of goal, whether high-level or low-level goals, has both a self model in the frontal neocortex and a model-free system in the basal ganglia. The neocortex offers a slower but more flexible system for training, and the basal ganglia offers a faster but less flexible version for well-trained paths and movements.

Most vertebrates at the time, as with modern lizards and fish, could still move quickly, remember patterns, track the passage of time, and intelligently learn through model-free reinforcement learning, but their movements were not planned.

The greater the predation risk experienced by primates, the larger the social group they created in response.

This process of ever-escalating deceptions and counter-deceptions reveals that both Rock and Belle were able to understand the other’s intent

The degree to which animals can infer the intent and knowledge of others continues to be controversial in animal psychology.

Primates are extremely sensitive to interactions that violate the social hierarchy.

If you examined the social hierarchy of different monkey groups, you would notice that it often isn’t the strongest, biggest, or most aggressive monkey who sits at the top. Unlike most other social animals, for primates, it is not only physical power that determines one’s social ranking but also political power.

Higher-ranking monkeys get their pick of food, grooming partners, mates, and resting sites.

One’s evolutionary fitness improves with one’s rank; higher-ranking monkeys have more children and are less likely to die from disease.

higher-ranking monkeys tend to be better at recruiting allies from unrelated individuals, and hierarchy reversals most often happen when monkeys fail to recruit such allies.

Those with power benefit from forging a sufficient coalition of lower-ranking allies, and low-ranking monkeys can substantially improve their lives by forging friendships with the right high-ranking families.

High-ranking monkeys also exhibit a cleverness in which lower-ranking members they choose to befriend. In a study where different low-ranking monkeys were trained to do specific tasks to obtain food, high-ranking monkeys quickly befriended those who had specialized skills,

Any primate born with better tricks for currying favor and gaining allies would survive better and have more babies.

Indeed, neocortex size of primates is correlated not only with social-group size but also with social savviness.

primitive version of theory of mind—only through this ability can individuals infer what others want and thereby figure out whom to cozy up to and how.

Only through this ability of theory of mind can you figure out who is likely to become powerful in the future, whom you need to make friends with, and whom you can deceive.

There have now been numerous experiments confirming this. The granular prefrontal cortex becomes uniquely active during tasks that require self-reference, such as evaluating your own personality traits, general self-related mind wandering, considering your own feelings, thinking about your own intentions, and thinking about yourself in general.

In other words, the gPFC constructs explanations of the simulation itself, of what the animal wants and knows and thinks.

metacognition: the ability to think about thinking.

These systems are all bootstrapped on one another. Reflexes drive valence responses without any learning required, making choices based on evolutionarily hard-coded rules. The vertebrate basal ganglia and amygdala can then learn new behaviors based on what has historically been reinforced by these reflexes, making choices based on maximizing reward. The mammalian aPFC can then learn a generative model of this model-free behavior and construct explanations, making choices based on imagined goals (e.g., drinking water). This could be considered a first-order model. The primate gPFC can then learn a more abstract generative model (a second-order model) of this aPFC-driven behavior and construct explanations of intent itself, making choices based on mind states and knowledge (I’m thirsty; drinking water when thirsty feels good, and when I simulate going down this way, I find water in my simulation, hence I want to go in this direction).

And indeed, revealing their importance in understanding others, the size of these granular prefrontal areas is correlated with social-network size in primates. The bigger a primate’s granular prefrontal area, the higher in the social hierarchy it tends to be.

Modeling your own mind and that of others is interwoven.

People tend to project their own personality traits onto others.

Over the subsequent twenty years, Rizzolatti’s mirror neurons have been found in numerous behaviors (grasping, placing, holding, finger movements, chewing, lip smacking, sticking one’s tongue out), across multiple areas of the brain (premotor cortex, parietal lobe, motor cortex), and across numerous species of primates.

When a primate watches another primate do an action, its premotor cortex often mirrors the actions it is observing.

By imagining yourself doing what others are doing, you can begin to understand why they are doing what they are doing:

people with impairments in performing specific movements, also show impairments in understanding the intentions of those very same movements in others.

