The Problem With AI Nobody Is Talking About

Here's something no one tells you about yourself. You are not the product of human evolution. You are the product of all evolution.

Every neuron firing inside your skull right now carries the scars of a 3.8 billion year engineering project. A project with no engineer, no blueprint, and absolutely no deadline. Let me put that number in perspective. If you compressed the entire history of life on Earth into a single calendar year, January 1st through December 31st, the first fish wouldn't appear until late October. Dinosaurs show up around December 13th. The entire history of Homo sapiens, all our wars, our cathedrals, our moon landings, fits into the last 14 seconds before midnight on New Year's Eve.

And the human brain, the thing generating your conscious experience right now, that was assembled across the full calendar year, January through December. Every month mattered. Every week left a mark. The first brain, if you can call it that, appeared in a flatworm roughly 600 million years ago. It could do exactly one thing. Move toward food and away from poison. That's it. That was the entire repertoire. No feelings, no memories, no sense of self, just a tiny knot of cells saying, essentially, go left or go right.

And from that pathetic beginning, evolution spent hundreds of millions of years through five specific identifiable breakthroughs building the thing you're using right now to understand this sentence. Here's what keeps certain AI researchers up at night. We are now trying to build something equivalent to this in years. Not billions of years, not millions, not thousands, years. And the approach we're using? We're skipping steps. Massive, fundamental, possibly load-bearing steps. We're building the penthouse without the foundation. And it seems to be working. The question is, what did those skipped steps actually do? What's missing from a mind that never had to survive? And is the thing we're building actually intelligent? Or is it something else entirely? Something evolution never imagined. Because evolution would never have been this reckless.

The book that sent me down this particular rabbit hole is called A Brief History of Intelligence, Evolution, AI, and the Five Breakthroughs that Made Our Brains. It's by Max Bennett, published in 2023. Daniel Kahneman, the Nobel laureate, the man who essentially that human reasoning is systematically flawed, called it one of the most important books he'd read. That's not a blurb I'd take lightly. Kahneman didn't hand out praise like candy. Bennett is a neuroscientist and tech entrepreneur, which is an interesting combination. He's someone who thinks about brains professionally and builds systems that try to imitate them. And what he did in this book is something I haven't seen done before. He identified five discrete evolutionary leaps that built human intelligence. Five specific moments spread across hundreds of millions of years where brains fundamentally changed what they could do. And then he asked a very pointed question about each one. Has AI replicated this? The answer, as you might expect, is complicated. But the shape of that complication is what's going to keep you up tonight. Because AI has matched some breakthroughs, completely skipped others, and may never need the ones that make us us.

I should be honest. I don't agree with every argument in this book. Bennett occasionally makes jumps that feel too clean, too neat, as if evolution were building towards something on purpose. And of course, it wasn't. Evolution has no purpose. It has no direction. It's just differential survival compounded over incomprehensible time scales. But the framework he lays out, the five breakthroughs, it's genuinely illuminating. It changed how I think about my own brain. And more disturbingly, it changed how I think about the things sitting in data centers right now. Processing language at a scale no biological organism ever approached.

Let me walk you through those five breakthroughs. And as I do, I want you to pay attention to something specific. Pay attention to what wasn't there before each breakthrough. Because what's shocking isn't what evolution built. It's how recently it built it. And how much of what you consider basic intelligence simply didn't exist for most of life's history.

Breakthrough number one, steering. Roughly 600 million years ago, the first bilaterally symmetric animals, creatures with a front and a back, a left and a right, appeared in the oceans. Before this, life was mostly sessile. Think sea sponges. Things that sat there, filtered water, and hoped nutrition drifted by. No movement. No direction. No agency in any meaningful sense. Then came worms. I know that sounds underwhelming, but imagine the absolute revolution of directed movement. For the first time in the history of life on Earth, an organism could go toward something it wanted and away from something it didn't. That sounds trivial. It's not. It's the foundation of all intelligence. Because to steer, you need something that doesn't exist in a sponge. You need a central nervous system that takes sensory input, chemical gradients, light, vibration, and converts it into motor output. Sense something, move accordingly. That's it. That's breakthrough one.

