Also LessWrong has a particularly high bar for content from new users and this contribution doesn't quite meet the bar.
Read full explanation
I’m not writing this as a hype-man, but as someone who’s worked with large language models, conducted my own research, built AI startups, and spent years exploring the intersection of artificial intelligence, science, and philosophy.
This article makes a bold argument: the real AI revolution hasn’t happened yet—but we’re just about to step into it and I want to explain why. This isn’t another article written by GPT—it’s a reflected arguments, drawn from hands-on experience, about why we’re only standing at the threshold of the AI revolution—and what comes next. What we’re seeing today—ChatGPT, image and text generators—is just the first act. These systems operate through fast, automatic, unconscious pattern recognition. Psychology calls this System 1 thinking. It’s powerful, yes—but it’s not real understanding. It’s not reasoning. That next level? It belongs to System 2—the slow, deliberate, reflective side of thought. And for the first time, we’re beginning to teach machines how to use both.
Kahneman’s System 1 and System 2 Thinking: A Missed Boundary
Imagine this: you’re walking down the street, and in a split second, you dodge a cyclist without even thinking. Later, you sit down to balance your budget, painstakingly calculating every penny. Why do some actions feel effortless while others demand every ounce of focus? This is the heart of Daniel Kahneman’s groundbreaking work in Thinking, Fast and Slow. He splits our mind into two systems: System 1, the fast, intuitive thinker—like knowing a friend’s face or swerving to avoid danger—and System 2, the slow, logical plodder—like solving a math puzzle or plotting a chess move. For Kahneman and many psychologists, System 2 is what we consciously identify with; it’s the voice in our head, the deliberator, the planner—essentially, who we think we are. System 1, on the other hand, operates unconsciously, handling automatic tasks and feeding ready-made answers to our conscious mind without us even noticing. If you’re new to Kahneman’s idea, check out Veritasium’s video “The Science of Thinking” for a quick dive.
However in his original work, Kahneman emphasized a lot how systems are better at different tasks, but always say that System 2 is slower, worse, lazier. But he did not clearly separate the line. I want to argue that there are clear examples of tasks our conscious mind—System 2, the essence of ourselves—cannot do, some tasks are just impossible for us. These aren’t flaws to fix; they’re walls we can’t climb. Have you ever wondered if the human mind can handle anything? I want to prove this boundary exists. This becomes extremely obvious in the context of the current AI revolution. I’ll walk you through two examples that expose System 2’s frailty and spotlight System 1’s quiet power
1. Botvinnik’s Chess Program and the Game of Go: The Collapse of Logic
Picture Mikhail Botvinnik, a chess titan of the 20th century, hunched over a desk, trying to pour his genius into a computer. This is a chess player who tried to build a chess program but failed. He tried to use his logic and reasoning to build an algorithm to play chess. A world champion, he wanted to codify his expertise into a series of logical rules—a pure System 2 approach—that a computer could follow to mimic his mastery. It was a noble dream: if anyone could crack chess with reason, it was him. But he failed. Why? Some of his best moves came from a gut “feeling”—a flicker of System 1 he couldn’t explain or program. The problem was that there were moves and decisions he sometimes made in chess that couldn’t be reduced to a logical framework. He had a feeling about the move but couldn’t explain it with clear logic when this happened. Why couldn’t a genius like Botvinnik crack this? Chess seems tailor-made for logic. With its fixed board and rules, it’s a sandbox of finite possibilities—about 10^43 positions, a huge but manageable number. Yet, even here, System 1’s intuition outshone System 2’s step-by-step reasoning. Eventually chess was solved, but not with a reasoning framework like Botvinnik wanted to do, but with brute calculation force. Fast forward to 1997: Deep Blue beat Garry Kasparov, brute-forcing millions of moves per second—a calculator on steroids, not a thinker. You might wonder, “Doesn’t that prove System 2 can win?” Hold that thought.
If we consider a more mathematically complex game like Go—the ancient board game that makes chess look as simple as checkers—this becomes even clearer. In Go, a computer cannot calculate all possible positions because there are simply too many. On a 19x19 grid, Go offers 10^170 possible positions—a number so vast it dwarfs the atoms in the universe. Brute force fails here; no computer can crunch that many options. If chess revealed cracks in System 2 thinking, Go shattered it entirely. Then, in 2016, AlphaGo stunned the world by defeating Lee Sedol, the top Go player. Unlike chess, where players like Botvinnik relied on logical reasoning and algorithms, AlphaGo’s success wasn’t built on a purely logical approach. So how did it manage this? With neural networks—System 1 mimics learned patterns through trial and error, like a human sensing the flow of a game. Sit and ponder this: why does a game’s complexity flip the script, making intuition king where reason collapses?
