A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains

by Max Solomon Bennett

pockets and furnish their homes. The Jetsons correctly predicted video calls, flat-screen TVs, cell phones, 3D printing, and smartwatches; all technologies that were unbelievable in 1962 and yet were ubiquitous by 2022.

the autonomous robot named Rosey. Rosey was a caretaker for the Jetson family, watching after the children and tending to the home.

All attempts to do this have failed. It isn’t fundamentally a mechanical problem; it’s an intellectual one—the ability to identify objects in a sink, pick them up appropriately, and load them without breaking anything has proven far more difficult than previously thought.

If only we could go back in time and examine this first brain to understand how it worked and what tricks it enabled. If only we could then track the complexification forward in the lineage that led to the human brain, observing each physical modification that occurred and the intellectual abilities it afforded. If we could do this, we might be able to grasp the complexity that eventually emerged. Indeed, as the biologist Theodosius Dobzhansky famously said, “Nothing in biology makes sense except in the light of evolution.”

left the following on a blackboard shortly before his death: “What I cannot create, I do not

understand.” The brain is our guiding inspiration for how to build AI, and AI is

Let’s start with Artificial Rat–level intelligence (ARI), then move on to Artificial Cat–level intelligence (ACI), and so on to Artificial Human–level Intelligence (AHI). —YANN LECUN, HEAD OF AI AT META

After countless random nucleotide chains were constructed and destroyed, a lucky sequence was stumbled upon, one that marked, at least on Earth, the first true rebellion against the seemingly inexorable onslaught of entropy. This new DNA-like molecule wasn’t alive per se, but it performed the most fundamental process by which life would later emerge: it duplicated itself. 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.

With these foundations in place, the process of evolution was initiated in full force: variations in DNA led to variations in proteins, which led to the evolutionary exploration of new cellular machinery, which, through natural selection, were pruned and selected for based on whether they further supported survival. By this point in life’s story, we have concluded the long, yet-to-be-replicated, and mysterious process scientists call abiogenesis: the process by which

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

animals that engage in such gastrulation also have neurons and muscles and seem to derive from a common neuron-enabled animal ancestor. Gastrulation, neurons, and muscles are the three inseparable features that bind all animals together and separate animals from all other kingdoms of life.

AT FIRST GLANCE, the diversity of the animal kingdom appears remarkable—from ants to alligators, bees to baboons, and crustaceans to cats, animals seem varied in countless ways. But if you pondered this further, you could just as easily conclude that what is remarkable about the animal kingdom is how little diversity there is. Almost all animals on Earth have the same body plan. They all have a front that contains a mouth, a brain, and the main sensory organs (such as eyes and ears), and they all have a back where waste comes out.

Radially symmetrical animals have only one opening—a mouth-butt if you will—which swallows food into their stomachs and spits

Even modern human engineers have yet to find a better structure for navigation. Cars, planes, boats, submarines, and almost every human-built navigation machine is bilaterally symmetric. It is simply the most efficient design for a movement 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.

Nematodes have a negative-valence neuron that triggers turning when temperatures increase, but only if the temperature is already above a certain threshold; it is a Too hot! neuron. Nematodes also have a Too cold! neuron; it triggers turning when temperatures decrease, but only when temperatures are already below a certain threshold. Together, these two negative-valence neurons enable nematodes to quickly steer away from heat when they’re too hot and away from cold when they’re too cold. Deep in the human brain is an ancient structure called the hypothalamus that houses temperature-sensitive neurons that work in the same way.

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.

While dopamine has no impact on liking reactions, serotonin decreases both liking and disliking reactions. When given drugs that increase serotonin levels, rats smack their lips less to good food and shake their heads less to bitter food. This is also what we would expect given the evolutionary origin of serotonin: serotonin is the satiation, things-are-okay-now, satisfaction chemical, designed to turn off valence responses.

renders even the most exciting stimuli entirely unmotivating. Psychologists call this canonical symptom of depression

Anhedonia in animals like nematodes seems to be a trick to preserve energy in the presence of inescapable stressors. Animals no longer respond to stressors, good food smells, or nearby mates. In humans, this ancient system robs its sufferers of the ability to experience pleasure and motivation. This is the blah or blues of depression. And like all affective states, chronic stress persists after the negative stimuli have gone away. Such learned helplessness, where animals stop trying to escape from negatively valenced stimuli, is seen even in many bilaterians, including cockroaches, slugs, and fruit flies.

We have invented drugs that hack these ancient systems. The euphoria provided by natural opioids is meant to be reserved for that brief period after a near-death experience. But humans can now indiscriminately trigger this state with nothing more than a pill. This creates a problem. Repeatedly flooding the brain with opioids creates a state of chronic stress when the drug wears off—adaptation is unavoidable. This then traps opioid users in a vicious cycle of relief, adaptation, chronic stress requiring more drugs to get back to baseline, which causes more adaptation and thereby more chronic stress. Evolutionary constraints cast a long shadow on modern humanity.

