What Is Intelligence? by Blaise Aguera y Arcas Lessons from AI About Evolution, Computing, and Minds
What's it about?
What is Intelligence? (2025) repositions AI not as a looming alien mind, but as a natural continuation of life’s long, messy story of evolution, cooperation, and prediction. It weaves together bacteria, brains, cities, and neural networks to show how intelligence emerges wherever systems learn to model themselves and their world. It takes us through the past, present, and future of AI, while describing our place in it.
As the title suggests, in this lesson we’re going to take a big-picture look at the concept of intelligence. But not only that. We’re going to lay out the author’s bold message, which is that modern AI isn’t some clever imitation – it’s a real expression of intelligence, built on the same underlying principle that drives all living minds: prediction. The main argument is that intelligence shouldn’t be thought of in terms of biology, but rather computation. Understanding prediction as the engine of intelligence reframes everything – from how life began, to how organisms survive, to why large AI models suddenly work as well as they do.
This perspective pushes aside old debates about whether AI is “really” intelligent and instead turns our attention toward the bigger story: how prediction scaled from molecules to minds, how it now extends into machines, and how these new symbiotic intelligences might reshape our world. So let’s get into it by going back some four billion years ago to when life first took shape.
When we rewind Earth’s history to its earliest scenes, around 4.6 billion years ago, we’re looking at a world that barely resembles today. It’s known as the Hadean Eon, when the planet consisted of oceans of lava, a sky thick with volcanic gases, and a surface that was battered by asteroids big enough to boil the atmosphere. Somehow, within this chaos, the chemistry of life began to brew.
While the exact spark may never be known, one of the more attractive theories takes place in the deep sea. At hydrothermal vents – towering constructions built from minerals and heat – water, gas, and rock interact inside tiny porous chambers. These pores naturally create the kind of microscopic batteries that drive chains of chemical reactions that form self-sustaining cycles that look a lot like the metabolic “motor” still humming inside every cell today. It’s as though our own cells carry a memory of those early oceans.
As life took shape, it did so in bold leaps. Once a bacterium was engulfed by an archaeal cell, and instead of being digested, the two formed an alliance that resulted in mitochondria – which in turn opened the door to multicellular life, nervous systems, animals, and eventually us.
Symbiosis, and big, evolutionary leaps aren’t limited to biology. Technology evolved in similar bursts, each breakthrough forming when separate parts came together to create something new, like combining stone, wood, and sinew to make hunting tools.
Now, you may be hesitant to draw parallels between biological life and computing, but early computer pioneers like Alan Turing and John von Neumann saw the link. They imagined machines that could read instructions, modify those instructions, and even build copies of themselves – which is precisely the kind of logic of self-replication that DNA serves. So it’s hardly a stretch to say that a cell is both part chemistry and part computation.
To put it simply: reproduction requires computation. In nature, millions of tiny agents – ribosomes, enzymes, morphogens – operate in parallel, using randomness and local rules to achieve a larger unified function.
That same principle shows up in recent experiments. In 2023, researchers tried to recreate a sort of digital primordial soup environment, to see if the same kind of mutations and self-replicating patterns could emerge. Into the soup went thousands of simple digital tapes that contained both rudimentary code and data, and the ability to modify itself. Sure enough, after millions of interactions, the tapes do begin to reproduce. All of a sudden, a phase-transition happens, and they shift from noise to self-replicating patterns.
From randomness, replicators emerge; once replicators exist, evolution takes over. And as we’ll see in the next section, that rising complexity can give way to intelligence.
At a very basic level, intelligence is all about survival. If we zoom way, way in and observe a lone bacterium drifting through its liquid world, we can see the core of intelligence forming. Not the kind we usually picture – with neurons firing or thoughts bubbling up – but a minimal, stripped-down version whose only real job is to keep the cell alive from one moment to the next.
Life can’t just replicate blindly. It needs a way to build itself, maintain itself, and do all of that inside a real environment that never stays still. That messiness changes everything. It means a living system needs a boundary – a membrane – to protect its delicate chemistry. But the boundary can’t be sealed shut, because energy and matter have to flow in and out in order to survive.