The subregions of premotor cortex required for controlling a given set of motor skills are the same subregions required for understanding the intentions of others performing those same motor skills.

This suggests that people mentally simulate themselves picking up a box when seeing someone else pick up a box (“ I would turn my arm that way only if the box was heavy”).

The main benefit is that it helps us, as it helped early primates, learn new skills through observation.

If you temporarily inhibit his premotor cortex during this task, he becomes specifically impaired at imitating hand motions but performs normally at following the red dots. Premotor activation is not just correlated with imitation learning; it seems to be, at least in some contexts, necessary for imitation learning. And here we can begin to unravel why primates are such great tool users.

Transmissibility Beats Ingenuity

if at least one member of a group figures out how to manufacture and use a termite-catching stick, the entire group can acquire this skill and continuously pass it down throughout generations.

Teaching is possible only with theory of mind. Teaching requires understanding what another mind does not know and what demonstrations would help manipulate another mind’s knowledge in the correct way.

Theory of mind enables a chimp child to realize that the reason it is not getting food with its stick while its mother is getting food is that its mother has a skill it does not yet have. This enables a continuous motivation to acquire the skill, even if it takes a long time to master.

Understanding the intentions of movements is essential for observational learning to work; it enables us to filter out extraneous movements and extract the essence of a skill.

Directly copying expert behaviors turned out to be a dangerously brittle approach to imitation learning.

This technique is called “inverse reinforcement learning” because these systems first try to learn the reward function they believe the skilled expert is optimizing for (i.e., their “intent”), and then these systems learn by trial and error, rewarding and punishing themselves using this inferred reward function.

Part of what makes this frugivore strategy so challenging is that it requires not only simulating differing navigational paths but also simulating your own future needs.

we have already seen where brains need to infer an intent—a “want”—of which it does not currently share: when they’re trying to infer the wants of other people. Might brains be able to use the same mechanism of theory of mind to anticipate a future need? Put another way: Is imagining the mind of someone else really any different from imagining the mind of your future self?

These may not, in fact, have been separate abilities but rather emergent properties of a single new breakthrough: the construction of a generative model of one’s own mind, a trick that can be called “mentalizing.”

Concepts, ideas, and thoughts, just like episodic memories and plans, are not unique to humans. What is unique is our ability to deliberately transfer these inner simulations to each other, a trick possible only because of language.

Money, gods, corporations, and states are imaginary concepts that exist only in the collective imaginations of human brains.

Coordinating behavior using mentalizing alone works only by each member of a group directly knowing each other. This mechanism of cooperation doesn’t scale; the limit of human group size maintained only by direct relationships has been estimated to be about one hundred fifty people.

An analogy to DNA is useful. The true power of DNA is not the products it constructs (hearts, livers, brains) but the process it enables (evolution). In this same way, the power of language is not its products (better teaching, coordinating, and common myths) but the process of ideas being transferred, accumulated, and modified across generations.

All human inventions, both technological and cultural, require an accumulation of basic building blocks before a single inventor can go “Aha!,” merge the preexisting ideas into something new, and transfer this new invention to others.

Many people who become linguistically impaired are otherwise intellectually typical. And people can be linguistically gifted while otherwise intellectually impaired.

Language is a specific and independent skill that evolution wove into our brains.

The human brain has parallel control of facial expressions; there is an older emotional-expression system that has a hard-coded mapping between emotional states and reflexive responses. This system is controlled by ancient structures like the amygdala. Then there is a separate system that provides voluntary control of facial muscles that is controlled by the neocortex.

The emotional-expression system and the language system have another difference: one is genetically hardwired, and the other is learned.

A skill as sophisticated as flying is too information-dense to hard-code directly into a genome. It is more efficient to encode a generic learning system (such as a cortex) and a specific hardwired learning curriculum (instinct to want to jump, instinct to flap wings, and instinct to attempt to glide). It is the pairing of a learning system and a curriculum that enables every single baby bird to learn how to fly.

TD-Gammon was trained by playing against itself. TD-Gammon always had an evenly matched player. This is the standard strategy for training reinforcement learning systems.

