The thing about these lists that put me on alert
Over the past six months I’ve read maybe twenty or thirty “AI-era super individual gear lists” seriously. Chinese ones, English ones, WeChat posts, long Twitter threads. They all look strikingly alike: a few large models first, then a few AI coding tools, then automation platforms, knowledge bases, editing tools, and finally a scoring table.
I use the vast majority of these tools myself. But somewhere around the fifteenth list, I started getting suspicious: the differences between these lists are far smaller than the advantages each of them claims.
If ten thousand people all receive the same list, install the same toolset, and lift their development speed to the same order of magnitude — who exactly did this gear give an advantage to?
The answer: it gave those ten thousand people an advantage over “the people who haven’t installed it yet.” Inside the ten thousand, nothing happened.
This isn’t a gotcha or a debating trick. It points at a fact I think is being systematically ignored: the biggest change of the AI era isn’t “individuals got stronger” — it’s that “getting stronger became a public good.” Every model upgrade ships to you and all of your competitors at the same moment. The lead you built yesterday on tools gets erased today by a single version update — not erased by any specific rival, erased for everyone simultaneously.
So I want to talk about gear differently. Not sorted by function, not tiered by price, but layered by a colder criterion.
The only test worth applying
For any piece of gear, I ask one question:
Does this layer’s progress help only me, or does it help all of my competitors at the same time?
This question splits all gear into two categories, and the line is remarkably clean.
The first category is entry tickets. Their progress is public, exogenous, and distributed to everyone at once. Models get stronger, AI coding tools get better, deployment gets cheaper — these improvements land on you without you doing anything, and they land on everyone else just the same. Investment at this layer can keep you from falling behind; it cannot put you ahead.
The second category is assets. Their progress is private, endogenous, and paid for in time. The two hundred essays you’ve written, the ten thousand readers you accumulated over three years, the first thing that comes to mind when someone hears your name, the contributions people remember you for in an open-source community — no model upgrade can hand these to somebody else. They can only be piled up by time and consistent behavior.
Telling these two apart matters a hundred times more than telling apart “which model is smarter.” Because returns on entry-ticket investment decay to zero as adoption spreads, while returns on asset investment compound with time.
There’s a counterintuitive corollary here that I want to state plainly: chasing the newest thing at the entry-ticket layer is, in essence, a liability. Not because the tools are useless — they’re extremely useful — but because the attention you spend there has an opportunity cost, and the marginal return of that layer is being diluted by the entire industry’s progress. If you spend two weeks tuning your toolchain to perfection, and then reinvest the time you saved into tuning tools again, you’re permanently stuck at that layer.
I know this reading is easy to distort into “don’t learn new tools,” so let me add: the necessity of the entry ticket and its unsustainability are both true at once. Without the ticket you can’t even get into the arena; holding the ticket doesn’t mean you win. The real question isn’t “should I invest” — it’s “at what point do I stop.” Almost nobody asks that question, because progress at the tool layer produces the illusion of personal progress so easily. You switch to a better model, delivery genuinely speeds up, the positive feedback is completely real — and completely capable of trapping you in that loop.
There’s an even sneakier part: entry tickets spread faster than you can iterate on yourself. Today you’re ahead on a carefully tuned configuration; three months later someone installs the defaults and lands at eighty percent of your position — because half the value of your configuration came from potholes you personally hit, and potholes vanish automatically as the ecosystem matures, with the gains split evenly across the whole industry. That’s why a tool-layer lead is always depreciating: its value isn’t taken by a rival, it’s evaporated by time.
The four-layer gear stack
Cut along this criterion and gear falls naturally into four layers. I’ve ordered them by how hard AI can flatten them, easiest to hardest — note the numbering direction: Layer 4 is the easiest to flatten, Layer 1 the hardest.
Hard for AI to flatten · strongest compounding
▲
│ ┌─────────────────────────────────────────┐
Layer 1 │ Reputation Why people trust you │ Asset
│ │ open-source credibility / consistent │
│ │ public output / real relationships │
│ └─────────────────────────────────────────┘
│ ┌─────────────────────────────────────────┐
Layer 2 │ Distribution How your work gets seen │ Asset
│ │ audience / content pipeline / being │
│ │ retrievable by AI │
│ └─────────────────────────────────────────┘
│ ┌─────────────────────────────────────────┐
Layer 3 │ Judgment What to build, what not to, │ Semi-asset
│ │ and to what standard │
│ │ demand gate / taste / judgment as rules │
│ └─────────────────────────────────────────┘
│ ┌─────────────────────────────────────────┐
Layer 4 │ Production Building the thing │ Entry ticket
▼ │ AI coding / agent orchestration / │
│ deployment / automation │
└─────────────────────────────────────────┘
Easy for AI to flatten · public good
I know this kind of layer diagram can easily become yet another scoring table, so let me say it immediately: these four layers are not parallel options, and you don’t need to pick between them. What you need is to figure out how much to invest in each layer, where to stop, and the direction that connects them.
