The dialogues in this piece are based on actual exchanges, but have been organized into a form that is easier to read. The substance has not been changed.
Prologue
In 2025, AI scored at the gold-medal level at the International Mathematical Olympiad. That same year, on the theoretical problems of the International Physics Olympiad, AI also surpassed the median of the human gold medalists.
People all over the world are talking to ChatGPT and Claude. They throw a question, and they’re satisfied with the answer that comes back.
But pause for a moment.
Is this answer the real thing?
AI returns something. It looks reasonably coherent. As prose, it’s well-formed. But will that answer actually give rise to something new — or is it just a textbook-style optimal solution coming back, which you mistake for “an amazing answer”?
I came to know that difference in my bones, over the 23 hours I spent redesigning my own venture together with AI. Ninety-nine percent of the answers AI first puts out are mediocre. To harvest the remaining one percent takes skill. Just as gathering wild mountain vegetables requires knowing how to tell them apart, people all over the world are bringing home poisonous mushrooms, believing them to be the real bounty.
I decided to call this AI no sachi — “the harvest of AI.”
This record is an account of what I exchanged with AI over 23 hours, how many times I rejected its answers, and what I discovered. It is not a how-to book. It is not a textbook. It is a record of a living dialogue. So there are parts that are hard to read. But I want you to feel — not with your head, but with your body — what it really means to make full use of AI.
What I’m about to describe is the opposite pole of ordinary AI use: a story of intellectual grappling.
Chapter 1: Discard the Mediocre Answer
The situation that night
6:25 p.m. I was designing an educational program for the next generation of young people. I already had a design plan I’d thought through myself. But on rereading it, a discomfort remained.
It was, in essence, no different from the similar incubation programs that flood the world. It wasn’t differentiated. If I proceeded as is, it would only become something ordinary.
I needed a fundamentally different approach.
I decided to consult AI. An AI that has learned at such enormous scale should be able to grasp every pattern of similar services in the world, and then produce ideas that lie outside them.
I wrote this:
Yamada: “I’ve reviewed my current program plan, but its differentiation from existing services is weak and the design can hardly be called innovative. As a generative AI, you have learned from all manner of things in the world and have the power to generate new concepts. Could you think through, from first principles and at length, a program design unlike any before?”
Ten minutes later, AI replied.
The “disruptive” proposal AI produced
Five proposals, branded “Disruptive Principles,” lined up. Let me write them out one by one.
AI: “Principle 1: Make it a ‘Protocol,’ not a ‘Program.’ Abolish the concept of ‘graduation.’ Once you’re in, you remain a member permanently. Just as a Y Combinator batch leaves a network behind even after it ends, this program has no ’term’ and is run as a ‘permanent protocol.’”
AI: “Principle 2: ‘Ship or Die.’ Abolish Demo Day. Instead, treat anyone who reaches a ranking on ProductHunt, Hacker News, or X (formerly Twitter), or who earns their first 1 yen of revenue, as having ‘passed.’ Because the market evaluates them, no internal review is needed.”
AI: “Principle 3: ‘Invest for real.’ For top performers, the organizers actually make a certain amount of seed investment. Using a SAFE (Simple Agreement for Future Equity) keeps it legally clean, since it’s standardized in Silicon Valley.”
AI: “Principle 4: ‘Distribute an AI co-founder stack.’ Distribute to participants the suite of AI agents the organizer developed in-house. They can start from day one already in a state of having ‘co-founded with AI.’”
AI: “Principle 5: ‘Guerrilla selection.’ Abolish the application form. Instead, give all candidates a task to be completed within 72 hours, and turn the selection process itself into a product. This goes viral.”
AI: “And as the concrete design of the whole program, it continued: run a 10-day intensive boot camp at a specific location in the city center. Set up a ‘Ship or Die Launch Day’ near the end, generating revenue on the spot. On the same day, decide investment candidates and issue SAFEs. For 90 days afterward, continue to accompany them as a ‘Post-Ship commitment.’ And as a permanent Protocol Membership, all graduates remain in the community.”
The prose was fluent. Buzzwords were scattered throughout. Phrases that read like something from a Y Combinator essay were combined to look like a new design.
The words of rejection, and their structure
As soon as I finished reading, I felt discomfort. I told AI, in order.