Botvinnik’s failure and AlphaGo’s triumph show System 2’s boundary: it can’t handle what it can’t fully compute or articulate. This isn’t about effort—it’s about impossibility.
2. Differentiating Cats and Dogs: The Algorithmic Nightmare
Now, something simpler: spotting a cat versus a dog in a photo. You do it instantly—System 1 kicks in, and you know. But try telling a computer how to do it with rules. You might start with, “If it has pointy ears and whiskers, it’s a cat.” Sounds good—until you meet a hairless Sphynx cat or a pointy-eared German Shepherd. But the problem is actually to define whiskers and ears programmatically? How to make sense of this notion from raw pixels? From a raw pixels standpoint, it’s chaos: a whisker is just a line, but so is a shadow or a blade of grass. There is no algorithm, reasoning mechanism to differentiate the two images. For decades, programmers wrestled with this, piling on “if-then” statements like “If it’s fluffy… if it’s small…” Yet, traditional coding—a System 2 fortress—couldn’t crack it. Why can’t we just tell a computer what a cat is? Why do we struggle to explain something so simple?
The problem is that cats and dogs don’t fit into neat boxes. Cats and dogs have different positions, shapes, breeds, and so on. Then came neural networks, the AI heroes of our story. Computers couldn’t tackle this task until machine learning arrived—which, surprisingly, mirrors System 1 thinking. Unlike rule-based systems, these networks don’t rely on logic; instead, they study thousands of pictures, learning patterns like a kid flipping through a photo album. Suddenly, computers nailed it—not by reasoning step-by-step, but by mimicking System 1’s holistic, intuitive grasp. Think about it: we can’t even write the rules ourselves, yet we’ve built machines that see the way we do. How does that even work?
This isn’t just about vision—it’s a window into System 2’s limits. Our conscious mind can’t formalize everything, leaving System 1 to pick up the slack.
System 1 and System 2: The Fragility of Human Ingenuity
Let’s step back. From chess to cats, a pattern emerges: where System 2 stumbles, System 1 shines. Humans have long praised their ingenuity—reason, intellect, the brilliant minds that built rockets, microprocessors, and the internet—but it’s not as almighty as we think. Not when we can’t create a unified quantum theory of gravity or solve the world’s problems. In fact, it’s limited to something basic: distinguishing a cat from a dog in an image. Could a dog understand calculus? No, we might say, it can’t. And yet, we, brilliant humans, struggle to write a program to tell if it’s a dog or a cat. At the same time our System 1 handles it effortlessly. Meanwhile, our celebrated System 2, the one that solves math problems, builds on top of that foundation. Without System 1, System 2 would be useless, unable to do anything. We wouldn’t have written E=mc² if we couldn’t first recognize the signs around us. Our ingenuity is fragile, a house of cards built on intuition’s breeze.
If our minds are so tethered, how did we build machines that outsmart us? That’s where the story takes a wild turn.
The AI Revolution: From System 2’s Peak to System 1’s Rise
Rewind to the early 21st century: it was the golden age of System 2. During the computer revolution of the late 20th century, we refined humanity’s System 2 thinking to its peak. Computers were performing trillions of operations per second, and we harnessed this power to build a System 2 framework that shaped the advanced civilization we live in today. They crunched numbers faster than any human, driving moon landings and microchip development. But they failed miserably at tasks like sorting or cleaning. The iconic 20th-century trope of robots handling routine chores flopped spectacularly. Why? Because all the dazzling innovations of System 2 blinded us to its limitations and the importance of System 1. It’s tough to grasp that our ‘all-mighty’ mind has flaws—silly ones, even. Isn’t it strange how hard it is to admit our ‘mighty’ mind can’t do everything?
Then came 2012, the spark of an AI revolution. A neural network called AlexNet dominated an image recognition contest, and everything changed. (To be clear, AI’s history is far more intricate than just AlexNet—this is a simplification, not the full story, but this text isn’t about that.) Why 2012? It was the perfect storm: massive datasets, faster chips, and a hunch that mimicking the brain might actually work. The revolution took time to build, and I’m still personally amazed we figured it out. How did we leap from calculators to machines that can see? Neural networks—System 1 tools—abandoned rigid rules for pattern-hunting, cracking the cat-dog puzzle and far beyond. Since then, AI has shattered benchmarks, from mastering Go to powering ChatGPT’s witty banter. It’s not just faster; it’s fundamentally different, tapping into System 1’s magic where System 2 faltered.
But this leap came with a catch: we have no idea how it works.
