Go is a fascinating game. It has only a handful of rules, yet it has obsessed some of the brightest minds in the world for millennia. The board is a 19x19 grid. You place stones to surround territory. Simple enough that a child can learn it in an afternoon, yet complex enough that even the best players dedicate their entire lives to mastery.
This was supposed to be the game that computers couldn't crack. When Deep Blue defeated Garry Kasparov in 1997, computer scientists estimated Go was at least three decades away. The branching factor is too large. The positions are too subtle. You can't brute-force your way to victory when the search space is larger than the number of atoms in the universe.
So it came as a genuine shock when AlphaGo defeated Lee Sedol, the world's top ranked player at the time, in 2016. Not just that it won, but how it won. The game changed that day, and it hasn't changed back.
Joseki: The Accumulated Wisdom of Generations
To understand what happened, you need to understand Joseki.
In the opening phase of Go, players fight for position in the corners. These corner skirmishes follow patterns that have been refined over centuries. A Joseki is a sequence of moves that, when played correctly by both sides, results in a roughly equal position. Neither player gains an advantage, but neither falls behind. They're the Go equivalent of chess openings. Except where chess openings number in the dozens, Joseki number in the thousands.
Professionals memorize hundreds of Joseki. They know the variations, the timing, when to deviate. This isn't lazy pattern-matching; it's the accumulated wisdom of the game. Masters spent their lives discovering these sequences. Students spend years internalizing them. The Joseki are Go knowledge.
For thousands of years, Joseki evolved slowly. The only significant shift came in the 20th century, when the rules changed to give the second player a slight point advantage. Players realized the first player needed to be more aggressive to compensate. Some Joseki fell out of favor. New ones emerged. But the fundamental idea held: these patterns represent optimal play, discovered through human intuition and tested through centuries of competition.
Then AlphaGo arrived.
The Rejection
AlphaGo didn't just beat Lee Sedol. It played moves that no professional would consider. Moves that violated established Joseki. Moves that looked, to human eyes, like mistakes.
In the second game, AlphaGo played Move 37: a shoulder hit on the third line, in an area of the board that seemed almost irrelevant to the main fight. Commentators were confused. This wasn't how you played Go. It wasn't wrong, it wasn't illegal, but it wasn't correct either. It wasn't something you would catch any professional playing.
The move turned out to be brilliant. It set up pressure that paid off fifty moves later. Lee Sedol himself called it "a beautiful move." But it wasn't a move any human had discovered.
Move 37 was even more striking coming from Lee Sedol's opponent. Lee Sedol was famous for a game called "the ladder game": a match where he played a sequence so complex and counterintuitive that it seemed to violate basic Go principles, yet worked through sheer calculation and reading ability. Move 37 was AlphaGo's ladder game moment: a move that required seeing further than humans could see, played against the very player known for seeing further than anyone else.
More disturbing was what AlphaGo didn't play. It refused certain Joseki entirely. Patterns that professionals considered balanced, even favorable, AlphaGo evaluated as losing propositions. It prioritized other sequences, quickly dubbed "AI Joseki" by the players.
The 3-3 invasion is the clearest example. Invading the opponent's corner at the 3-3 point had always been considered small, safe, and slightly suboptimal. It secured territory, yes, but it gave the opponent strong influence toward the center. Professionals avoided it in serious games. It was something you played when you were behind and desperate.
AlphaGo played it constantly. Not as a desperation move, but as a standard opening. And when professionals started testing it, they discovered something uncomfortable: it worked. The influence it gave away wasn't worth as much as everyone thought. The territory it secured was more valuable. Thousands of years of wisdom was, in specific cases, simply wrong.
Territory vs. Influence
This is the core insight, and it's worth dwelling on.
Territory in Go is concrete. You can count it. It's yours, guaranteed, as long as you can defend it. Influence is potential. It's power radiating toward empty areas of the board. It's vague, artistic, harder to teach. "This stone has good influence" was the kind of thing a master might say to a student, and the student would nod and try to absorb the intuition through osmosis.