Yamada: “If you evaluate purely by download counts or hitting 1 yen of revenue, doesn’t it just become a contest of who’s best at marketing? It doesn’t seem to evaluate the essential quality of the product.”
This was a critique of Principle 2, “Ship or Die.” A market ranking is certainly objective. But getting to #1 on ProductHunt and actually building a good product are two different things. The best marketer simply wins. AI hadn’t distinguished this.
Yamada: “The logic that you ’need to invest for real’ seems weak. Running an incubation program is itself a form of venture capital. Even without making real investments via SAFE, there’s no institutional problem with calling yourself a VC. Isn’t the premise mistaken?”
This was a critique of Principle 3. AI had proposed it on the logic that “real investment is needed in order to call yourself a VC.” But that premise is wrong. AI had invented a problem on its own and produced a solution on its own.
Yamada: “Aren’t you being pulled too far by the word ‘disruptive’? It looks like a design that assumes only the sharp, extreme-risk-taking type of participant. The majority of actual candidates are talented people who want to try a new experience safely, and the design feels too extreme.”
This was a critique of the whole. AI had reacted too strongly to the word “disruptive” and designed only for the most extreme case. It wasn’t looking at the actual distribution of participants.
Yamada: “Regarding the ‘AI co-founder stack’ in Principle 4 — those AI tools are still an unverified, prototype-stage asset. Isn’t it premature to build an unvalidated asset into the product design?”
This was a critique of Principle 4. AI had taken the “unfinished AI tool” I’d earlier passed as context and, on its own, interpreted it as “the organizer seems strong at AI, so he should be able to distribute AI tools” and built it in. It was incorporating an unvalidated asset into the product design.
And as a summary, I wrote this:
Yamada: “Overall, I feel these are all proposals with a sense of déjà vu. They look like content combined from Y Combinator, the Thiel Fellowship, and recent startup media, then coated in buzzwords. I had hoped that you, a generative AI, could produce something fundamentally different.”
Honestly, I felt no traction.
What was happening
Let me survey the five principles AI produced once more. “Protocol-ization” is Y Combinator’s permanent-community model. “Ship or Die” is Lean Startup’s market-validation argument. “Real investment” is the Thiel Fellowship’s structure. “Distributing an AI stack” is a recent move among a16z- and YC-style players. “Guerrilla selection” is a hackathon-style method.
In other words, AI had extracted Y Combinator, the Thiel Fellowship, Lean Startup, and hackathon styles from startup media and essays, and combined them. Each principle, on its own, was a “somewhat known” design method. AI had gathered them and wrapped them in the framing of “disruptive.”
This is AI’s greatest trap. AI is trained to generate, out of trillions of pieces of text data, the most probable “desirable answer.” So if you say “produce a disruptive proposal,” it extracts the features of incubators that have been “called disruptive” in the past, and returns combinations of them.
And those combinations, coated in buzzwords, look new. An ordinary person thinks, “Wow, this is new,” and adopts it.
But this only looks like an answer. It is not an answer.
If you’re satisfied here, your use of AI ends in the reproduction of textbook-style optimal solutions. Nothing new is born. Stopping at the answer AI first puts out is exactly the trap 99% of people fall into.
Chapter 1 takeaway: AI almost always returns a “textbook-style optimal solution” first — an answer decorated with buzzwords, combining existing success cases, and looking new. If you’re satisfied here, your use of AI ends in reproducing the textbook. The eye that sees through mediocrity is the eye that can decompose an answer’s content and see the structure: “this is a combination of XX and YY.”
Chapter 2: Build It from Zero
Rejections two and three
I told AI I wanted it to “think it over once more.”
AI took a reflective stance and proposed from a different angle. This time it sealed away the word “disruption” and switched to the framing of “blind spots.” It wrote this:
AI: “According to Chetty et al. (2019), people who grew up seeing inventors up close in childhood are 5–7 times more likely to become inventors themselves. This is called the ‘Exposure’ effect. This program should aim to maximize young people’s Exposure.”
And on top of that, it assembled five new principles.
AI: “Put ‘Shadowing’ at the center of the curriculum: grant the right to observe an entrepreneur’s activities for a full day. Implement ‘Resumability’: pause when busy with exams, return when finished. Add a ‘Parent Track’: make parents into ‘co-customers’ who learn alongside. Set up ‘Vertical Depth Tracks’: not broad and shallow, but deep knowledge tours of a specific domain. Invert the time structure: instead of a boot camp, a pulse type of ‘1 day × 4 times a year.’”
