At the forefront of 21st century brain science
The Origins of Humanity Reimagined Through AI's Eyes

“This book beautifully encapsulates all the discoveries made in neuroscience,
“It reveals everything you were too scared to ask.”
Carl Friston, the world's most cited neuroscientist

“It’s not just a simple brain science book.
“It contains unique insights into the future of artificial intelligence and human intelligence.”
Jaeseung Jeong, Professor of Brain and Cognitive Sciences at KAIST and Dean of the School of Convergence Studies

The function of the first brain was very simple.
It was simply a manipulation of the organism to bring it closer to its prey or to move it away from its predators.
Since then, the brain has gone through five innovations: repeating learning, imagining, guessing, and using language.
The human brain is not as special as we vaguely think, and it did not evolve specifically for higher-level thinking.
But the result was, unexpectedly, the birth of a thinking being.
And these beings are creating a new intelligence called AI with their own hands.

Max Bennett, a scientist and AI entrepreneur currently attracting the attention of world-renowned scholars, says that AI was able to surpass human intelligence, and that all the secrets we need to know to move forward lie in five innovations that occurred in the human brain.
His book, "The Origins of Intelligence," a blend of academic curiosity and entrepreneurial pragmatism, will not only stir the hearts of those seeking to understand the brain and the nature of humanity, but also provide practical assistance to those leading the AI ​​industry and those eager to understand future changes.
What do you envision as the next revolution in artificial intelligence and human intelligence? Envision the future with this next-generation introduction to brain science, which integrates evolutionary perspectives and neurological mechanisms to provide new insights.

If you want to reverse-engineer how the brain works, if you want to build a robot named Logy, if you want to uncover the hidden nature of human intelligence, the human brain might be the last place you should look.
Perhaps the best place to start is with dusty fossils buried deep in the Earth's crust, tiny genes embedded in cells throughout the animal kingdom, and the brains of many other animals on Earth.
In other words, the answer may not lie in the present, but in the hidden remnants of the long past.

How wonderful it would be if we could go back in time and examine the original brain, understanding how it worked and the functions it enabled. Furthermore, if we could trace the brain's gradual complexity throughout the human lineage, observing each physical change and the intellectual abilities it enabled. If we could do so, we might be able to understand the resulting complexity.
In fact, biologist Theodosius Dobzhansky famously said:
“In biology, nothing can be understood without seeing it through the lens of evolution.”
--- From "Introduction: Reconstructing the History of Human Intelligence Through the Eyes of AI"

There is a more important observation about bilateral symmetry.
The only animals with brains are bilaterally symmetrical.
This is no coincidence.
The original brain and the bilateral symmetrical system initially shared the same evolutionary purpose.
It allows animals to explore their surroundings through manipulation.
Control is innovation #1.
--- 「2.
From “The Birth of Good and Bad”

Fish are much smarter than we think.
Fish can learn to find and press specific buttons to get food.7
You can also learn how to escape through small escape hatches to avoid getting caught in the net.
You can even learn how to get food through rings.
Fish can remember how to perform these tasks for months or even years after they are trained.
In all these experiments, the learning process is the same.
Fish gradually refine their behavior by trying relatively random behaviors and figuring out which ones are reinforced.
In fact, Thorndike's trial-and-error learning is more often called by another name.
It is ‘reinforcement learning’.
--- 「5.
From “Learning from Trial and Error”

Curiosity and reinforcement learning have evolved together because reinforcement learning requires curiosity to work.
The discovery of the ability to recognize patterns, remember locations, and flexibly change behavior based on past rewards and punishments opened up new opportunities for the first vertebrates.
For the first time, learning became an extremely valuable activity in itself.
The more patterns a vertebrate can recognize and the more locations it can remember, the better its chances of survival.
And the more you try, the more likely you are to learn the right connections between your actions and the consequences that follow.
So curiosity first emerged 500 million years ago in the tiny brains of ancestors similar to modern fish.
--- 「8.
From "Why did I become curious about life?"

For those working in the field of AI, Mountcastle's hypothesis is an incomparable scientific gift.
(Omitted) Instead of needing to understand all the trillions of connections in the entire cortex, we might only need to understand the million or so connections in the cortical columns.
Furthermore, if Mountcastle's theory is correct, it means that the neocortical columns implement a very general and universal algorithm that can be applied to a wide variety of functions across all sensory modalities, including motor, language, and perception.