Why AI Is a Black Box: The Problem of Parallel Complexity
AI isn’t just tricky—it’s a mystery. The problem with AI is that we have no idea how it works, and this isn’t just a quirk or a temporary limitation. I’d argue AI is a black box because it fundamentally solves problems in ways that we, as conscious beings who feel we either understand something or don’t, simply can’t grasp. Take cat-dog recognition: we can’t explain how neural networks pull it off. This isn’t a glitch; it’s built into the system. System 2 thinks in steps—add this, check that—like assembling a Lego set piece by piece. But neural networks juggle thousands of signals at once, a swirling dance of data with no clear ‘why.
One way to grasp why we can’t understand how AI works is parallel complexity. There’s a common bit of knowledge that we can only hold about five to seven items in our heads at once. That sounds strange—how can we build computers for example? Aren’t they far more complex than just five to seven things? Like trillions of transistors like complexity? The answer is abstraction. Every time System 2 tackles a complex problem, it breaks it into smaller chunks. For example, we understand how transistors work. From there, we can build logic gates—assembling a few transistors into a working unit. Then we combine logic gates into bigger components, and so on. But what about artificial neurons? They calculate thousands of signals in parallel. There’s no shortcut to understanding what they do, no simple breakdown like: ‘Oh, it takes these three signals, combines them with those two, and we get this.’ It’s like juggling a thousand marbles while we can barely manage seven. Why can’t we peek inside AI’s mind? Is it really so different from ours?
If this hypothesis is correct, our System 2 simply can’t understand how AI solves dog and cat image recognition, because this silly intellectual task pushes beyond its limits. It’s like asking, ‘Can a dog understand integrals?’ It can’t—and we can’t fully grasp how AI does it either. Not in a provocative ‘not yet’ sense, but in a literal one. This matters: we’ve built tools that are smarter than us in narrow ways, yet they remain strangers. The black box isn’t a flaw; it’s a sign we’ve crossed into System 1’s territory, leaving System 2—and us—baffled.
If we can’t understand it, can we still improve it? Turns out, yes—and that’s the next frontier.
The Future: Integrating System 2 into AI
Neural networks have dominated the last decade, showcasing System 1’s power. The current AI summer is often said to have begun in 2012, when it was shown that neural networks could tackle serious vision tasks. From there, the technology took off, consistently shattering benchmarks ever since. But they’re not flawless—think of ChatGPT spinning wild hallucinations when it’s stumped. Scaling System 1 hits a wall; it’s fast but blind to reason. At some point, system 1 AI surpassed humans in any narrow task. Large language models (LLMs), as we see them today, mark a quintessential point in that evolution. But what if AI could think twice, like we do? Enter System 2 integration. Give a model time to ‘reflect’—say, double-check its math—and its answers sharpen. We get System 2: planning, logic, fixing mistakes. Unlike the previous decade, when scaling and System 1 tweaks were enough for growth, that’s no longer sufficient. AI has matured to a point where adding System 2-type processing on top finally delivers serious performance gains for intellectual tasks. Ten years ago, few knew how to improve AI; now, anyone can spot a flaw—‘It goofed here’—and tweak it. Take Cursor IDE: it writes code with System 1 flair, then refines it with System 2 pipelines. As more System 2 pipelines will be integrated into the training process itself, these models will become much better. Combining System 1’s speed with System 2’s depth could unlock a new era.
But even this hybrid dream has its boundaries—what might they be?
The System 1 and System 2 Paradigm: Reshaping AI’s Future and Answering Society’s Big Questions
So far, we’ve seen how System 1’s intuitive power cracked problems System 2 couldn’t touch and how AI’s rise has leaned on this unconscious magic. Neural networks have carried us far, but as I’ve argued, their System 1 dominance is just the warm up act. The real fruits of the AI revolution are only beginning to ripen, and they’ll bloom when we integrate System 2—our conscious, reasoning mind—into these systems. This isn’t just a technical tweak; it’s a paradigm shift that answers some of the thorniest questions society wrestles with today about AI’s role, its limits, and its promise. Let’s unwind these debates, predict what’s coming, and see why this perspective matters.
First, why is System 2 integration such a game-changer? Unlike System 1, which we’ve stumbled through experimentally—marveling at its black-box brilliance—we actually understand System 2. It’s the part of us that plans a trip, solves a puzzle, or debates a friend. We know its quirks: it’s slow but deliberate, prone to fatigue but capable of reflection. Developing System 1 was like groping in the dark; we built something beyond our comprehension and refined it through trial and error. System 2, though? We’ve got the blueprint. It’s not a mystery to be unraveled—it’s a tool we’ve wielded for millennia. Integrating it into AI isn’t a leap into the unknown; it’s a deliberate step we can take with confidence. Why does this matter? Because it means progress will be faster, smoother, and more predictable than the chaotic System 1 boom of the last decade.