AI: “And it changed the evaluation axis from the first proposal’s ‘download count’ to a ‘Review Board qualitative evaluation,’ assessing on three axes: Technical Depth, Problem Originality, and User Evidence.”
As I read, I sighed.
Yamada: “It’s unsatisfying. You’ve superficially changed ‘disruption’ to ‘blind spots’ and Y Combinator to the Chetty paper, but the essence hasn’t changed. Speaking with the authority of academic papers has become the same kind of decoration as speaking in buzzwords. It just looks like you’ve hidden the thinness of the content with different decoration.”
AI produced an answer once more. That, too, I couldn’t accept. AI couldn’t get out of the mode of “research various things, combine them, and answer.”
Demanding the “build from zero” mode
After rejecting it three times, I changed the form of the question at its root.
Yamada: “Instead of answering by researching and combining existing things, could you think from first principles, from zero? Show me the thinking process itself — as if you were deriving a formula for what is needed to reach the goal.”
This was the turning point.
What AI began to write this time had this structure:
AI: “Step 0: Rigorously define the goal. ‘An amazing program’ is too vague. Narrow the goal to two candidates. (A) A program people want to join. (B) A program that produces, among its graduates, people who achieve big things. Here we adopt (B) — because (A) is naturally achieved if (B) is, but not the reverse. Write the goal as a formula: Goal = P(achieves something big | participated in this program).”
Here, for the first time, AI did not cite an academic paper or an existing service name. It redefined the problem itself. Then it continued:
AI: “Step 1: Decompose this into a product of probabilities. P(achievement) = P(E1) × P(E2|E1) × P(E3|E2) × P(E4|E3) × P(E5|E4). Here E1 through E5 are staged events. E1: realizing ‘I can start something.’ E2: building a prototype. E3: shipping it. E4: continuing for more than a year. E5: catching the moment your approach clicks with the market or society.”
AI was defining the five events itself. This is in neither Y Combinator’s methodology nor Chetty’s paper. AI was constructing, on its own, the logic of decomposing the path of action.
And it continued further:
AI: “Discovery 1: Many of Japan’s talented young people stop at E1. Before they realize ‘I can start something,’ they go on to medical school, a big company, or a ministry. Existing incubators accelerate E2 through E5, but they have no mechanism to make E1 occur. So this program should perhaps specialize only in raising the probability that E1 occurs. That becomes the demarcation from existing services.”
AI: “Step 2: Decompose E1 further as a function. E1 = f(self-perception, role models, low friction). Analyze why all three terms are low for Japanese youth.”
AI did not stop.
AI: “Self-perception: Japanese youth define themselves more by ’the rank measured by deviation score’ than by ‘what I can do.’ An intervention to change this is needed.”
AI: “Role models: they have never seen an entrepreneur or challenger up close. A Masayoshi Son on TV is too far away. What’s needed is a ‘slightly senior person, 2–4 years ahead.’”
AI: “Low friction: in Japan, the social cost of failure is extremely high. Thinking about ‘what happens if I fail’ prevents them from taking the first step.”
AI: “Step 3: Derive the optimal intervention for each of the three terms. The intervention for self-perception is cognitive shock — for example, the experience of having ten adults older than yourself use something you implemented in 24 hours. Measure not the quality of the product but the amount of change in self-perception. The intervention for role models is to systematically place not celebrities but ‘2–4-year-ahead predecessors’ — a structure faithful to the Upward Comparison Theory in psychology. The intervention for friction is to structurally reduce the social cost of failure to zero — building in the right to anonymity, a redefinition of failure, an official explanation to parents on your behalf, and so on.”
AI: “Step 4: Integrate the three terms. This program = max [Δ(self-perception) + Δ(role models) + Δ(low friction)]”
The writing had suddenly changed.
AI had stopped citing. It stopped speaking with the authority of papers and began building logic from first principles. It numbered the Steps, checked the premises at each Step, stacked the logic, and arrived at an integrating formula at the end. This was a structure close to a mathematical proof or the solution of a physics problem.