Humans constantly miscalculated influence because we couldn't calculate the follow-up. We couldn't trace the ripple effects of a strong position tens of moves ahead with any precision. So we treated influence as soft judgment, something you developed through experience and intuition.
AlphaGo showed that influence could be calculated. The vague artistic sense had underlying structure. The AI could see, actually see, in a computable way, how a stone's power propagated across the board. It could weigh territory against influence with quantitative precision.
This flipped the game. Openings changed. Mid-game aggression shifted. Even life-and-death reading, the tactical heart of Go, was affected, because the AI found sequences humans had dismissed as too risky, and proved they worked.
The Training Asymmetry
The change didn't happen evenly. Some players adapted faster than others. Some resisted longer. The key variable was access.
The Chinese national team embraced AI training early. They reviewed AI move suggestions obsessively. They trained against superhuman opponents who never tired, never made simple mistakes, never got tilted. They developed an intuition for what the AI would do in positions that had never occurred in human games.
Older players, even some former champions, struggled. The patterns they'd spent decades mastering were suddenly suspect. The judgment they'd developed through thousands of games was being second-guessed by something that didn't play the way humans played.
Lee Chang-ho, who dominated the top of the charts for decades and was known as "Stone Buddha" for his calm, classical style, captured the frustration directly: "I do not enjoy playing games with AI. It keeps jumping around the board." The AI's move selection felt erratic, chaotic, unreadable. But it was also winning.
This created what we might call a training asymmetry. The players who trained with AI didn't just memorize moves. They were open to developing intuition in unconventional ways, training against superhuman opponents who exposed them to positions no human had explored. They learned to feel their way through chaotic board states that earlier generations would have dismissed. The Chinese national team embraced this early, reviewing AI suggestions obsessively, building new intuitions for territory vs. influence through exposure rather than traditional study.
The results were immediate and strange. Top players in Go used to stay dominant for decades as mastery was a lifetime achievement. After AlphaGo, the top ranks started changing frequently. Players who trained with AI could beat players who hadn't, not because they memorized patterns without understanding, but because they had developed new intuitions that the traditional training couldn't replicate.
No amateur magically beat a professional. The fundamentals still mattered: reading ability, tactical sharpness, psychological resilience. But professionals who used AI started beating other professionals more consistently. Pattern recall became the differentiator.
The Annoying 3-3 Invasion
Here's where it gets uncomfortable.
Go tutors started reporting a pattern. Their students, amateurs, kyu players, people still learning fundamentals, would play the 3-3 invasion and other AI tactics in games. They'd win. But they didn't actually understand why they were winning. They couldn't read the follow-up sequences. They couldn't evaluate when the invasion was appropriate versus when it was a mistake. They just knew the move was "good now" because the AI played it.
The tutors found this maddening. Their students were winning games they shouldn't win, not because they'd improved at Go, but because they'd memorized a pattern that worked in contexts they couldn't actually assess. The students skipped the hard work of understanding influence, of learning to calculate follow-ups, of developing the judgment that makes a 3-3 invasion meaningful rather than just a trick.
This isn't a criticism of the students. The professionals who adopted AI had something the students lacked: ten thousand games of foundation to contextualize the new patterns. They were open to training their intuition in unconventional ways, letting the AI expose them to positions that redefined their understanding of territory vs. influence. They didn't memorize patterns blindly; they built new intuitions on top of deep expertise.
The students had a hundred games and a cheat code. They were skipping the hard work of developing the underlying judgment that makes AI patterns meaningful. What looks like "knowing the right move" in a professional is actually deep intuition. In an amateur who has skipped the basics, it's just pattern matching without comprehension.
The AI Joseki spread anyway. Commentators started teaching AI insights on streams and videos. "Play this, not that." The patterns propagated faster than the understanding. And players who learned the patterns without the foundation started beating players who had the foundation but hadn't memorized the patterns yet.
The Synthesis Problem
Not everyone wanted to adapt. Some players resisted. They found the AI style ugly, mechanical, inhuman. They insisted on playing "real Go," the game as it had been played for millennia.