Inputs and outputs are basically the cost of being alive, but every input is also information. Every organism needs to know what to consume and what to avoid. That’s the first glimmer of intelligence: using information to stay alive.
On the inside, a bacterium has its own tiny control network – a biochemical “brain” made of genes and proteins that monitor hunger, energy levels, and all sorts of microscopic states. These internal signals blend with external ones to steer behavior. And that behavior is simple but clever. Intelligent, even. Many bacteria swim with a little rotating propeller called a flagellum, and they manage to swim toward food by comparing now with just-a-moment-ago.
When the scent of food is rising, they run longer. When it’s falling, they tumble more often. Whether it’s food, heat, or light, bacteria follow the same playbook: infer a pattern using only a sequence of tiny, random events over time. In other words, they solve problems using statistics and probability.
The whole point of this is homeostasis – keeping internal conditions in the comfort zone. So we can even represent it all in a computation model that knits together three streams at once: external inputs (X), internal state (H), and possible actions (O). This joint distribution P(X,H,O) is the organism’s entire worldview. It predicts not just what’s out there, but what might happen if it runs, tumbles, or switches a gene on or off.
Here’s where evolution comes in. Over generations, the population tunes itself toward better and better internal models – leaner, more general, more predictive. And once an organism has a model of the world that includes itself – a model that shapes its actions, which in turn shape its future – the bacterium has a goal: staying alive. That goal is the seed of purposiveness, which in turn is the beginning of intelligence, and the foundation for everything that comes next.
Let’s fast forward to the turn of the eighteenth century, when one of the stars of the Enlightenment era, the German mathematician and philosopher Gottfried Wilhelm Leibniz, helped usher in the age of the mechanical calculator.
Leibniz was one of the first people to see broader, loftier uses for calculations. He dreamed of an algorithm that could give a true or false answer to any statement – a day when philosophers would settle debates the way accountants balance ledgers.
But for the next few centuries Leibniz’s dream was put on the back burner in favor of commercialism. In the eighteenth and nineteenth centuries, calculations and early computers helped usher in the Industrial Revolution. Thanks to machines like the Analytical Engine, created by British mathematician Charles Babbage, mechanical computers became extensions of the factory floor.
Still, there were a couple of people who recognized a greater potential. For the British mathematician Ada Lovelace, computers like the Analytical Engine could lead us to a general theory that linked machines, mathematics, and the very patterns that shape life. She saw a future where we might discover the “calculus of the nervous system.”
By the early 1900s, neuroscience finally caught up. Experiments indeed picked up on electrical spikes in the nervous system, which led to a wild idea: maybe neurons were logic gates and the brain was indeed a kind of analytical engine. In 1943, two American scientists, Warren McCulloch and Walter Pitts, went all in on that view, sketching a neuron-as-logic-gate model that would end up influencing the design for digital computers.
Those computers took off and became the fast, precise, obedient engines of binary thinking. And they set the tone for GOFAI, or good old-fashioned artificial intelligence: machines were the logical, rational robots while humans were the intuitive, emotional beings.
Eventually, as scientists continued to zoom in on the human brain, they found that neurons didn’t behave like straight-forward logic gates. They behaved like complex, nonlinear, dynamic systems. The brain looked less like a theorem prover and more like the living, breathing, messy organism it belonged to.
But, the GOFAI model was easy to implement, so it moved ahead while the more human-like AI found a home in the mostly forgotten branch of cybernetics. This was a school of thought that saw minds and machines as prediction engines – systems that learn from experience, regulate themselves, adapt to uncertainty, and close feedback loops with the world. It’s a worldview grounded in biology, behavior, feedback, and prediction, rather than pristine logic.
Instead of batch processing, cybernetics focused on real-time feedback loops – systems that sense the world, act on it, and constantly update their behavior. Gun turrets, bomb-sight computers, and early flight simulators all worked this way.