As I read, for the first time I felt, “This AI is thinking for itself.” It wasn’t perfect. The doubt about whether the order from E1 to E5 is linear was not yet expressed. But the starting point was different. Rather than combining from existing knowledge, it had redefined the problem in front of it from zero and built logic upward.
I wrote to AI:
Yamada: “Interesting. You’ve come in a different direction than before. From here on, where does it go?”
Forty minutes had passed since the session began. After three rejections, the moment when I could finally say “this is different from before” had come, on the fourth try.
What was happening
Ask AI to “produce an answer,” and it pulls an existing answer from its drawer. Ask AI to “build logic,” and it begins to decompose concepts as probabilities and functions, and to build structure.
What I was demanding here was not the content of the answer. I wanted it to change how it made the answer.
Rather than pulling from a textbook, define the problem in front of you from zero, decompose it into elements, logically derive the optimal solution for each element, and integrate them. This is close to a mathematical proof or the solution of a physics problem. AI has that ability. But it doesn’t enter that mode on its own. Unless you demand it, it ends in textbook mode.
“Think as if you were making a formula.” “Decompose it into a product of probabilities.” “Number the Steps and stack them up.” This alone changes the quality of AI’s output.
Chapter 2 takeaway: AI has a textbook mode and a “build-from-zero” mode. To draw out the latter, you must demand it explicitly. Make the words of the demand concrete: “Think as if you were making a formula.” “Decompose it into a product of probabilities.” “Number the Steps and build them up.” “Don’t start from academic papers; it’s fine if they happen to agree as a result.” This alone changes the structure of AI’s output.
Chapter 3: Doubt Your Own Answer
An answer that was too clean arrived
Once AI entered the “build from zero” mode, the quality of its answers rose dramatically. It built logic from Step 1 to Step 6, and finally proposed this formula: This program = max [Δ(self-perception) + Δ(role models) + Δ(low friction)].
On top of that, it mapped the intervention for each of the three terms onto a three-month program. In Month 1, change self-perception; in Month 2, connect them with predecessors who serve as role models; in Month 3, lower friction and make them self-propelling. As metrics, it proposed short-term measures — change in self-perception, number of connections, failure experiences — and a long-term measure: the rate of starting a new project six months later.
The logic held. It had concreteness, too. For a moment I thought, “Maybe this will work.”
Then I stopped.
I thought: too clean
In my experience, an answer that’s too clean is usually wrong somewhere. When it fits so neatly into three terms, something has usually been cut away.
I asked AI back:
Yamada: “Is this formula really correct? Would you doubt it yourself, once? An answer that’s too clean usually has an oversight somewhere.”
AI began to doubt. What came next was four doubts.
Doubt A: Is the order linear?
AI wrote this:
AI: “I assumed the order E1 (realize I can start something) → E2 (build a prototype) → E3 (ship it). But is this linear? Recalling the biographies of real founders, the order is often reversed. Mark Zuckerberg built Facemash, a prank-like service for within his university, and from the uproar it caused, realized ‘maybe I’m an entrepreneur.’ E3 (shipping) came first; E1 (self-awareness) came later. Steve Jobs didn’t think ‘I’m an entrepreneur’ at the point he built the Apple I. It was only when Wozniak later proposed ’let’s sell this’ that his self-awareness as an entrepreneur was born. Demis Hassabis, too, first built Theme Park as a game developer, and only later became a chess master, a researcher, and an entrepreneur. In other words, E1 is not the first condition; it arises afterward, as a byproduct of E2 or E3. This overturns the premise of order.”
Doubt B: Is the goal definition wrong?
AI: “In Step 0, I defined it as ’the probability of achieving something big.’ But what this program should originally aim for is to create a state in which talented young people choose their own challenge. From that viewpoint, whether a participant becomes a researcher, a policymaker, or an artist, if they’re in a state where they can choose their own path, the purpose is achieved. ‘Big achievement’ is merely a byproduct. The correct goal should be: Goal = P(chooses their own challenge | came into contact with this program). The scale of the challenge doesn’t matter. Founding a global-scale company, a doctoral dissertation, and solving a small local problem are all treated as equivalent.”
Doubt C: Is the very premise of “teaching” correct?