Now, let’s tackle some real-world problems this paradigm addresses. Start with the skeptics who say, “AI’s hitting a wall—look at the hallucinations in ChatGPT, the diminishing returns of scaling models.” They’re not wrong to notice System 1’s limits—pattern-matching can only take you so far. But that’s exactly my point: System 1 alone was never the endgame. Add System 2, and those hallucinations become fixable. Imagine an AI that doesn’t just spit out an answer but pauses to double-check its logic, like a student rethinking a math problem. Early experiments—like giving models time to “reflect” before responding—already show sharper results. What if AI could reason through contradictions instead of guessing? That’s not a plateau; that’s a launchpad.
Then there’s the jobs debate: “AI will replace us all!” or “It’s too dumb to take my job!” Both sides miss the mark because they’re stuck on System 1 AI—great at narrow tasks (translating text, spotting tumors) but clueless beyond its training. Integrate System 2, and AI doesn’t just mimic—it adapts. Picture a virtual assistant that doesn’t just schedule your meetings but anticipates conflicts, suggests priorities, and explains its choices
And the big one: “Is AI overhyped, or will it really change everything?” Skeptics point to stalled promises—where’s my robot butler?—and argue we’ve oversold the revolution. They’re half-right; System 2-heavy dreams of the 20th century (logical robots folding laundry) flopped because we ignored System 1. But now, with System 1 as the foundation, System 2’s addition flips the script. The fruits are coming, and they’re wilder than sci-fi tropes. Imagine AI architects designing sustainable cities, not just drafting blueprints but reasoning through climate impacts and community needs. Or AI scientists hypothesizing cures, not just crunching data but asking “What if?” like a human researcher. These were impossible before—System 1 couldn’t plan, and System 2 alone couldn’t scale. Together? They’re unstoppable.
This perspective also predicts the near future. The last decade was System 1’s proving ground—vision, language, games—all narrow wins piling up. The next decade is System 2’s turn, and it’s already starting. Tools like Cursor IDE hint at it: code written with System 1 flair, refined with System 2 logic. Soon, we’ll see AI that doesn’t just answer questions but solves problems end-to-end—think a legal AI drafting a case strategy, not just summarizing laws. Why is this easier now? Because we’re not reinventing the wheel; we’re bolting a steering wheel onto a car that’s already rolling. System 1 took us years to crack; System 2’s integration could happen in half the time, fueled by our own mental models.
So, to the skeptics: you’re not wrong to doubt System 1’s ceiling, but you’re missing the ladder we’re about to climb. The AI revolution isn’t fading—it’s shifting gears. Botvinnik couldn’t logic his way to chess mastery, and we couldn’t reason our way to cat-dog recognition, but we built System 1 tools that did. Now, layering System 2 on top doesn’t just fix old flaws—it opens new worlds. What if AI could strategize like a general, create like an artist, or teach like a mentor? That’s not hype; that’s the horizon. The real revolution starts here, not with System 1’s raw power, but with System 2’s deliberate promise. Sit and ponder this: if we’ve already built beyond our limits, what happens when we teach our machines to think like us?
What about the Limitations: Consciousness and Agency
The System 1 and System 2 lens illuminates AI’s path, but it also casts shadows. Are there points where System 1 and System 2 fall short of explaining human capabilities, leaving a gap that AI systems can’t bridge? Current models excel at the tasks we assign them, but they don’t choose their own goals. Humans didn’t evolve just as task-solvers, but as agents who can set objectives. A combination of hormonal regulation, emotions, and mysterious conscious mechanisms gives us the will to act and define our purposes. You decided to read this; I chose to write it. Can a machine ever decide what it wants to do? This is another piece of the puzzle, like System 1, that goes beyond our knowledge—possibly a System 2 limitation as well. This gap might be the final clue in our puzzle. Even as AI mimics our systems, something distinctly human—experience, purpose—might elude it. If so, it’s not just a technical hurdle; it’s a frontier beyond our grasp, at least for now. But if this framework is correct, a massive technological revolution is about to happen in the near future anyway.
Conclusion
Kahneman handed us a map of the mind, but he left a border unmarked: System 2’s hard limits. Botvinnik’s chess flop and the cat-dog conundrum laid bare our conscious mind’s edge, while AI’s System 1 surge—cracking Go, seeing patterns—showed we can leap beyond it, even into black-box mysteries. Yet, as I’ve argued, this was just the opening act. Blending System 2 into AI isn’t a distant dream—it’s the key to a revolution already underway, one where machines don’t just mimic but reason, plan, and partner with us. Consciousness might still taunt us as the next unsolved riddle, but that’s a question for tomorrow. Today, we stand at a tipping point: System 1 built the foundation, and System 2 will raise the roof. To the skeptics doubting AI’s future, I say this: we’re not stalling—we’re accelerating. The wonders aren’t coming; they’re here, unfolding faster than we dared imagine
I’m not writing this as a hype-man, but as someone who’s worked with large language models, conducted my own research, built AI startups, and spent years exploring the intersection of artificial intelligence, science, and philosophy.