Here's the problem: even the resisters ended up playing AI Go.
If your opponent opens with a 3-3 invasion, you have to respond. If they follow AI Joseki, you have to know the counters or you lose. Even if you never touched a training bot, even if you refused to review AI games, your opponents didn't. Their style became your reality. The synthesis problem meant that traditionalists absorbed AI insights secondhand, filtered through their opponents' play, without the training that would have made those insights intuitive.
There's no going back. Even if we lost access to all AI training tools tomorrow, if every Leela Zero server went offline, every KataGo weight was deleted, the game has changed. The knowledge is out there. It's in the opening books, in the commentary, in what strong players teach their students. The Joseki that AlphaGo rejected are gone forever. The ones it favored are now standard.
Lee Sedol understood this. After losing to AlphaGo, he won one game, the famous Game 4, with Move 78, the "God Move" that AlphaGo hadn't predicted. It was a glorious moment. It was also temporary. AlphaGo kept improving. The gap widened. In 2019, Lee Sedol retired, saying that AI had made the game meaningless. The thing he had dedicated his life to mastering had become, in his view, unbeatable by humans.
But here's the thing: Go is still played. The game didn't end. It just became different.
Software Joseki
Let's talk about code.
Software engineering has its own Joseki. Design patterns. Folder structures. How you name your functions. Where documentation lives. How you structure a code review. What makes a commit "too big." We've developed these patterns through decades of collective experience, and we teach them to newcomers as accumulated wisdom.
Some of these are enforced by compilers. Most aren't. They're conventions, preferences, best practices. They're the software equivalent of Joseki, sequences that, when followed, lead to acceptable outcomes. Not optimal, necessarily. Not the only way. But safe. Tested. The way things are done.
AI is starting to disrupt them.
Take Tailwind CSS. The utility-first approach: putting styles directly in your HTML, repeating classes instead of abstracting them, was controversial when it emerged. It violated DRY (Don't Repeat Yourself). It made templates verbose. It looked ugly.
But it has a property that's extremely valuable for AI: locality of behavior. Everything you need to know about an element's appearance is right there in the element's markup. You don't need to parse external stylesheets. You don't need to understand a naming convention. You don't need to trace a class name back to its definition. The AI can read, modify, and generate styles without understanding your design system's abstraction layers.
Is Tailwind objectively better? That's not the point. The point is that AI finds it easier to work with. And as more code is written with AI assistance, patterns that AI handles well will spread faster than patterns that require human abstraction skills.
This is the 3-3 invasion of software. Something that looks suboptimal by traditional metrics, verbose, repetitive, not following best practices, but turns out to have hidden value in the new environment.
Territory and Influence, Revisited
In software, territory is what ships. It's features, bug fixes, revenue. It's concrete and countable. Your sprint velocity. Your commit count. Your lines of code.
Influence is harder to measure. It's architectural leverage. It's the ease of future changes. It's developer experience. It's the thing that makes making money easier, even if it doesn't make money directly.
Humans have miscalculated or undervalued influence for the same reason Go players did: we couldn't calculate the follow-up. We couldn't trace how a clean abstraction today affects feature velocity six months from now. We couldn't quantify technical debt. So we treated it as soft judgment, something you developed through experience. .AI is starting to show structure in these soft judgments
This will flip priorities. The patterns that maximize immediate territory at the expense of influence will look different when AI can actually calculate the trade-off. Our Joseki will change.
The New Stone Buddhas
There will be resistance. There already is.
"I don't enjoy coding with AI. It jumps around too much. It doesn't follow the structure. It writes code that works but doesn't feel right."
This is the Stone Buddha complaint, and it's valid. Working with AI feels different. The code it generates often violates conventions we find aesthetically important. It doesn't always respect the abstraction layers we've carefully constructed. It produces working solutions that we wouldn't have written ourselves.