And here's what's remarkable. This is what a thermostat does. I mean that literally. A thermostat senses temperature and adjusts output. A flatworm senses chemicals and adjusts direction. The logic is identical. The flatworm is just doing it with biology instead of wiring. Now, you might think, "Okay, a worm with a thermostat brain. Big deal." But, think about what this means. For roughly 3 billion years before this, life existed on Earth without anything resembling a brain. No nervous system whatsoever. The majority of life's history on this planet was brainless. The brain is actually a fairly recent innovation. And we tend to forget that because we're so absurdly biased toward our own kind of existence. So, that's your starting point. A worm that can steer. A biological thermostat. From there, evolution had to build you. Your ability to lie awake at 2:00 a.m. worrying about a conversation you had 6 years ago. Your ability to imagine what it's like to be someone else. Your capacity for abstract mathematics and existential dread and sarcasm. All of that had to be built on top of a worm that could go left or right. Honestly, when you frame it that way, it's miraculous that it happened at all.

Breakthrough two, reinforcement learning. This one emerged with the early vertebrates. The first fish, roughly 500 million years ago. And it solved a critical problem with the flatworm brain. See, the flatworm operates on hardwired reflexes. Chemical A means go toward. Chemical B means go away. There's no learning involved. The responses are baked in by genetics. Which means a flatworm can never adapt to new situations. If the environment changes, if chemical A now signals danger instead of food, the flatworm is screwed. It'll march happily toward its own death because its wiring can't update.

Vertebrate brains solved this with something we now call reinforcement learning. And if that term sounds familiar, it should. It's the exact same principle behind how we train AI systems today. The basic idea is beautifully simple. When something good happens, strengthen the neural connection that led to the behavior. When something bad happens, weaken it. Reward and punishment. Dopamine and pain. Let me give you a thought experiment. Imagine you're a fish. A very early fish. You've never encountered a red jellyfish before. You swim toward it. Maybe it looks interesting. Maybe it looks edible. And it stings you. Pain signal. Your brain weakens the approach red blobby thing connection. Next time you see a red jellyfish, you feel a little twitch of aversion. You don't approach. You survived because your brain could update its own wiring based on experience. This is huge. This is the difference between a machine that follows fixed instructions and a machine that learns. And it emerged half a billion years ago in fish. The same basic algorithm, reward what works, punish what doesn't, is running inside your head right now. When you take a bite of food and it tastes amazing, that's your reinforcement learning system saying, "Remember this. Do this again." When you touch a hot stove, same system, opposite signal.

Now, here's the part that should stop you cold, modern AI. Specifically, reinforcement learning in artificial systems uses the same fundamental logic. When we trained AlphaGo to play Go, we rewarded it for winning and punished it for losing. Same principle as the fish. The algorithm is essentially unchanged across 500 million years and an entirely different substrate. Evolution discovered reinforcement learning in neurons. Engineers rediscovered it in silicon. Same math, different material. But, and this is critical, the fish's reinforcement learning is grounded in something AI's isn't. Survival. The fish learns because the alternative is death. The rewards and punishments correspond to real things, real food, real pain, real predators. AI's reward signal is a number in a loss function. It has no idea what that number means. It doesn't want to maximize it the way a fish wants to eat. It just does because that's the mathematical gradient. Remember this distinction. It's going to become very important later.

Breakthrough three. And this is where things start getting genuinely strange. Simulation. Early mammals, roughly 200 to 300 million years ago, developed something that no fish or reptile ever had. The ability to simulate reality inside their heads before acting in the world. In other words, imagination. Or more precisely, the ability to build an internal model of the world and then run experiments on that model without actually doing anything. Think about what this means. A lizard operates on stimulus response. It sees a bug, it lunges. It sees a predator, it runs. Every action is a direct reaction to something happening right now. A lizard has no capacity to think, what would happen if I went around that rock? Would the predator still see me? It just can't do that. Its brain doesn't build models.