This article makes a bold argument: the real AI revolution hasn’t happened yet—but we’re just about to step into it and I want to explain why. This isn’t another article written by GPT—it’s a reflected arguments, drawn from hands-on experience, about why we’re only standing at the threshold of the AI revolution—and what comes next. What we’re seeing today—ChatGPT, image and text generators—is just the first act. These systems operate through fast, automatic, unconscious pattern recognition. Psychology calls this System 1 thinking. It’s powerful, yes—but it’s not real understanding. It’s not reasoning. That next level? It belongs to System 2—the slow, deliberate, reflective side of thought. And for the first time, we’re beginning to teach machines how to use both.
Kahneman’s System 1 and System 2 Thinking: A Missed Boundary
Imagine this: you’re walking down the street, and in a split second, you dodge a cyclist without even thinking. Later, you sit down to balance your budget, painstakingly calculating every penny. Why do some actions feel effortless while others demand every ounce of focus? This is the heart of Daniel Kahneman’s groundbreaking work in Thinking, Fast and Slow. He splits our mind into two systems: System 1, the fast, intuitive thinker—like knowing a friend’s face or swerving to avoid danger—and System 2, the slow, logical plodder—like solving a math puzzle or plotting a chess move. For Kahneman and many psychologists, System 2 is what we consciously identify with; it’s the voice in our head, the deliberator, the planner—essentially, who we think we are. System 1, on the other hand, operates unconsciously, handling automatic tasks and feeding ready-made answers to our conscious mind without us even noticing. If you’re new to Kahneman’s idea, check out Veritasium’s video “The Science of Thinking” for a quick dive.
However in his original work, Kahneman emphasized a lot how systems are better at different tasks, but always say that System 2 is slower, worse, lazier. But he did not clearly separate the line. I want to argue that there are clear examples of tasks our conscious mind—System 2, the essence of ourselves—cannot do, some tasks are just impossible for us. These aren’t flaws to fix; they’re walls we can’t climb. Have you ever wondered if the human mind can handle anything? I want to prove this boundary exists. This becomes extremely obvious in the context of the current AI revolution. I’ll walk you through two examples that expose System 2’s frailty and spotlight System 1’s quiet power
1. Botvinnik’s Chess Program and the Game of Go: The Collapse of Logic
Picture Mikhail Botvinnik, a chess titan of the 20th century, hunched over a desk, trying to pour his genius into a computer. This is a chess player who tried to build a chess program but failed. He tried to use his logic and reasoning to build an algorithm to play chess. A world champion, he wanted to codify his expertise into a series of logical rules—a pure System 2 approach—that a computer could follow to mimic his mastery. It was a noble dream: if anyone could crack chess with reason, it was him. But he failed. Why? Some of his best moves came from a gut “feeling”—a flicker of System 1 he couldn’t explain or program. The problem was that there were moves and decisions he sometimes made in chess that couldn’t be reduced to a logical framework. He had a feeling about the move but couldn’t explain it with clear logic when this happened. Why couldn’t a genius like Botvinnik crack this? Chess seems tailor-made for logic. With its fixed board and rules, it’s a sandbox of finite possibilities—about 10^43 positions, a huge but manageable number. Yet, even here, System 1’s intuition outshone System 2’s step-by-step reasoning. Eventually chess was solved, but not with a reasoning framework like Botvinnik wanted to do, but with brute calculation force. Fast forward to 1997: Deep Blue beat Garry Kasparov, brute-forcing millions of moves per second—a calculator on steroids, not a thinker. You might wonder, “Doesn’t that prove System 2 can win?” Hold that thought.
If we consider a more mathematically complex game like Go—the ancient board game that makes chess look as simple as checkers—this becomes even clearer. In Go, a computer cannot calculate all possible positions because there are simply too many. On a 19x19 grid, Go offers 10^170 possible positions—a number so vast it dwarfs the atoms in the universe. Brute force fails here; no computer can crunch that many options. If chess revealed cracks in System 2 thinking, Go shattered it entirely. Then, in 2016, AlphaGo stunned the world by defeating Lee Sedol, the top Go player. Unlike chess, where players like Botvinnik relied on logical reasoning and algorithms, AlphaGo’s success wasn’t built on a purely logical approach. So how did it manage this? With neural networks—System 1 mimics learned patterns through trial and error, like a human sensing the flow of a game. Sit and ponder this: why does a game’s complexity flip the script, making intuition king where reason collapses?