But while the cybernetics theory was grand, it would take some time before the technology caught up. In that time, however, the core cybernetic insight – that intelligence emerges from prediction, feedback, learning, and continuous values – never went away.
So how did the early, wobbly steps of cybernetics lead to modern AI advancements? While we can’t get into every detail, we have to mention the perceptron. The perceptron is the simplest kind of artificial neuron, the kind you can bundle together, feed a ton of photos, and ask it to recognize something like bananas.
But the real ability for a machine to learn, through trial-and-error, how to identify a banana, is computation. The computation is what allows the machine to learn – for the neuron to fire in a positive way when it makes the right decisions. This is known as an activation function, and one of the most commonly used is called ReLU, short for Rectified Linear Unit.
The name is a little misleading, because ReLu is a feat of nonlinear problem solving that makes it possible for a machine to recognize a banana no matter how it’s lit, angled, ripened, or photographed. In biological terms, it’s like neurons firing only when their preferred pattern shows up strongly enough.
When the process becomes this repeated predictive routine, what clicks into place is something called transfer learning, where the AI has learned enough to start applying some of the same patterns for completely different objects. That’s why you can successfully add an apple perceptron with barely any examples.
And that sets the stage for unsupervised learning. Instead of labeling millions of images, you mask out random chunks and ask the network to inpaint them – to, fill in the blanks and predict what’s missing. Do that well, and you’ve implicitly learned the world: edges, colors, objects, depth, categories. You’ve built the representations needed for almost any downstream task. It’s the same trick language models use with missing words. Prediction and modeling are two sides of the same coin.
Our own visual system runs a kind of lifelong masked-autoencoder training. Most of what we “see” is low-resolution, noisy, or obscured – but because we’re constantly flicking our eyes around, checking our guesses, and filling in gaps, we build a remarkable internal model. What reaches consciousness is the reconstruction, not the raw input.
And while the human brain has a huge, overlapping set of neurons, only a small fraction of those are firing at any given moment. It’s a biological necessity, but also a matter of efficiency. The brain itself is a system built to move signals forward but learning itself has to run backward. It’s all about feedback. Muscles, senses, neuromodulators… all of them feed information upstream, changing how neurons behave.
Evolutionary learning, neural learning, and machine learning – are really versions of the same principle: dynamically stable symbiotic prediction. Living things survive by predicting the future, predicting themselves, and predicting each other, and doing so across many timescales. That’s the root of intelligence – not labels, not single rewards, but continual modeling of the world as it unfolds.
Before we wrap this lesson up, let’s get into the one area of AI that most of us have some familiarity with: language models.
Products like ChatGPT were seen as quite the breakthrough, but let’s knock language off its pedestal a little bit. Sure, humans use it at a very sophisticated level, but plenty of animals communicate richly – dolphins, whales, parrots, even prairie dogs have surprisingly detailed alarm calls. What’s ultimately important about language is what it does: it lets minds reveal themselves to each other. It’s basically a social compression system. Your brain takes this huge, messy, sensory-soaked inner world and squeezes it into a compact code that someone else can make sense of. Sometimes that code is a yelp of pain; sometimes it’s Hamlet.
Without language, we could still guess at what others think, but with it, we can exchange memories, plans, abstractions, and all the layers of theory-of-mind – our thoughts about other people’s thoughts.
There are a few milestones to language, with the hardest tricks being compositionality, and the capacity to express abstractions. But, once a system can symbolize things like “self,” “other,” “next week,” “maybe,” or “if-then,” then you’ve got something that can describe pretty much any mental universe it can imagine.
Language modeling used to rely on grammar rules or symbolic logic, and that never worked because natural language isn’t tidy. Neural nets changed everything, especially when they were trained on next-token prediction – which is the process of having the model predict the next word, or token, in a sequence based on the preceding words. This kind of prediction forces a model to learn almost everything humans know about the world. If a system can guess the right word after “I dropped the bowling ball on the violin, so I had to get it (blank),” it must understand both language and physics enough to know which object would need to be repaired.