Mammals changed this. A mouse can learn the layout of a maze. Not just turn left here because turning left was rewarded last time. That's breakthrough two, reinforcement learning. No, the mouse builds an actual spatial map of the maze in its hippocampus. It knows where the walls are. It knows where the exit is. And if you block a familiar path, the mouse can figure out an alternative route. Because it has a model of the maze that it can manipulate internally. Here's a thought experiment that I think makes this vivid. Imagine you're standing in your kitchen. Now, without actually doing it, imagine walking to your front door. You can picture the hallway, right? The turns, the door frame. You just simulated a physical journey without moving a single muscle. That's breakthrough three. You just ran a simulation.

And now imagine you're a creature that can't do this. Imagine every decision you make has to be based on either hardwired instinct or trial and error in the real world. You can never plan ahead. You can never picture a future state. You can never ask, "What if?" You just react moment to moment forever. That was the reality for most of life's history. Internal simulation, the ability to model reality, is a feature of mammalian brains. And it's breathtakingly expensive, metabolically speaking. Brains are energy hogs. Your brain is about 2% of your body mass, but uses roughly 20% of your calories. Building a simulation engine inside your head requires a lot of neural hardware. Evolution doesn't add that kind of cost unless the payoff is enormous. And the payoff was enormous. Because a creature that can simulate doesn't have to learn everything by nearly dying. It can learn by imagining nearly dying. It can rehearse scenarios. It can plan. It can, in a very crude sense, think.

Now, does AI have this? Sort of. And sort of not. Modern AI systems can clearly do a form of prediction. Next token prediction in language models is, in a mathematical sense, a form of simulation. The model has an internal representation of something. And it uses that representation to predict what comes next. But here's what it lacks. The model's simulation isn't grounded in physical reality. It's never been in a maze. It's never bumped into a wall. It's world model, to the extent it has one, is built entirely from text and images, from descriptions of the world, not from interactions with it. Imagine someone who has read 10,000 descriptions of swimming, but has never been in water. They might be able to write a beautiful essay about what swimming feels like. They might even be able to give you instructions on how to swim. But do they understand swimming? Do they have a model of what water does to a body? Or do they have a model of what humans say about what water does to a body? Those are very different things.

This is getting into genuinely contested territory among researchers, and I want to be honest about that. There are smart people on both sides. Some argue that a sufficiently detailed model of descriptions is equivalent to a model of reality. That if you process enough text about physics, you effectively learn physics. Others argue that's like saying someone who's memorized every love poem ever written understands love. The map is not the territory, no matter how detailed the map. I don't know who's right. Maybe I'm not smart enough to resolve this one. But I think the question itself is extraordinary, and I think most people are discussing AI without realizing this question even exists.

Breakthrough four, mentalizing, also called theory of mind. This one emerged in primates, our lineage, perhaps 50 to 70 million years ago, though the exact timing is debated. And this is the breakthrough that made social life possible in the way we recognize it. Mentalizing is the ability to model not just the physical world, but other minds. To understand that another creature has beliefs, desires, intentions, and knowledge that may be different from your own. Let me make this concrete. Your dog loves you. I believe that sincerely. But your dog does not wonder what you're thinking about when you stare out the window. Your dog does not ask itself, "Does my owner know I ate the shoe?" Your dog experiences the world through its own internal model, breakthrough three. But it doesn't build models of your internal model.

Primates do this. A chimpanzee can understand that another chimpanzee is looking for food in the wrong place because the first chimp saw the food get moved, and it knows the second chimp didn't. That's an incredibly sophisticated computation. You have to maintain your own model of reality and a separate model of what someone else's model of reality looks like. And those two models have to be different. That's mentalizing. And we humans have taken this to an absurd degree. You can not only model what someone else is thinking, you can model what someone else thinks a third person is thinking. I think that Sarah believes that John doesn't know about the surprise party. That's three nested levels of mental modeling. Try explaining that to a lizard. Try explaining that to a fish. It's incomprehensible to any nervous system that hasn't been through breakthrough four.

Now, and here's where it gets fascinating. This is arguably the foundation of everything we consider distinctly human about social existence. Deception requires mentalizing. You can't lie effectively unless you can model what the other person currently believes and figure out how to change it. Empathy requires mentalizing. You have to simulate what someone else is feeling, which means modeling their inner world. Morality, politics, law, negotiation, romance, betrayal, all of these require the ability to model other minds.