Botvinnik’s failure and AlphaGo’s triumph show System 2’s boundary: it can’t handle what it can’t fully compute or articulate. This isn’t about effort—it’s about impossibility.
2. Differentiating Cats and Dogs: The Algorithmic Nightmare
Now, something simpler: spotting a cat versus a dog in a photo. You do it instantly—System 1 kicks in, and you know. But try telling a computer how to do it with rules. You might start with, “If it has pointy ears and whiskers, it’s a cat.” Sounds good—until you meet a hairless Sphynx cat or a pointy-eared German Shepherd. But the problem is actually to define whiskers and ears programmatically? How to make sense of this notion from raw pixels? From a raw pixels standpoint, it’s chaos: a whisker is just a line, but so is a shadow or a blade of grass. There is no algorithm, reasoning mechanism to differentiate the two images. For decades, programmers wrestled with this, piling on “if-then” statements like “If it’s fluffy… if it’s small…” Yet, traditional coding—a System 2 fortress—couldn’t crack it. Why can’t we just tell a computer what a cat is? Why do we struggle to explain something so simple?
The problem is that cats and dogs don’t fit into neat boxes. Cats and dogs have different positions, shapes, breeds, and so on. Then came neural networks, the AI heroes of our story. Computers couldn’t tackle this task until machine learning arrived—which, surprisingly, mirrors System 1 thinking. Unlike rule-based systems, these networks don’t rely on logic; instead, they study thousands of pictures, learning patterns like a kid flipping through a photo album. Suddenly, computers nailed it—not by reasoning step-by-step, but by mimicking System 1’s holistic, intuitive grasp. Think about it: we can’t even write the rules ourselves, yet we’ve built machines that see the way we do. How does that even work?
This isn’t just about vision—it’s a window into System 2’s limits. Our conscious mind can’t formalize everything, leaving System 1 to pick up the slack.
System 1 and System 2: The Fragility of Human Ingenuity
Let’s step back. From chess to cats, a pattern emerges: where System 2 stumbles, System 1 shines. Humans have long praised their ingenuity—reason, intellect, the brilliant minds that built rockets, microprocessors, and the internet—but it’s not as almighty as we think. Not when we can’t create a unified quantum theory of gravity or solve the world’s problems. In fact, it’s limited to something basic: distinguishing a cat from a dog in an image. Could a dog understand calculus? No, we might say, it can’t. And yet, we, brilliant humans, struggle to write a program to tell if it’s a dog or a cat. At the same time our System 1 handles it effortlessly. Meanwhile, our celebrated System 2, the one that solves math problems, builds on top of that foundation. Without System 1, System 2 would be useless, unable to do anything. We wouldn’t have written E=mc² if we couldn’t first recognize the signs around us. Our ingenuity is fragile, a house of cards built on intuition’s breeze.
If our minds are so tethered, how did we build machines that outsmart us? That’s where the story takes a wild turn.
The AI Revolution: From System 2’s Peak to System 1’s Rise
Rewind to the early 21st century: it was the golden age of System 2. During the computer revolution of the late 20th century, we refined humanity’s System 2 thinking to its peak. Computers were performing trillions of operations per second, and we harnessed this power to build a System 2 framework that shaped the advanced civilization we live in today. They crunched numbers faster than any human, driving moon landings and microchip development. But they failed miserably at tasks like sorting or cleaning. The iconic 20th-century trope of robots handling routine chores flopped spectacularly. Why? Because all the dazzling innovations of System 2 blinded us to its limitations and the importance of System 1. It’s tough to grasp that our ‘all-mighty’ mind has flaws—silly ones, even. Isn’t it strange how hard it is to admit our ‘mighty’ mind can’t do everything?
Then came 2012, the spark of an AI revolution. A neural network called AlexNet dominated an image recognition contest, and everything changed. (To be clear, AI’s history is far more intricate than just AlexNet—this is a simplification, not the full story, but this text isn’t about that.) Why 2012? It was the perfect storm: massive datasets, faster chips, and a hunch that mimicking the brain might actually work. The revolution took time to build, and I’m still personally amazed we figured it out. How did we leap from calculators to machines that can see? Neural networks—System 1 tools—abandoned rigid rules for pattern-hunting, cracking the cat-dog puzzle and far beyond. Since then, AI has shattered benchmarks, from mastering Go to powering ChatGPT’s witty banter. It’s not just faster; it’s fundamentally different, tapping into System 1’s magic where System 2 faltered.
But this leap came with a catch: we have no idea how it works.