The evolution of AI language models can be seen as going from RNNs to Transformers. RNNs was a machine learning framework that struggled with long-range memory. When that framework was upgraded to Transformers that problem was solved. It created a new kind of attention, where the model could let each word “look at” any other word – in context. It made room for rich, parallel reasoning. Oddly enough, even though Transformers weren’t designed to mimic biology, their tools echoed things we see in the hippocampus and grid-cell systems.
But Transformers also have a weird quirk: they don’t remember how they computed something. Each token is generated fresh, without internal continuity. That’s why they sometimes solve a math problem correctly, then give a bogus explanation. If this happens, all you need to do is prompt the model to “think step by step.” Just like a student writing out their work, the model performs better when it breaks reasoning into pieces.
It goes to show you, language is the scaffolding of thought. Step-by-step reasoning improves AI performance just as it helps humans build and accumulate knowledge.
Let’s close things out by looking at what the future might hold for artificial intelligence.
Right now, AI isn’t exactly at the same level of human intelligence, but it is actually quite close to the kind of intelligence we see in other lifeforms.
For example, bees show a whole menu of abilities we’d consider cognitive: flexible learning, generalization, long-term working memory, even restraint and delayed gratification. They build and repair hives in ways that adapt to weird situations, recognize shapes and patterns across senses, and make cautious choices based on past bad experiences. In other words, they do the small-brain version of what we might call reasoning. All of this is the kind of intelligence Transformers are capable of as well. Even small models can read little stories and reason about them. That puts them roughly in the bee-brain range.
Now, when people talk about AI’s trajectory, they often picture a neat three-act play: narrow AI, then general AI, then superintelligence and the Singularity beyond it. In practice, the jump to generality already happened – quietly – when we started pretraining giant sequence models and talking to them interactively.
Once that switch flipped, progress stopped looking like a series of thresholds and started looking like a long, steep climb. It’s a pattern we’ve seen before. Early computers were specialized contraptions, until the programmable, digital era of general-purpose computing arrived. After that, everything accelerated on a smooth exponential curve. AI is following the same script: the big transition has already occurred, and what we’re seeing now is the acceleration phase.
But perhaps the more interesting question is what this means for life on Earth. These technological transitions reshape things. Agriculture, cities, industry, electricity – each leap rewires how we do things, who depends on whom, and how information flows.
AI could very well be the next one. Not because machines are replacing us, but because of how humanity has weaved itself into a larger and larger collective mind. Each major transition has made individual humans less self-sufficient and more dependent on infrastructure – sometimes to a frightening degree. A hunter-gatherer can survive alone for days in the wild; a city dweller can’t survive a day without power, water pumps, supply chains, and communication systems. Every new layer of technological symbiosis raises both our capabilities and our vulnerability. The rise of AI continues that pattern: immense new powers wrapped in new dependencies we barely understand yet.
This may be the accurate take: to treat AI as a shift in our evolutionary story that’s already underway. Intelligence is becoming more collective, more hybrid, and much more distributed. The pressing questions now aren’t about some hypothetical superintelligence arriving out of nowhere, but about how this new layer of cognition changes our risks, reshapes our institutions, and alters what we can build together.
The main takeaway of this lesson to What Is Intelligence? by Blaise Aguera y Arcas is that intelligence isn’t a mysterious essence reserved for humans – it’s a natural outcome of life’s constant struggle to predict, adapt, and stay dynamically stable. Over time, the pressures of staying alive produced richer nervous systems, more flexible behavior, and ultimately the social, self-reflective minds we recognize in ourselves. Intelligence is just the gradual layering of better prediction, richer internal models, and more complex feedback loops. Our modern AI systems evolved from the same principles. Cybernetics and early computers pushed engineers toward machines that sense the world, act on it, and update their behavior – mirroring nature’s own solutions. Perceptrons and neural networks eventually picked up this thread and gave us architectures that can generalize and approximate the messy nonlinear functions real cognition depends on. While today’s systems still lack continual learning and on-the-fly adaptation, they reflect the same pattern that shaped biological intelligence. AI isn’t an alien rival but the latest chapter in a very old evolutionary story – one that links molecules, animals, societies, and machines through the universal logic of prediction.