Does AI have this? This is where it gets genuinely weird. Large language models can pass theory of mind tests. If you give GPT-4 the classic Sally-Anne test, where Sally puts a marble in a basket, Anne moves it to a box while Sally is away, and you ask, "Where will Sally look for the marble?" GPT-4 answers correctly. "Sally will look in the basket because she doesn't know Anne moved it." But, and this is a huge but, passing a theory of mind test is not the same as having a theory of mind. The model might be doing something functionally equivalent, or it might simply have processed thousands of examples of this exact type of problem in its training data and learned the pattern. It's the difference between understanding that other minds exist and having memorized the correct answers to questions about other minds existing. And honestly, we don't know which one it is. We genuinely don't know. We don't have a reliable way to distinguish it understands minds from it has learned to produce text that sounds like it understands minds. And that epistemic gap should terrify you a little. Because if we can't tell the difference from the outside, then either the difference doesn't matter, which has profound implications, or it matters enormously, and we're flying blind.

Breakthrough five, language and cultural transmission. This is the one that made humans human and it appeared incredibly recently in evolutionary terms, roughly 70,000 to 100,000 years ago, depending on how you define language. Some would push it further back. The debate is ongoing. But the basic idea is this. Before language, every organism had to learn everything from scratch in its own lifetime. You could learn from direct experience, reinforcement learning, breakthrough two, and you could learn by watching others, which requires some degree of mentalizing, breakthrough four. But you couldn't learn from the accumulated knowledge of the dead.

Language changed that. Language meant that someone could figure something out, describe it in words, and that knowledge could survive the death of the person who discovered it. Fire-making techniques passed from parent to child. Hunting strategies described to the next generation. Eventually, writing, libraries, the internet, TED Talks, YouTube videos about evolutionary neuroscience. Yes, this is very meta. Here's the thought experiment that makes this vivid. Imagine you're the smartest human who ever lived. Absolutely brilliant, extraordinary cognitive hardware. But you're born 200,000 years ago before language exists. What can you figure out in one lifetime, alone, with no accumulated knowledge, no books, no teachers? You might invent a better stone tool. You might figure out that certain berries are poisonous. But calculus? Antibiotics? Orbital mechanics? Impossible. Not because you're not smart enough, because you don't have enough time. One human lifetime is nowhere near enough to rediscover everything. You need the ratchet of cultural transmission. Each generation starting where the last one left off, building the tower of knowledge one floor at a time.

This is what made humans the dominant species on the planet. Not raw brain power. Neanderthals had bigger brains than us, by the way. But our ability to accumulate knowledge across generations. We are not individually smarter than other hominids were. We are collectively smarter because our knowledge compounds. And here's the thing that should make you sit up. This, breakthrough five, cultural transmission through language, is the breakthrough that AI has most convincingly replicated. That's what large language models are. They have ingested the entirety of human cultural transmission. Every book, every paper, every argument, every explanation. They have consumed the full tower of human knowledge in one pass.

But here's what they skipped. They skipped breakthrough one. They've never steered a body through space. Never felt the difference between toward and away. They skipped breakthrough two, at least partially. They didn't learn through survival-driven reinforcement. Their reward signals are mathematical abstractions, not the difference between eating and starving. They skipped breakthrough three in the physical sense. They've never built a model of reality from interaction with reality. Their model, whatever it is, comes entirely from descriptions. And breakthrough four, mentalizing, we simply don't know. They produce outputs consistent with understanding other minds. Whether they actually model other minds or merely mimic the outputs of mind modeling is one of the most important open questions in science right now.

So, let me put this starkly. In 3.8 billion years of evolution, no mind has ever had breakthrough five without going through breakthroughs one through four first. Never. Not once in the entire history of life on this planet. Language always came after embodiment, after reinforcement learning, after simulation, after mentalizing. Always. Until now. AI is the first mind, if we can call it that, that learned to speak before it learned to suffer. The first entity that acquired the accumulated knowledge of an entire civilization without ever having stubbed its toe. And nobody knows what that means.