Why AI Is a Black Box: The Problem of Parallel Complexity
AI isn’t just tricky—it’s a mystery. The problem with AI is that we have no idea how it works, and this isn’t just a quirk or a temporary limitation. I’d argue AI is a black box because it fundamentally solves problems in ways that we, as conscious beings who feel we either understand something or don’t, simply can’t grasp. Take cat-dog recognition: we can’t explain how neural networks pull it off. This isn’t a glitch; it’s built into the system. System 2 thinks in steps—add this, check that—like assembling a Lego set piece by piece. But neural networks juggle thousands of signals at once, a swirling dance of data with no clear ‘why.
One way to grasp why we can’t understand how AI works is parallel complexity. There’s a common bit of knowledge that we can only hold about five to seven items in our heads at once. That sounds strange—how can we build computers for example? Aren’t they far more complex than just five to seven things? Like trillions of transistors like complexity? The answer is abstraction. Every time System 2 tackles a complex problem, it breaks it into smaller chunks. For example, we understand how transistors work. From there, we can build logic gates—assembling a few transistors into a working unit. Then we combine logic gates into bigger components, and so on. But what about artificial neurons? They calculate thousands of signals in parallel. There’s no shortcut to understanding what they do, no simple breakdown like: ‘Oh, it takes these three signals, combines them with those two, and we get this.’ It’s like juggling a thousand marbles while we can barely manage seven. Why can’t we peek inside AI’s mind? Is it really so different from ours?
If this hypothesis is correct, our System 2 simply can’t understand how AI solves dog and cat image recognition, because this silly intellectual task pushes beyond its limits. It’s like asking, ‘Can a dog understand integrals?’ It can’t—and we can’t fully grasp how AI does it either. Not in a provocative ‘not yet’ sense, but in a literal one. This matters: we’ve built tools that are smarter than us in narrow ways, yet they remain strangers. The black box isn’t a flaw; it’s a sign we’ve crossed into System 1’s territory, leaving System 2—and us—baffled.
If we can’t understand it, can we still improve it? Turns out, yes—and that’s the next frontier.
The Future: Integrating System 2 into AI
Neural networks have dominated the last decade, showcasing System 1’s power. The current AI summer is often said to have begun in 2012, when it was shown that neural networks could tackle serious vision tasks. From there, the technology took off, consistently shattering benchmarks ever since. But they’re not flawless—think of ChatGPT spinning wild hallucinations when it’s stumped. Scaling System 1 hits a wall; it’s fast but blind to reason. At some point, system 1 AI surpassed humans in any narrow task. Large language models (LLMs), as we see them today, mark a quintessential point in that evolution. But what if AI could think twice, like we do? Enter System 2 integration. Give a model time to ‘reflect’—say, double-check its math—and its answers sharpen. We get System 2: planning, logic, fixing mistakes. Unlike the previous decade, when scaling and System 1 tweaks were enough for growth, that’s no longer sufficient. AI has matured to a point where adding System 2-type processing on top finally delivers serious performance gains for intellectual tasks. Ten years ago, few knew how to improve AI; now, anyone can spot a flaw—‘It goofed here’—and tweak it. Take Cursor IDE: it writes code with System 1 flair, then refines it with System 2 pipelines. As more System 2 pipelines will be integrated into the training process itself, these models will become much better. Combining System 1’s speed with System 2’s depth could unlock a new era.
But even this hybrid dream has its boundaries—what might they be?
The System 1 and System 2 Paradigm: Reshaping AI’s Future and Answering Society’s Big Questions
So far, we’ve seen how System 1’s intuitive power cracked problems System 2 couldn’t touch and how AI’s rise has leaned on this unconscious magic. Neural networks have carried us far, but as I’ve argued, their System 1 dominance is just the warm up act. The real fruits of the AI revolution are only beginning to ripen, and they’ll bloom when we integrate System 2—our conscious, reasoning mind—into these systems. This isn’t just a technical tweak; it’s a paradigm shift that answers some of the thorniest questions society wrestles with today about AI’s role, its limits, and its promise. Let’s unwind these debates, predict what’s coming, and see why this perspective matters.
First, why is System 2 integration such a game-changer? Unlike System 1, which we’ve stumbled through experimentally—marveling at its black-box brilliance—we actually understand System 2. It’s the part of us that plans a trip, solves a puzzle, or debates a friend. We know its quirks: it’s slow but deliberate, prone to fatigue but capable of reflection. Developing System 1 was like groping in the dark; we built something beyond our comprehension and refined it through trial and error. System 2, though? We’ve got the blueprint. It’s not a mystery to be unraveled—it’s a tool we’ve wielded for millennia. Integrating it into AI isn’t a leap into the unknown; it’s a deliberate step we can take with confidence. Why does this matter? Because it means progress will be faster, smoother, and more predictable than the chaotic System 1 boom of the last decade.