Google’s AlphaZero was also trained by playing itself. The curriculum used to train a model is as crucial as the model itself.

It seems conversation is not a natural consequence of the ability to learn language; rather, the ability to learn language is, at least in part, a consequence of a simpler genetically hard-coded instinct to engage in conversation. It seems to be this hardwired curriculum of gestural and vocal turn-taking on which language is built. This type of turn-taking evolved first in early humans; chimpanzee infants show no such behavior.

What is one of the first things that parents do once they have achieved a state of joint attention with their child? They assign labels to things. The more joint attention expressed by an infant at the age of one year, the larger the child’s vocabulary is twelve months later.

Humans may have also evolved a unique hardwired instinct to ask questions to inquire about the inner simulations of others.

This is why humans deprived of contact with others will develop emotional expressions, but they’ll never develop language. The language curriculum requires both a teacher and a student.

It is hard to get chimps to engage in joint attention; it is hard to get them to take turns; and they have no instinct to share their thoughts or ask questions. And without these instincts, language is largely out of reach, just as a bird without the instinct to jump would never learn to fly.

Humans are born not when they are ready to be born, but when their brains hit the maximum size that can fit through the birth canal.

Homo erectus was our meat-eating, stone-tool-using, (possibly) fire-wielding, premature-birthing, (mostly) monogamous, grandmothering, hairless, sweating, big-brained ancestor. The million-dollar question is, of course, did Homo erectus speak?

Language doesn’t directly benefit an individual the way eyes do; it benefits individuals only if others are using language with them in a useful way.

It turns out most group behaviors in animals aren’t altruistic; they are mutually beneficial arrangements that are net-positive for all participants. Fish swim in shoals because it benefits all of them, and the movements are actually best explained by the fish on the edges all fighting to get into the center where it is safest. Wildebeests band together because they are all safer when they are in a group.

What do we humans naturally have an instinct to talk about? What is the most natural activity we use language for? Well, we gossip.

And amid all the altruistic instincts and behaviors that began to form from this dynamic, the most powerful was, undoubtedly, the use of language to share knowledge and cooperatively plan among non-kin.

Not only did the pressure for bigger brains continue to ratchet up, but so too did the frontier of how big it was biologically possible for brains to get. As brains expanded, humans became better hunters and cooks, which provided more calories and thereby expanded the frontier of how big brains could get. And as brains got bigger, births became earlier, which created even more opportunity for language learning, which put even more pressure on altruistic cooperation to support child-rearing, which again expanded the frontier of how big brains could get as it became possible to evolve longer time periods of childhood brain development.

Our language, altruism, cruelty, cooking, monogamy, premature birthing, and irresistible proclivity for gossip are all interwoven into the larger whole that makes up what it means to be human.

When humans use language with each other, there is an ungodly number of assumptions not to be found in the words themselves. We infer what people actually mean by what they say.

When a human makes a request like “Maximize the production of paper clips” or “Be nice to Rima” or “Eat breakfast,” he or she is not actually providing a well-defined goal. Instead, both parties are guessing what is going on in the other’s head.

The intertwining of mentalizing and language is ubiquitous. Every conversation is built on the foundation of modeling the other minds you are conversing with—guessing what one means by what he said and guessing what should be said to maximize the chance the other knows what you mean.

A calculator performs arithmetic better than any human, but still lacks the same understanding of math as a human. Even if LLMs correctly answer commonsense and theory-of-mind questions, it does not necessarily mean it reasons about these questions in the same way.

But without incorporating an inner model of the external world or a model of other minds—without the breakthroughs of simulating and mentalizing—these LLMs will fail to capture something essential about human intelligence.

In the human brain, language is the window to our inner simulation. Language is the interface to our mental world. And language is built on the foundation of our ability to model and reason about the minds of others—to infer what they mean and figure out exactly which words will produce the desired simulation in their mind.

And speaking was possible only because mentalizing came before; without the ability the infer the intent and knowledge in the mind of another, you could not infer what to communicate to help transmit an idea or infer what people mean by what they say. And without the ability to infer the knowledge and intent of another, you could not engage in the crucial step of shared attention whereby teachers identify objects for students.