What is Intelligence? (2025) repositions AI not as a looming alien mind, but as a natural continuation of life’s long, messy story of evolution, cooperation, and prediction. It weaves together bacteria, brains, cities, and neural networks to show how intelligence emerges wherever systems learn to model themselves and their world. It takes us through the past, present, and future of AI, while describing our place in it.
As the title suggests, in this lesson we’re going to take a big-picture look at the concept of intelligence. But not only that. We’re going to lay out the author’s bold message, which is that modern AI isn’t some clever imitation – it’s a real expression of intelligence, built on the same underlying principle that drives all living minds: prediction. The main argument is that intelligence shouldn’t be thought of in terms of biology, but rather computation. Understanding prediction as the engine of intelligence reframes everything – from how life began, to how organisms survive, to why large AI models suddenly work as well as they do.
This perspective pushes aside old debates about whether AI is “really” intelligent and instead turns our attention toward the bigger story: how prediction scaled from molecules to minds, how it now extends into machines, and how these new symbiotic intelligences might reshape our world. So let’s get into it by going back some four billion years ago to when life first took shape.
When we rewind Earth’s history to its earliest scenes, around 4.6 billion years ago, we’re looking at a world that barely resembles today. It’s known as the Hadean Eon, when the planet consisted of oceans of lava, a sky thick with volcanic gases, and a surface that was battered by asteroids big enough to boil the atmosphere. Somehow, within this chaos, the chemistry of life began to brew.
While the exact spark may never be known, one of the more attractive theories takes place in the deep sea. At hydrothermal vents – towering constructions built from minerals and heat – water, gas, and rock interact inside tiny porous chambers. These pores naturally create the kind of microscopic batteries that drive chains of chemical reactions that form self-sustaining cycles that look a lot like the metabolic “motor” still humming inside every cell today. It’s as though our own cells carry a memory of those early oceans.
As life took shape, it did so in bold leaps. Once a bacterium was engulfed by an archaeal cell, and instead of being digested, the two formed an alliance that resulted in mitochondria – which in turn opened the door to multicellular life, nervous systems, animals, and eventually us.
Symbiosis, and big, evolutionary leaps aren’t limited to biology. Technology evolved in similar bursts, each breakthrough forming when separate parts came together to create something new, like combining stone, wood, and sinew to make hunting tools.
Now, you may be hesitant to draw parallels between biological life and computing, but early computer pioneers like Alan Turing and John von Neumann saw the link. They imagined machines that could read instructions, modify those instructions, and even build copies of themselves – which is precisely the kind of logic of self-replication that DNA serves. So it’s hardly a stretch to say that a cell is both part chemistry and part computation.
To put it simply: reproduction requires computation. In nature, millions of tiny agents – ribosomes, enzymes, morphogens – operate in parallel, using randomness and local rules to achieve a larger unified function.
That same principle shows up in recent experiments. In 2023, researchers tried to recreate a sort of digital primordial soup environment, to see if the same kind of mutations and self-replicating patterns could emerge. Into the soup went thousands of simple digital tapes that contained both rudimentary code and data, and the ability to modify itself. Sure enough, after millions of interactions, the tapes do begin to reproduce. All of a sudden, a phase-transition happens, and they shift from noise to self-replicating patterns.
From randomness, replicators emerge; once replicators exist, evolution takes over. And as we’ll see in the next section, that rising complexity can give way to intelligence.
At a very basic level, intelligence is all about survival. If we zoom way, way in and observe a lone bacterium drifting through its liquid world, we can see the core of intelligence forming. Not the kind we usually picture – with neurons firing or thoughts bubbling up – but a minimal, stripped-down version whose only real job is to keep the cell alive from one moment to the next.
Life can’t just replicate blindly. It needs a way to build itself, maintain itself, and do all of that inside a real environment that never stays still. That messiness changes everything. It means a living system needs a boundary – a membrane – to protect its delicate chemistry. But the boundary can’t be sealed shut, because energy and matter have to flow in and out in order to survive.