But here's the synthesis problem: even if you don't use AI, your collaborators do. Even if you write every line by hand, you're working in codebases increasingly shaped by AI-assisted decisions. The style that feels wrong to you becomes the environment you have to navigate. The APIs you consume, the libraries you depend on, the documentation you read are all increasingly touched by AI generation. Youwill even have AI interacting with your tool on behalf of users. Even if you personally write every line of code by hand, you're building on top of abstractions that were increasingly designed with AI assistance. Your users are making decisions based on AI-generated summaries. Your dependencies were written by developers using AI copilots. The AI-shaped world becomes the water you swim in, whether you chose to drink from that particular stream or not.
There's no going back. Even if every AI coding assistant vanished tomorrow, the knowledge is out there. The patterns AI favors are being taught. The conventions AI finds easy are spreading. The Joseki that require human abstraction skills are being forgotten.
What To Do About It
So what should you actually do? Here's where the Go analogy becomes useful.
First: Don't memorize patterns. Learn principles. The Joseki that AlphaGo rejected were memorized by thousands of players who suddenly found their knowledge obsolete. The players who adapted weren't the ones who memorized more AI Joseki, they were the ones who understood the underlying trade-offs. Territory vs. influence. Local vs. global. Concrete vs. potential. These principles persist even when the specific patterns change. The Chinese national team didn't just copy AI moves; they eventually developed intuitions for why certain positions worked, even if the original discovery came from incomprehensible calculation.
In software: understand why design patterns exist, not just which ones to use. Understand the trade-offs between duplication and abstraction, coupling and cohesion, explicitness and conciseness. The specific recommendations will change. The underlying tensions won't.
Second: Embrace chaotic positions. AlphaGo found value in board states that humans avoided because they looked messy, complicated, hard to evaluate. The AI could calculate through the chaos; humans couldn't, so we steered clear.
In software, this might mean code that looks "wrong" by conventional standards but has properties AI finds valuable. It might mean accepting solutions you wouldn't have written yourself. It might mean developing taste for what works rather than what looks right.
Third: Recognize the training asymmetry. The developers who train with AI, who review its suggestions, develop intuition for its patterns, learn to read its chaotic style, will have advantages over those who don't. This isn't about replacing human judgment. It's about augmenting it with something that calculates differently than you do.
But be careful: don't become the student who wins with the 3-3 invasion without understanding why. Don't let AI-generated solutions substitute for actually learning the fundamentals. The patterns will keep changing. The principles last longer.
Fourth: Accept the synthesis problem. Even if you resist AI-assisted coding, you'll be working in an environment shaped by it. The new Joseki will become standard not because everyone adopts them, but because enough people do that the rest have to adapt. There's no pre-AI world to retreat to. The synthesis extends through your dependencies, your tools, your users' expectations as the entire ecosystem is being reshaped.
The Game Continues
Lee Sedol retired because he thought Go had become soulless. But the game is still played. There are still professional tournaments. People still find meaning in it, even when he doesn't.
What changed is the nature of mastery. It used to be about pattern memorization, about knowing thousands of Joseki and their variations. Now it's still about patterns, the patterns have just changed, and the way you learn them has changed too. Mastery is still pattern recognition, still tactical sharpness, still the ability to read complex positions. But the patterns you recognize are different, and you learn them through AI training in addition to human games. The fundamentals haven't disappeared; they've just found new expression.
Software is heading toward the same transition. The Joseki are the design patterns, the conventions, the accumulated wisdom - will change. Some will be discarded. Others will be transformed. The AI will find value in approaches we've dismissed, and dismiss approaches we've cherished.
The game doesn't end. It becomes different.
The developers who thrive won't be the ones who memorize the new patterns fastest. They'll be the ones who understand what the patterns mean. Who can read a chaotic codebase and see structure in it. Who can work with AI without surrendering to it. Who remember that territory matters, but influence matters too. And who can tell the difference even when the calculation gets hard.
This is the new meta. Adaptation speed, not just pattern recall. Understanding why, not just knowing what.
The Joseki are dead. Long live the Joseki.