And yet, while associative learning is found across bilaterians, our most distant animal cousins—the radially symmetric jellyfish, anemones, and coral—are not capable of learning associations.* Despite many pairings of a light with an electric shock, an anemone will never learn to withdraw in response to just the light. They withdraw only from the shock itself. The

Your self-driving car doesn’t automatically get better as you drive; the facial-recognition technology in your phone doesn’t automatically get better each time you open your phone. As of 2023, most modern AI systems go through a process of training, and once trained, they are sent off into the world to be used, but they no longer learn. This has always presented a problem for AI systems—if the contingencies in the world change in a way not captured in the training data, then these AI systems need to be retrained, otherwise they will make catastrophic mistakes.

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

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.

The beautifully conserved circuitry of the basal ganglia, first emerging in the minuscule brain of early vertebrates and maintained for five hundred million years, seems to be the biological manifestation of Sutton’s actor-critic system. Sutton discovered a trick that evolution had already stumbled upon over five hundred million years ago.

The result of all these minor perturbations is that the next encounter will be similar but not the same. In figure 7.3 you can see three examples of the olfactory patterns that the next encounter with the predator smell might activate. 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.

vertebrate memory differs from computer memory. Auto-association suggests that vertebrate brains use content-addressable memory—memories are recalled by providing subsets of the original experience, which reactivate the original pattern. If I tell you the beginning of a story you’ve heard before, you can recall the rest; if I show you half a picture of your car, you can draw the rest. However, computers use register-addressable memory—memories that can be recalled only if you have the unique memory address for them. If you lose the address, you lose the memory.

Cohen and McCloskey referred to this property of artificial neural networks as the problem of catastrophic forgetting. This was not an esoteric finding but a ubiquitous and devastating limitation of neural networks: when you train a neural network to recognize a new pattern or perform a new task, you risk interfering with the network’s previously learned patterns.

How do modern AI systems overcome this problem? Well, they don’t yet. Programmers merely avoid the problem by freezing their AI systems after they are trained. We don’t let AI systems learn things sequentially; they learn things all at once and then stop learning.

architecture first decomposed input pictures into multiple feature maps, like V1 seemed to do. Each feature map was a grid that signaled the location of a feature—such as vertical or horizontal lines—within the input picture. This process is called a convolution, hence the name applied to the type of network that Fukushima had invented: convolutional neural networks.

In Montezuma’s Revenge, you start in a room filled with obstacles. In each direction is another room, each with its own obstacles. There is no sign or clue as to which direction is the right way to go. The first reward is earned when you find your way to a hidden door in a faraway hidden room. This makes the game particularly hard for reinforcement learning systems: the first reward occurs so late in the game that there is no early nudging of what behavior should be reinforced or punished. And yet somehow, of course, humans beat this game.

There is an alternative approach to tackling the exploitation-exploration dilemma, one that is both beautifully simple and refreshingly familiar. 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.

exhibit curiosity; only the most advanced invertebrates, such as insects and cephalopods, show curiosity, a trick that evolved independently and wasn’t present in early bilaterians.

The emergence and mechanisms of curiosity help explain gambling, which is an irrational oddity of vertebrate behavior. Gamblers violate Thorndike’s law of effect—they continue to gamble their money away despite the fact that the expected reward is negative.

their long-range vision on land, were never endowed with the gift of simulating. The only nonmammals that have shown evidence of the ability to simulate actions and plan are birds. And birds are, conspicuously, the only nonmammal species alive today that independently evolved warm-bloodedness.

What they do have is an ability to learn powerful “world models” that allow them to predict the consequences of their actions and to search for and plan actions to achieve a goal. The ability to learn such world models is what’s missing from AI systems today. The simulation rendered in the neocortices

In other words, the path toward salt was vicariously reinforced before the rat even acted. I am unaware of any studies showing that fish or reptiles can perform such a task.

contains my favorite food, cherry, and gamble that it will also be released quickly? When rats chose to forgo quick access to a banana treat to try the cherry door and the next tone signaled a long wait of forty-five seconds, rats showed all the signs of regretting their choice. They paused and looked back toward the corridor that they had passed and could no longer go back to. And the neurons in the taste area of the neocortex reactivated the representation of banana, showing that rats were literally imagining a world where they had made a different choice and got to eat the banana.

of the aPFC of a rat, you can see patterns of activity that encode the task a rat is performing—with specific populations of neurons selectively firing only at specific locations within a complex task sequence, reliably tracking progress toward an imagined goal.

What is the evolutionary usefulness of this model of self in the frontal cortex? Why try to “explain” one’s own behavior by constructing “intent”? It turns out, this might be how mammals choose when to simulate things and how to select what to simulate. Explaining one’s own behavior might solve the search problem. Let’s see how.

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. When such automation occurs at the right times, it enables us to complete complex behaviors easily; when it occurs at the wrong times, we make bad choices.

Perhaps the motor cortex doesn’t generate motor commands but rather motor predictions. Perhaps the motor cortex is in a constant state of observing the body movements that occur in the nearby somatosensory cortex (hence why there is such an elegant mirror of motor cortex and somatosensory cortex) and then tries to explain the behavior and use these explanations to predict what an animal will do next. And perhaps

As Yann Lecun said, “the weak reasoning abilities of LLMs are partially compensated by their large associative memory capacity. They are a bit like students who have learned the material by rote but haven’t really built deep mental models of the underlying reality.”

All the evolutionary innovations that followed the first string of DNA have been in this spirit, the spirit of persisting, of fighting back against entropy, of refusing to fade into nothingness. And in this great battle, ideas that float from human brain