Inputs and outputs are basically the cost of being alive, but every input is also information. Every organism needs to know what to consume and what to avoid. That’s the first glimmer of intelligence: using information to stay alive.
On the inside, a bacterium has its own tiny control network – a biochemical “brain” made of genes and proteins that monitor hunger, energy levels, and all sorts of microscopic states. These internal signals blend with external ones to steer behavior. And that behavior is simple but clever. Intelligent, even. Many bacteria swim with a little rotating propeller called a flagellum, and they manage to swim toward food by comparing now with just-a-moment-ago.
When the scent of food is rising, they run longer. When it’s falling, they tumble more often. Whether it’s food, heat, or light, bacteria follow the same playbook: infer a pattern using only a sequence of tiny, random events over time. In other words, they solve problems using statistics and probability.
The whole point of this is homeostasis – keeping internal conditions in the comfort zone. So we can even represent it all in a computation model that knits together three streams at once: external inputs (X), internal state (H), and possible actions (O). This joint distribution P(X,H,O) is the organism’s entire worldview. It predicts not just what’s out there, but what might happen if it runs, tumbles, or switches a gene on or off.
Here’s where evolution comes in. Over generations, the population tunes itself toward better and better internal models – leaner, more general, more predictive. And once an organism has a model of the world that includes itself – a model that shapes its actions, which in turn shape its future – the bacterium has a goal: staying alive. That goal is the seed of purposiveness, which in turn is the beginning of intelligence, and the foundation for everything that comes next.
Let’s fast forward to the turn of the eighteenth century, when one of the stars of the Enlightenment era, the German mathematician and philosopher Gottfried Wilhelm Leibniz, helped usher in the age of the mechanical calculator.
Leibniz was one of the first people to see broader, loftier uses for calculations. He dreamed of an algorithm that could give a true or false answer to any statement – a day when philosophers would settle debates the way accountants balance ledgers.
But for the next few centuries Leibniz’s dream was put on the back burner in favor of commercialism. In the eighteenth and nineteenth centuries, calculations and early computers helped usher in the Industrial Revolution. Thanks to machines like the Analytical Engine, created by British mathematician Charles Babbage, mechanical computers became extensions of the factory floor.
Still, there were a couple of people who recognized a greater potential. For the British mathematician Ada Lovelace, computers like the Analytical Engine could lead us to a general theory that linked machines, mathematics, and the very patterns that shape life. She saw a future where we might discover the “calculus of the nervous system.”
By the early 1900s, neuroscience finally caught up. Experiments indeed picked up on electrical spikes in the nervous system, which led to a wild idea: maybe neurons were logic gates and the brain was indeed a kind of analytical engine. In 1943, two American scientists, Warren McCulloch and Walter Pitts, went all in on that view, sketching a neuron-as-logic-gate model that would end up influencing the design for digital computers.
Those computers took off and became the fast, precise, obedient engines of binary thinking. And they set the tone for GOFAI, or good old-fashioned artificial intelligence: machines were the logical, rational robots while humans were the intuitive, emotional beings.
Eventually, as scientists continued to zoom in on the human brain, they found that neurons didn’t behave like straight-forward logic gates. They behaved like complex, nonlinear, dynamic systems. The brain looked less like a theorem prover and more like the living, breathing, messy organism it belonged to.
But, the GOFAI model was easy to implement, so it moved ahead while the more human-like AI found a home in the mostly forgotten branch of cybernetics. This was a school of thought that saw minds and machines as prediction engines – systems that learn from experience, regulate themselves, adapt to uncertainty, and close feedback loops with the world. It’s a worldview grounded in biology, behavior, feedback, and prediction, rather than pristine logic.
Instead of batch processing, cybernetics focused on real-time feedback loops – systems that sense the world, act on it, and constantly update their behavior. Gun turrets, bomb-sight computers, and early flight simulators all worked this way.
But while the cybernetics theory was grand, it would take some time before the technology caught up. In that time, however, the core cybernetic insight – that intelligence emerges from prediction, feedback, learning, and continuous values – never went away.