Modern AI models are often considered narrow AI.
This means that it can only work in limited situations where it has been specially trained.
The human brain appears to be universal.
It can work in a variety of situations.
So far, the focus of research has been on creating artificial general intelligence.
But it may be that we have been researching it backwards.
The reason the new cortex can do its job so well may be that in some ways it is much less general-purpose than current artificial neural networks.
(Omitted) For example, the cortex may be pre-wired to assume that sensory data, whether visual, auditory, or kinesthetic, represent three-dimensional objects that exist independently of ourselves and can move independently.
Then there would be no need to learn things like time, space, and the difference between me and other beings.
--- 「11.
From "The Gift of the New Skin"

If we succeed in creating a robot with a mammal-like locomotion system, it will also bring with it several desirable properties.
These robots will use new, complex technologies to learn autonomously and adjust their movements in real time to respond to changes in the world.
If we give them a super-goal, the robots will come up with all the sub-goals necessary to achieve it.
When they first learn a new task, they will be slow and careful, simulating each movement before performing it, but as they get better, the actions will become automatic.
Robots will learn new skills at an increasingly faster rate because they will reapply previously learned low-level skills to newly experienced higher-level goals.
If their brains actually function like mammalian brains, they might not need massive supercomputers to accomplish these tasks.
In fact, the entire human brain only needs about the energy that goes into a single light bulb to function.

If we succeed in creating a robot with a mammal-like locomotion system, it will also bring with it several desirable properties.
These robots will use new, complex technologies to learn autonomously and adjust their movements in real time to respond to changes in the world.
If we give them a super-goal, the robots will come up with all the sub-goals necessary to achieve it.
When they first learn a new task, they will be slow and careful, simulating each movement before performing it, but as they get better, the actions will become automatic.
Robots will learn new skills at an increasingly faster rate because they will reapply previously learned low-level skills to newly experienced higher-level goals.
If their brains actually function like mammalian brains, they might not need massive supercomputers to accomplish these tasks.
In fact, the entire human brain only needs about the energy that goes into a single light bulb to function.
--- 「14.
From "Why Dishwashing Robots Haven't Been Developed"

Imagine putting a primate ancestor into a maze.
When the animal reached the fork in the road, it turned left.
Suppose we could ask different brain regions of the animal why it turned left.
Then you will hear very different answers step by step.
The reflex will respond like this:
“Because evolution has engraved in me a rule that tells me to turn left where the scent of food is coming from,” the vertebrate brain structure would respond.
“Because going left maximizes the predicted future reward.” This is how the mammalian brain structure would respond.
“Because the left side leads to food,” the primate brain structure would say.
“I’m hungry, and when I’m hungry, I feel better when I eat, and as far as I know, the left path leads to food.”

Advertising platforms can use human behavior to predict what people will buy next.
There is also AI that can identify emotions expressed on the face (the system is trained by showing it countless photos of faces categorized by emotion).
But all of this is a far cry from the complex theory of mind observed in the brains of humans and other primates. If we want AI systems and robots to live alongside us, inferring our intentions from our words, anticipating our needs and wants before we speak, and navigating social relationships in human groups with their myriad hidden rules and etiquette—in other words, if we want AI systems that are similar to humans, they must possess a theory of mind.
--- 「16.
From "Understanding Other People's Minds"

In some ways, it may become increasingly difficult to discern the difference between the way LLMs think and the way people think.
But although calculators are better at math than anyone else in the world, they don't understand math like people do.
Likewise, just because an LLM answers commonsense and theory of mind questions correctly doesn't mean it reasons in the same way as a human.
(Omitted) In fact, LLM has a memory capacity as large as a supercomputer, so it has read more books and articles than a person can read in 1,000 lifetimes.
So, while it may seem like common-sense reasoning on the surface, it actually works more like pattern matching in an astronomically large text corpus.
--- 「22.
From "Chat GPT and a Window into the Mind"

Each innovation was possible because of the basic components that came before it.
Control was possible thanks to the evolution of nerve cells before that.
Reinforcement learning was possible because bootstrapping was performed based on previously evolved emotional neurons.
The simulation was possible because there was prior trial-and-error learning in the basal ganglia.
Without the basal ganglia that enable trial-and-error learning, the mechanisms by which imagined simulations can influence behavior would not have been developed.
The evolution of actual trial-and-error learning in vertebrates allowed for the later emergence of vicarious trial-and-error in mammals.
Mentalization was possible because simulation had evolved before then.
Mentalization, in a word, is simulating the old mammalian areas of the neocortex.
The same calculations are simply made internally.
The reason language was possible was because mentalization appeared before it.
If we didn't have the ability to infer the intentions and knowledge of others, we wouldn't be able to infer what we need to communicate to convey our ideas, nor would we be able to infer what others are saying and what their intentions are.
Without the ability to infer the other person's knowledge and intentions, we would not be able to engage in the crucial step of shared attention that teachers teach their students.

It took only 4.5 billion years for the raw material molecules on Earth to turn into the human brain.
So what level of intelligence can we reach in the remaining 7 billion years of evolution?

--- "Going out.
From “The Sixth Innovation”