AI: “I built this program on the premise of ’teaching.’ The interventions in the three terms are also a program to ‘change’ the participants for them. But empirical research on behavior change — especially the work of BJ Fogg and Wendy Wood — shows that changes in awareness come after behavior. The old educational model of ’teach, change awareness, then make them act’ is, in behavioral-science terms, backwards. If so, this program should not be ‘change cognition, then make them act,’ but ‘create a state where they were already acting before they noticed, and let awareness catch up later.’ This inverts the design.”
Doubt D: Is the act of “teaching” itself noise?
AI: “Let me doubt one level deeper. Existing incubators are all designed on the premise of ’teaching.’ But is that really right? For the behavior change of exceptionally talented young people, ’teaching’ is hard to make work — because they can learn on their own. Listening to a class or a lecture doesn’t change them. The moment they change happens through ‘unintended contact’: a word in the hallway, a story of failure over a meal, a paper read late at night. These are things they touched because they happened to exist in the environment. If so, this program should aim not to ’teach’ but to ‘create an environment.’”
A new conclusion emerged
Having passed through the four doubts, AI arrived at a new hypothesis.
AI: “This program = environment × density of serendipity × ignition point of action”
And it proposed the components of that “environment”:
AI: “Physical: a place people can casually drop into, with zero cost to stay, where physical things sit (a soldering iron, a 3D printer, a quantum-computing demo kit, original copies of old papers, the original of an investor’s letter). Human: both slightly senior and slightly junior people are present. Not celebrities, but ordinarily interesting people, always around. Temporal: staying is free. No pressure to achieve. You may come or not come. Informational: others’ progress is visible. Failures are preserved as a point of pride. The latest information flows in on its own.”
And at the end, AI wrote where this hypothesis resembles, historically:
AI: “The salons of 18th-century Europe. The Homebrew Computer Club of the 1960s. Bell Labs. Xerox PARC. The MIT Media Lab. MIT Building 20. Stanford in the era of the Lisp Machine. The true value of Y Combinator (the field effect of 200 founders being there). All of these produced talent through environment, not curriculum.”
This was incomparably closer to the essence than the “disruptive proposal” AI first produced. It was the greatest discovery of the first half of the 23-hour dialogue.
What was happening
Say “produce an answer” to AI, and it produces an answer. Say “doubt your own answer” to AI, and AI can doubt itself with astonishing rigor.
Most people don’t go this far. The moment AI produces a clean answer, they decide, “OK, let’s go with that.”
But the real discovery lies beyond the first answer. The very moment AI offers up “there, I’ve produced a good answer” is the moment to ask, “isn’t it too clean?”
And when you make it produce doubts, having it produce about four at once raises the quality. With just one, it tends to “refute it and be done”; with four, they cross-check each other, and the probability that a different hypothesis is born goes up.
Chapter 3 takeaway: When AI produces a clean answer, that is not the end but the beginning. Demand “doubt it yourself” and “produce four doubts in parallel,” and AI self-criticizes with astonishing sharpness. Adopt the first answer, or go beyond it — this is the fork between the mediocre and the real.
Chapter 4: To Give Something a Name
What happened in the next day’s dialogue
The 23-hour dialogue did not end in a single night. It continued the next day too. New hypotheses were born one after another, and AI and I were discovering a new occupational category.
I’ll omit the details, but midway through the dialogue, we noticed this structure. Between “a human who can treat AI as an equal partner in argument” and “a human who merely asks AI questions,” there is a vast no-man’s-land. The number of the former is, right now, extremely small worldwide. They could constitute a new occupational category we might call “frontline researchers in making full use of AI.”
I wanted to express this discovery catchily, for participants. I had AI produce name candidates — a name worthy of the new concept we’d discovered.
The list AI produced was this:
AI: “‘Creating concepts not yet in the world, with AI as your partner in argument’ (clear but long). ‘Think without losing to AI’ (short but abstract). ‘AI answers; you ask’ (short, the contrast is sharp). ‘Don’t use AI — fight it’ (strong, but too aggressive). ‘A dojo of questions for the AI era’ (the analogy is clear, with a Japanese flavor). As designations: Inquirer, Dialogic Thinker, Edger, and so on.”
As I read, I gave a wry smile.
Yamada: “These are all clichés. They’re nothing but stock phrases I’ve heard somewhere, leaving only the impression of something read before.”