So how did the early, wobbly steps of cybernetics lead to modern AI advancements? While we can’t get into every detail, we have to mention the perceptron. The perceptron is the simplest kind of artificial neuron, the kind you can bundle together, feed a ton of photos, and ask it to recognize something like bananas.
But the real ability for a machine to learn, through trial-and-error, how to identify a banana, is computation. The computation is what allows the machine to learn – for the neuron to fire in a positive way when it makes the right decisions. This is known as an activation function, and one of the most commonly used is called ReLU, short for Rectified Linear Unit.
The name is a little misleading, because ReLu is a feat of nonlinear problem solving that makes it possible for a machine to recognize a banana no matter how it’s lit, angled, ripened, or photographed. In biological terms, it’s like neurons firing only when their preferred pattern shows up strongly enough.
When the process becomes this repeated predictive routine, what clicks into place is something called transfer learning, where the AI has learned enough to start applying some of the same patterns for completely different objects. That’s why you can successfully add an apple perceptron with barely any examples.
And that sets the stage for unsupervised learning. Instead of labeling millions of images, you mask out random chunks and ask the network to inpaint them – to, fill in the blanks and predict what’s missing. Do that well, and you’ve implicitly learned the world: edges, colors, objects, depth, categories. You’ve built the representations needed for almost any downstream task. It’s the same trick language models use with missing words. Prediction and modeling are two sides of the same coin.
Our own visual system runs a kind of lifelong masked-autoencoder training. Most of what we “see” is low-resolution, noisy, or obscured – but because we’re constantly flicking our eyes around, checking our guesses, and filling in gaps, we build a remarkable internal model. What reaches consciousness is the reconstruction, not the raw input.
And while the human brain has a huge, overlapping set of neurons, only a small fraction of those are firing at any given moment. It’s a biological necessity, but also a matter of efficiency. The brain itself is a system built to move signals forward but learning itself has to run backward. It’s all about feedback. Muscles, senses, neuromodulators… all of them feed information upstream, changing how neurons behave.
Evolutionary learning, neural learning, and machine learning – are really versions of the same principle: dynamically stable symbiotic prediction. Living things survive by predicting the future, predicting themselves, and predicting each other, and doing so across many timescales. That’s the root of intelligence – not labels, not single rewards, but continual modeling of the world as it unfolds.
Before we wrap this lesson up, let’s get into the one area of AI that most of us have some familiarity with: language models.
Products like ChatGPT were seen as quite the breakthrough, but let’s knock language off its pedestal a little bit. Sure, humans use it at a very sophisticated level, but plenty of animals communicate richly – dolphins, whales, parrots, even prairie dogs have surprisingly detailed alarm calls. What’s ultimately important about language is what it does: it lets minds reveal themselves to each other. It’s basically a social compression system. Your brain takes this huge, messy, sensory-soaked inner world and squeezes it into a compact code that someone else can make sense of. Sometimes that code is a yelp of pain; sometimes it’s Hamlet.
Without language, we could still guess at what others think, but with it, we can exchange memories, plans, abstractions, and all the layers of theory-of-mind – our thoughts about other people’s thoughts.
There are a few milestones to language, with the hardest tricks being compositionality, and the capacity to express abstractions. But, once a system can symbolize things like “self,” “other,” “next week,” “maybe,” or “if-then,” then you’ve got something that can describe pretty much any mental universe it can imagine.
Language modeling used to rely on grammar rules or symbolic logic, and that never worked because natural language isn’t tidy. Neural nets changed everything, especially when they were trained on next-token prediction – which is the process of having the model predict the next word, or token, in a sequence based on the preceding words. This kind of prediction forces a model to learn almost everything humans know about the world. If a system can guess the right word after “I dropped the bowling ball on the violin, so I had to get it (blank),” it must understand both language and physics enough to know which object would need to be repaired.