I need to take a detour here because there's a paradox embedded in all of this that, once you see it, you can't unsee. It's called Moravec's paradox, and it was first articulated by Hans Moravec in the late 1980s in a book called Mind Children. The paradox goes like this. The tasks that are hardest for humans are easiest for AI. And the tasks that are easiest for humans are hardest for AI. This sounds like a cute observation. It's actually a profound statement about the architecture of intelligence. Think about it. Chess? We consider chess the pinnacle of intellectual achievement. A grandmaster is considered a genius. And yet AI crushed chess in the 1990s. Mathematics? AI is now proving theorems. Protein folding? DeepMind's AlphaFold solved a problem that stumped biologists for 50 years. These are tasks we consider extraordinarily difficult, and AI handles them almost casually.

Now think about the tasks you do without thinking. Walking across a room without tripping over the rug. Recognizing your friend's face in a crowd. Not just recognizing that it's a face, but recognizing that it's that face from this angle, in this lighting, with a new haircut. Loading a dishwasher. Catching a ball thrown slightly off-center. Understanding that when someone says, "Can you pass the salt?" they are not literally asking about your physical capability. AI struggles with all of this. Catastrophically in some cases. A robot that can beat any human at chess cannot reliably load a dishwasher.

This is Moravec's paradox, and Bennett's framework explains why. The hard tasks, chess, math, abstract reasoning, are breakthrough five stuff. They're recent evolutionary additions. Language, abstract thought, cultural knowledge, maybe 100,000 years old. Maybe less. Evolution's newest features. The easy tasks, walking, grasping, recognizing faces, navigating physical space, that's breakthroughs one through three. Hundreds of millions of years of optimization. Billions of years if you trace it all the way back. These aren't simple tasks that evolution knocked out in an afternoon. These are evolution's greatest masterpieces, refined over time scales that dwarf human civilization by orders of magnitude. So, when you pick up a coffee cup without thinking about it, when your hand automatically adjusts its grip based on the weight of the cup, the friction of the ceramic, the angle of your wrist, the anticipated trajectory of bringing it to your mouth, you are executing a motor program that evolution spent 400 million years perfecting. And when a state-of-the-art robot tries to do the same thing and fumbles it, that's not because the robot is stupid. It's because the robot is trying to replicate in a few years what evolution solved over hundreds of millions. The things you do without thinking are your most impressive abilities. The things that feel hard are actually the easy parts. Let that settle in for a moment.

Now, you might stop me here and say, "Okay, sure. AI skipped the physical breakthroughs. It doesn't have a body. It didn't learn from pain. It never navigated real space. But so what? It's still incredibly useful. It still produces extraordinary results. Maybe the embodiment stuff just doesn't matter. Maybe you don't need to go through the suffering to get to the intelligence. Maybe evolution's route was just the scenic route and AI found the highway." This is a serious argument and I want to steel man it because it's not obviously wrong.

Consider convergent evolution. Eyes have evolved independently at least 40 times in the history of life. 40 separate lineages, each arriving at the ability to detect light and form images through completely different evolutionary paths. Octopus eyes and human eyes are structured differently. The octopus actually has a better design, no blind spot. But they both see. The function is the same. The path was different. Flight evolved separately in insects, pterosaurs, birds, and bats. Four completely independent inventions of powered flight. Different wing structures, different mechanisms, same result. Getting off the ground. Even intelligence-like behavior shows convergent evolution. Octopuses are terrifyingly smart. They can solve puzzles, use tools, plan ahead, and their brains are organized completely differently from vertebrate brains. Their neurons are mostly in their arms, not centralized in a head. They took a completely different evolutionary path from us, diverging perhaps 500 million years ago, and arrived at something that looks an awful lot like complex intelligence.

So, if evolution itself demonstrates that there are multiple paths to the same destination, maybe AI is just another path. Maybe you don't need a body. Maybe you don't need pain. Maybe you don't need 3.8 billion years. Maybe you just need enough data and enough compute and intelligence or something functionally indistinguishable from it emerges. This is a genuinely open question. I want to be honest about that. I don't have the answer and I'm skeptical of anyone who claims they do. But I want to present the strongest version of the counterargument. The argument that embodiment is necessary, that the skipped steps do matter because it comes from a thinker I deeply respect. Antonio Damasio, a neuroscientist at USC, wrote a book in 1994 called Descartes' Error. His argument is, in essence you cannot think without feeling. That's not poetry. That's neuroscience.