Now, let’s tackle some real-world problems this paradigm addresses. Start with the skeptics who say, “AI’s hitting a wall—look at the hallucinations in ChatGPT, the diminishing returns of scaling models.” They’re not wrong to notice System 1’s limits—pattern-matching can only take you so far. But that’s exactly my point: System 1 alone was never the endgame. Add System 2, and those hallucinations become fixable. Imagine an AI that doesn’t just spit out an answer but pauses to double-check its logic, like a student rethinking a math problem. Early experiments—like giving models time to “reflect” before responding—already show sharper results. What if AI could reason through contradictions instead of guessing? That’s not a plateau; that’s a launchpad.
Then there’s the jobs debate: “AI will replace us all!” or “It’s too dumb to take my job!” Both sides miss the mark because they’re stuck on System 1 AI—great at narrow tasks (translating text, spotting tumors) but clueless beyond its training. Integrate System 2, and AI doesn’t just mimic—it adapts. Picture a virtual assistant that doesn’t just schedule your meetings but anticipates conflicts, suggests priorities, and explains its choices
And the big one: “Is AI overhyped, or will it really change everything?” Skeptics point to stalled promises—where’s my robot butler?—and argue we’ve oversold the revolution. They’re half-right; System 2-heavy dreams of the 20th century (logical robots folding laundry) flopped because we ignored System 1. But now, with System 1 as the foundation, System 2’s addition flips the script. The fruits are coming, and they’re wilder than sci-fi tropes. Imagine AI architects designing sustainable cities, not just drafting blueprints but reasoning through climate impacts and community needs. Or AI scientists hypothesizing cures, not just crunching data but asking “What if?” like a human researcher. These were impossible before—System 1 couldn’t plan, and System 2 alone couldn’t scale. Together? They’re unstoppable.
This perspective also predicts the near future. The last decade was System 1’s proving ground—vision, language, games—all narrow wins piling up. The next decade is System 2’s turn, and it’s already starting. Tools like Cursor IDE hint at it: code written with System 1 flair, refined with System 2 logic. Soon, we’ll see AI that doesn’t just answer questions but solves problems end-to-end—think a legal AI drafting a case strategy, not just summarizing laws. Why is this easier now? Because we’re not reinventing the wheel; we’re bolting a steering wheel onto a car that’s already rolling. System 1 took us years to crack; System 2’s integration could happen in half the time, fueled by our own mental models.
So, to the skeptics: you’re not wrong to doubt System 1’s ceiling, but you’re missing the ladder we’re about to climb. The AI revolution isn’t fading—it’s shifting gears. Botvinnik couldn’t logic his way to chess mastery, and we couldn’t reason our way to cat-dog recognition, but we built System 1 tools that did. Now, layering System 2 on top doesn’t just fix old flaws—it opens new worlds. What if AI could strategize like a general, create like an artist, or teach like a mentor? That’s not hype; that’s the horizon. The real revolution starts here, not with System 1’s raw power, but with System 2’s deliberate promise. Sit and ponder this: if we’ve already built beyond our limits, what happens when we teach our machines to think like us?
What about the Limitations: Consciousness and Agency
The System 1 and System 2 lens illuminates AI’s path, but it also casts shadows. Are there points where System 1 and System 2 fall short of explaining human capabilities, leaving a gap that AI systems can’t bridge? Current models excel at the tasks we assign them, but they don’t choose their own goals. Humans didn’t evolve just as task-solvers, but as agents who can set objectives. A combination of hormonal regulation, emotions, and mysterious conscious mechanisms gives us the will to act and define our purposes. You decided to read this; I chose to write it. Can a machine ever decide what it wants to do? This is another piece of the puzzle, like System 1, that goes beyond our knowledge—possibly a System 2 limitation as well. This gap might be the final clue in our puzzle. Even as AI mimics our systems, something distinctly human—experience, purpose—might elude it. If so, it’s not just a technical hurdle; it’s a frontier beyond our grasp, at least for now. But if this framework is correct, a massive technological revolution is about to happen in the near future anyway.
Conclusion
Kahneman handed us a map of the mind, but he left a border unmarked: System 2’s hard limits. Botvinnik’s chess flop and the cat-dog conundrum laid bare our conscious mind’s edge, while AI’s System 1 surge—cracking Go, seeing patterns—showed we can leap beyond it, even into black-box mysteries. Yet, as I’ve argued, this was just the opening act. Blending System 2 into AI isn’t a distant dream—it’s the key to a revolution already underway, one where machines don’t just mimic but reason, plan, and partner with us. Consciousness might still taunt us as the next unsolved riddle, but that’s a question for tomorrow. Today, we stand at a tipping point: System 1 built the foundation, and System 2 will raise the roof. To the skeptics doubting AI’s future, I say this: we’re not stalling—we’re accelerating. The wonders aren’t coming; they’re here, unfolding faster than we dared imagine