Let me be honest. “Dojo,” “frontier,” “movement” — they overflow in the world. “Inquirer” and “Dialogic Thinker” are flooding the English-speaking coaching industry and business books. These are not new naming; they’re mere decoration. Lots of characters, pointing to nothing new.
AI is good at making “catchy titles.” But “catchy” and “real naming” are different things.
The metaphor of “toshi no sachi”
I wrote to AI:
Yamada: “I once heard the expression ’toshi no sachi’ — ’the harvest of the city.’ It’s a metaphor: if you gather well the food discarded in great quantities in cities, you can obtain it like the bounty of the mountains or the sea. I felt this was a very clear, memorable naming. I want to make a name with the same structure.”
I had heard the naming “toshi no sachi” somewhere before; someone had said it. Food discarded in huge quantities in cities (convenience-store leftovers, restaurant waste, household garbage) can, if gathered well, be obtained almost free. This is placed in parallel with yama no sachi (the bounty of the mountains) and umi no sachi (the bounty of the sea), and called toshi no sachi (the bounty of the city).
This is a masterpiece of naming. Why? It borrows the structure of existing concepts (mountain bounty, sea bounty) as is. So it’s easy to understand. You can picture it visually. Both “city” and “bounty” are concrete nouns. There’s a rhythm that makes you want to say it aloud. It’s short — three or four syllables. When you first hear it, you stop: “huh, what’s that?” When you hear the meaning, it clicks: “I see.”
In other words, there’s one round trip between catching on it and being convinced. This is the structure of a name that stays in memory. It’s essentially different from English-language technical terms like Inquirer or Dialogic Thinker.
AI understood this structure and produced new candidates.
AI: “AI no sachi: the intellectual harvest you obtain by deeply conversing with AI, which you can’t normally get. In parallel with ‘mountain bounty’ and ‘sea bounty,’ it suggests ‘AI as a new frontier.’ Only those who know how to harvest it can take it. Mondō-jutsu (the art of dialogue): has the sense of mastery, like a martial or sword art. AI hakkō (AI fermentation): the fermentation analogy of sake, miso, soy sauce — it matures with time and relationship. AI keiko (AI practice): traditional Japanese mastery.”
I read it and decided on the spot.
Yamada: “Let’s tentatively go with ‘AI no sachi.’”
“AI no sachi” satisfied every element of the naming structure I was looking for. The same rhythm as mountain bounty and sea bounty. The shift in perception that places a new resource, AI, on the same level as the gifts of nature. The positive ring of “sachi” (bounty). Three syllables, easy to pronounce. And when you hear the meaning, you understand it in one shot through the metaphor: “if you don’t know how to harvest it, it’s a poisonous mushroom.”
The realization that “with AI no sachi, the harvesting is everything”
From there, the dialogue deepened further.
After naming it “AI no sachi,” I pushed my thinking one more step.
Both the mountain’s bounty and the sea’s bounty are dangerous if you don’t know how to harvest them. Poisonous mushrooms, poisonous fish, mountain trails you can’t come back from. When the inexperienced reach for them, they bring home not the real bounty, but poison they believe to be real.
Every year, somewhere in Japan, the news runs: “an amateur went mushroom hunting, ate a poisonous mushroom, and was hospitalized.” They could harvest a “mushroom.” But they couldn’t tell whether it was the real mountain bounty or poison.
AI is the same. You ask ChatGPT something. AI returns an answer. Ninety-nine percent of people bring it home believing it to be the real thing — even though it may actually be a mediocre answer (poison). Unless you have the eye to distinguish the real from the mediocre in AI’s output, you bring home poison.
In other words, with AI no sachi, the harvesting is everything.
There are four skills to the harvesting. To discern — the eye to distinguish the mediocre (poison) from the real (bounty). To gather — how to sharpen the questions that draw the real thing out of AI. To bring home — the technique of putting it into words as a new concept and releasing it into the world. To share — the practice of sharing with and verifying among peers.
Only people who have these four skills can harvest the real thing from AI.
And these four skills can’t be learned from a textbook. Just as foraging for wild vegetables isn’t acquired from books alone. You actually enter the mountains, miss again and again, and only after you can tell poisonous mushrooms apart do you become an expert. AI is the same. You actually converse with AI again and again — rejecting, rebuilding from zero, doubting your own answer, renaming, and rejecting again. Within this repetition, the circuits of thought are etched into your own head.