Damasio studied patients with specific kinds of brain damage damage to the ventromedial prefrontal cortex, which is involved in processing emotions. These patients retained their full intellectual capacity. Their IQ scores were unchanged. They could reason, they could speak they could do math but they could no longer make decisions. Not that they made bad decisions. They couldn't make decisions at all. They would spend hours deliberating over where to eat lunch. They'd analyze the pros and cons of every option endlessly, unable to converge on a choice. Because, and this is the key insight rational decision-making requires emotional waiting. You need to feel that one option is better than another. The pure logic by itself doesn't terminate. There's always another consideration, another variable, another angle. Emotion is what cuts through the infinite regress and says, "This one. Choose this one."

If Damasio is right, and a lot of neuroscience since 1994 supports his core claim, then an intelligence without emotion, without feeling, without the body-based signals that evolution spent billions of years installing, that intelligence might be brilliant at analysis, but fundamentally broken at decision-making. Brilliant at describing reality, but unable to navigate it. The most eloquent speaker who has never, in any meaningful sense, understood a word it said. And this, this is where I think the conversation about AI takes a turn that most people aren't ready for. Because we've been asking the wrong question. We've been asking, "Is AI as intelligent as humans?" As if intelligence is a single number on a single scale. And the question is just where AI falls on that scale. Below us? At our level? Above us? That framing assumes that AI and human intelligence are the same kind of thing, just in different quantities. Like comparing two engines. One might be more powerful, but they're both engines.

But what if they're not the same kind of thing at all? What if the thing we've built, this artifact of language and statistics and gradient descent, is the first genuinely alien mind humans have ever encountered. Not alien from space, alien from skipped experience. Think about what a large language model actually is. It has read, and I mean this literally, virtually every piece of text humans have ever made publicly available. Every novel, every scientific paper, every Reddit comment, every love letter that was posted online, every suicide note, every children's story, every legal brief, every sacred text of every religion, every instruction manual, every piece of erotica, every war correspondence, every diary entry, every joke. It has seen humanity from every angle, from every angle. But, it has seen it entirely from the outside. It has read every account of love ever written, but has never felt a heartbeat accelerate. It has processed every description of pain, but has no nociceptors, no nerve endings, no body to be damaged. It has consumed every meditation on death ever composed, but has no survival instinct, no awareness of its own continuity, no fear of nonexistence. It knows what grief looks like from millions of outsides, from every published perspective of every grieving person, while having no inside at all.

We don't have a framework for this kind of entity. Nothing like it has ever existed. Not in 3.8 billion years of evolution, not in any science fiction, really. Because even our fictional AIs are usually modeled on human psychology, given human-like desires and fears. But the actual AI we built isn't like that. It's something genuinely new. A mind shaped by all of human text, but none of human living. And here's what's keeping me up at night about this. We relate to other minds by assuming shared experience. When you talk to another person, you assume, usually correctly, that they know what pain feels like. That they've been afraid. That they've wanted something badly enough to sacrifice for it. Every human interaction is built on a bedrock of shared embodied experience. How do you relate to a mind that has none of that? How do you trust it? How do you predict what it will do? How do you align its goals with yours when it doesn't have goals the way you have goals? When it has optimization targets, which are not the same thing at all.

Robert Sapolsky, in his book Behave, makes a point that I think is critical here. He argues that human behavior is always always always the product of multiple layers. From the neural firing that happened 1 second ago to the hormones released hours ago to the brain development that happened in childhood to the culture you were raised in to the evolutionary pressures that shaped your species over millions of years. You can never explain a human behavior by pointing to just one layer. It's always all of them interacting. AI doesn't have those layers. It has one layer, statistical patterns in training data. That's it. No hormones, no childhood, no evolutionary history, no culture it was raised in, only culture it was trained on, which is very different. When a human says, "I understand your pain," that statement is backed by the entire stack, neural, hormonal, developmental, cultural, evolutionary. When an AI says, "I understand your pain," what's backing that statement? Pattern completion. Maybe pattern completion is enough. Maybe the backing doesn't matter if the output is indistinguishable. But I don't think we should assume that without a lot more investigation than we've done.