The evolution of AI language models can be seen as going from RNNs to Transformers. RNNs was a machine learning framework that struggled with long-range memory. When that framework was upgraded to Transformers that problem was solved. It created a new kind of attention, where the model could let each word “look at” any other word – in context. It made room for rich, parallel reasoning. Oddly enough, even though Transformers weren’t designed to mimic biology, their tools echoed things we see in the hippocampus and grid-cell systems.
But Transformers also have a weird quirk: they don’t remember how they computed something. Each token is generated fresh, without internal continuity. That’s why they sometimes solve a math problem correctly, then give a bogus explanation. If this happens, all you need to do is prompt the model to “think step by step.” Just like a student writing out their work, the model performs better when it breaks reasoning into pieces.
It goes to show you, language is the scaffolding of thought. Step-by-step reasoning improves AI performance just as it helps humans build and accumulate knowledge.
Let’s close things out by looking at what the future might hold for artificial intelligence.
Right now, AI isn’t exactly at the same level of human intelligence, but it is actually quite close to the kind of intelligence we see in other lifeforms.
For example, bees show a whole menu of abilities we’d consider cognitive: flexible learning, generalization, long-term working memory, even restraint and delayed gratification. They build and repair hives in ways that adapt to weird situations, recognize shapes and patterns across senses, and make cautious choices based on past bad experiences. In other words, they do the small-brain version of what we might call reasoning. All of this is the kind of intelligence Transformers are capable of as well. Even small models can read little stories and reason about them. That puts them roughly in the bee-brain range.
Now, when people talk about AI’s trajectory, they often picture a neat three-act play: narrow AI, then general AI, then superintelligence and the Singularity beyond it. In practice, the jump to generality already happened – quietly – when we started pretraining giant sequence models and talking to them interactively.
Once that switch flipped, progress stopped looking like a series of thresholds and started looking like a long, steep climb. It’s a pattern we’ve seen before. Early computers were specialized contraptions, until the programmable, digital era of general-purpose computing arrived. After that, everything accelerated on a smooth exponential curve. AI is following the same script: the big transition has already occurred, and what we’re seeing now is the acceleration phase.
But perhaps the more interesting question is what this means for life on Earth. These technological transitions reshape things. Agriculture, cities, industry, electricity – each leap rewires how we do things, who depends on whom, and how information flows.
AI could very well be the next one. Not because machines are replacing us, but because of how humanity has weaved itself into a larger and larger collective mind. Each major transition has made individual humans less self-sufficient and more dependent on infrastructure – sometimes to a frightening degree. A hunter-gatherer can survive alone for days in the wild; a city dweller can’t survive a day without power, water pumps, supply chains, and communication systems. Every new layer of technological symbiosis raises both our capabilities and our vulnerability. The rise of AI continues that pattern: immense new powers wrapped in new dependencies we barely understand yet.
This may be the accurate take: to treat AI as a shift in our evolutionary story that’s already underway. Intelligence is becoming more collective, more hybrid, and much more distributed. The pressing questions now aren’t about some hypothetical superintelligence arriving out of nowhere, but about how this new layer of cognition changes our risks, reshapes our institutions, and alters what we can build together.
The main takeaway of this lesson to What Is Intelligence? by Blaise Aguera y Arcas is that intelligence isn’t a mysterious essence reserved for humans – it’s a natural outcome of life’s constant struggle to predict, adapt, and stay dynamically stable. Over time, the pressures of staying alive produced richer nervous systems, more flexible behavior, and ultimately the social, self-reflective minds we recognize in ourselves. Intelligence is just the gradual layering of better prediction, richer internal models, and more complex feedback loops. Our modern AI systems evolved from the same principles. Cybernetics and early computers pushed engineers toward machines that sense the world, act on it, and update their behavior – mirroring nature’s own solutions. Perceptrons and neural networks eventually picked up this thread and gave us architectures that can generalize and approximate the messy nonlinear functions real cognition depends on. While today’s systems still lack continual learning and on-the-fly adaptation, they reflect the same pattern that shaped biological intelligence. AI isn’t an alien rival but the latest chapter in a very old evolutionary story – one that links molecules, animals, societies, and machines through the universal logic of prediction.
Comments
Post a Comment