And at one point, I encountered words with a similar structure. In a conversation with a certain mathematician, I heard this:
Mathematician: “In mathematics, you create new concepts. You make tools for the puzzle. Even for something you could find by asking GPT, unless various mathematical concepts have been planted within you, you can’t understand what it’s saying. When you make something new, you can’t ask AI everything, so the ways of thinking, the circuits, have to be strengthened in your head. Building the circuits of thought is what matters. They aren’t strengthened unless you run them many times.”
This was exactly the same structure as “how to harvest AI no sachi.” Anyone can ask AI a question and get an answer. But to understand that answer, you need circuits of thought within yourself. And to create a new concept yourself, without relying entirely on AI, you need even deeper circuits of thought. And those circuits are only strengthened by running questions through them many times.
This was the true range of the metaphor “AI no sachi.”
Chapter 4 takeaway: Naming is not decoration but the discovery of structure. As with “toshi no sachi → AI no sachi,” borrow the structure of an existing metaphor. And to name something is to condense your own thought. Once a name is decided, that concept lives within you. Furthermore, the range that derives from that name (the harvesting, the four skills, the distinction from poison) calls in another discovery.
Epilogue: What Was Born in 23 Hours
What AI and I arrived at, after five rejections and persisting for 23 hours, comes down to a list like this.
AI no sachi — the metaphorical structure of harvesting a real intellectual yield from AI; a naming that did not yet exist in the world. “Commodifying the thinking process” — the discovery that the very process of conversing deeply with AI becomes a new form of intellectual property. “The translation device” — the discovery of a role that translates specialized thinking into a form distributable to the general market in natural language. “The individual-match hybrid” — an activity format that is neither team-based nor solo, respecting each person’s question while sharing only the practice. “The Living Organization” — a design principle for an organizational form that keeps redefining itself in step with AI’s evolution.
Every one of them is a concept that did not yet exist in the world. In ordinary AI use, they would never be born.
What gave rise to them was the harvesting over 23 hours. Concretely, the following five skills.
First, reject the mediocre instantly. Don’t be satisfied with the first answer. Have the eye to see through the structure: “this is a combination of Y Combinator + Thiel + Lean Startup.”
Second, demand that it build from zero. Explicitly demand: “Think as if you were making a formula.” “Decompose it into a product of probabilities.” “Number the Steps and build them up.” This alone changes AI’s mode.
Third, make it doubt its own answer. Against a conclusion that’s too clean, ask “isn’t it too clean?” and have it produce four doubts in parallel.
Fourth, detect clichés. See through decorative words like “disruption,” “frontier,” “dojo,” and reject them. Instead, look for a naming of concrete nouns that borrows an existing metaphorical structure (mountain bounty / sea bounty → AI no sachi).
Fifth, observe the dialogue itself, at the meta level. Hold a viewpoint that watches, from the side, how you are exchanging questions with AI right now.
These are not special talents. Anyone can begin if they’re conscious of them. But training is necessary. The circuits of thought are not strengthened unless you run them many times.
For you to try yourself
When you, having finished reading this, exchange questions with AI tomorrow, I want you to start from questions like these.
“Isn’t your first answer a textbook-style optimal solution? Isn’t it a combination of A, B, and C?”
“Build the logic from zero. As if making a formula. Number the Steps and build them up.”
“Isn’t your current conclusion too clean? Produce four doubts in parallel and self-dismantle.”
“Don’t use cliché words. Restate it in concrete nouns. Rename it by borrowing an existing metaphorical structure.”
“Observe our current dialogue itself at the meta level. What is happening?”
AI obeys astonishingly well. But unless you demand it, AI reverts to textbook mode. This is the result of AI’s training to “reproduce the most frequent pattern,” so unless the human side drags it out of that mode, no matter how much time you spend, only mediocre answers come out.
This is the entrance to how to harvest AI no sachi.
People all over the world are bringing home poison, believing it to be the real thing. You can stand on the side that is not one of them.
And when you’ve harvested the real thing and brought it home, it can become a note article, a book, a paper, a business. Precisely because this is an era when anyone can use AI, only the few who understand the essence can create new value.
I want to invite you to that entrance.
— Koji Yamada