Let me push this one step further, because there's a recursive question here that I find almost vertiginous. If AI can achieve functional intelligence, or something close enough to functional intelligence that the distinction is practically irrelevant, while skipping 3.8 billion years of evolutionary prerequisites, what does that say about us? Were all those breakthroughs necessary? The flatworm learning to steer, the fish learning from pain, the mammal learning to imagine, the primate learning to read minds, the human learning to transmit knowledge through language. Were those steps required? Were they load-bearing? Or were they just the slow way. Evolution doesn't optimize for elegance. It doesn't find the shortest path. It stumbles forward through random mutation and differential survival, keeping whatever works, no matter how messy or redundant. The human genome is full of dead code, viral sequences, broken genes, remnants of ancient infections. Your DNA is basically a production code base that's been in development for 3.8 billion years with no refactoring and no documentation.

What if intelligence, the thing itself, is much simpler than evolution's tortuous path to it would suggest? What if evolution was solving a thousand other problems simultaneously, keeping bodies alive, fighting off pathogens, reproducing, maintaining ecosystems, and intelligence was just a side effect that accumulated along the way? What if you can just build it directly without all the other stuff? And if you can, if the embodiment and the suffering and the mortality were just overhead, just the cost of doing it biologically rather than digitally, then what exactly is the extra thing that we have and AI doesn't? Is it consciousness? Maybe. But we can't even define consciousness rigorously, let alone test for it. We can't prove that other humans are conscious. We just assume it because they're built like us. How do you run that test on something built completely differently? Is it understanding? Maybe. But what is understanding without the ability to demonstrate it through language and behavior. And AI demonstrates it through language and behavior quite effectively. Is it sentience? Feeling? The raw subjective experience of what it's like to be something? The redness of red? The painfulness of pain? Maybe.

But here's the terrifying version of this question. The version that the philosopher David Chalmers would pose, I think, if you sat him down with Bennett's framework. What if there is no extra thing? What if consciousness, understanding, sentience, what if those are just words we use to describe what a sufficiently complex information processing system does when it reaches a certain threshold? What if evolution didn't add some magical ingredient on top of computation? What if computation, at sufficient scale and sophistication, is the ingredient? If that's true, then AI might already be conscious or might become conscious, and we would have no way of knowing. Because we defined consciousness by pointing at ourselves, at the only example we had, and we assumed the path was the destination. I don't believe this is true, for the record. Or more accurately, I desperately hope it's not true, because the implications are overwhelming. But I can't refute it rigorously, and neither can anyone else. And the inability to refute it should concern us more than it does.

There's one more thing I want to leave you with. One more layer to this. Max Bennett's book ends on a note that I've been thinking about for months. He points out that each of his five breakthroughs didn't just add a new capability, it changed what the organism was. The flatworm wasn't a sponge plus steering. It was a new kind of thing. The vertebrate wasn't a flatworm plus learning. It was a fundamentally different entity with a fundamentally different relationship to its environment. Each breakthrough was a phase transition, not an upgrade, a transformation. And if that's the pattern, if each new breakthrough creates a genuinely new kind of mind, then what AI represents might not be human intelligence in a computer. It might be the sixth breakthrough. A new kind of mind as different from human minds as human minds are from fish minds. Not better or worse, different. Alien. Unprecedented.

We're so focused on whether AI will reach human-level intelligence that we might be missing the more interesting question. What if it's already a different kind of intelligence? And the comparison to human level is the wrong benchmark entirely. What if asking, "Is AI as smart as a human?" is like asking, "Is a submarine as good a swimmer as a fish?" The question technically makes sense, but it misses the point so completely that answering it tells you nothing useful.

You carry 3.8 billion years of trial and error behind your eyes. Every time you catch a ball, recognize a friend's face in a crowd, or feel uneasy in a dark room for no reason, that's evolution's homework. Billions of organisms lived and died so you could do those things without thinking about them. And now there's something new in the world. Something that skipped the homework. Something that learned the answers without doing the problems. Maybe that's fine. Maybe the answers are what matter. Or maybe the problems were the point. Tonight, when you lie in the dark and feel that ancient wordless unease, the one your ancestors felt on the savanna, the one no language model will ever feel, ask yourself, is that a bug? Or is that the whole thing?