Layer 4 · Production: provision to sufficiency, then stop
This layer is the entire content of most gear lists: AI coding tools, agent orchestration, deployment platforms, automation pipelines.
It’s the archetypal entry ticket. Every advance at this layer is a public event, synchronized across the whole industry. I’m not saying it doesn’t matter — the opposite: it’s a necessity. Without it you can’t even enter the arena. But its strategic position is “necessary but not sufficient,” and most people treat it as “necessary and sufficient.”
The correct way to use this layer: provision until it’s sufficient, provision cheaply, then stop. The test for “sufficient” is not “I’m on the latest tools” — it’s “my delivery speed is no longer my bottleneck.” Once the bottleneck moves — usually to “I can’t review fast enough” or “I don’t know what to build” — it’s time to move your attention up a layer.
I’ll take this layer apart in the second essay, including a fact most people haven’t noticed: once execution speed goes up, the real bottleneck shifts to review bandwidth — and review bandwidth is nearly impossible to expand the same way. That essay will cite some empirical data that isn’t comfortable to read.
Layer 3 · Judgment: once execution is free, this is where all the cost lives
AI can build what you’ve thought through. It cannot build what you haven’t thought through — but it will pretend to, and the result will look convincingly like the real thing.
This layer is the capacity for “what to build, what not to build, and to what standard.” It includes the demand gate (deciding which requests deserve a place in the queue), acceptance criteria (deciding whether output passes), and taste and curation (deciding what deserves to exist at all).
It’s a semi-asset: part of it can be externalized into rules and documents and reused; the rest can only live in a human head.
The value curve of this layer is peculiar: the cheaper execution gets, the higher the relative value of judgment. When building something took three months, the cost of pointing in the wrong direction was partially hedged by the long build itself — you had plenty of time to wake up midway. When building something takes three hours, the cost of a wrong direction is paid in full, and you’ll have finished five wrong things before you wake up.
Asking AI to generate requirements costs almost nothing. That sounds like good news; it’s actually the most dangerous link in this whole system — it fills your queue with tasks that look reasonable and are actually low-value, and then your agents will faithfully complete every one of them.
Layer 2 · Distribution: the only capacity still compounding
A product can be cloned overnight. The readership you spent two years building cannot.
This layer is the capacity for “how the things you make get seen”: audience, content pipeline, channel voice, and one item newly added in the age of AI retrieval — whether your content can be found and cited by models.
It’s a pure asset. Audience and trust are functions of time, and no tool can compress that function. A brand-new account and a three-year-old account posting identical content will get completely different responses, and that gap can only be closed with time.
This is the layer people with engineering backgrounds underestimate the most. I took this detour myself: I kept believing “make the product good enough and people will come.” That line held some truth back when building was expensive — because a good product was itself scarce. In an era where the cost of building trends toward zero, it has mostly stopped working: good products are no longer scarce; being seen is.
Layer 1 · Reputation: the last layer that cannot be flattened
“He’s the one who built this” — when that sentence alone can change how people judge a product, you have Layer 1 assets.
It includes your contribution record in open-source communities, the predictability you’ve built through long-term public output, and real human relationships.
It’s the hardest of the four to build, the slowest, and the hardest to flatten. The reason is plain: trust requires consistency across time, and time cannot be compressed. You can use AI to produce a hundred articles in a day; you cannot use AI to make a hundred people trust you in a day.
This layer usually isn’t a “tool.” It’s an asset — sometimes it’s just you, the person. And precisely for that reason, it’s the layer most often left off gear lists: a list can only hold things that can be bought.
That sentence is worth pausing on: anything a list can enumerate is already not scarce. The fact that a piece of gear can be written into a list, linked, and installed with one click proves it’s already a public good. So the property of “not being on any list” is itself the source of Layer 1’s scarcity. It can’t be listed because it can’t be bought; it can’t be bought because it can only be grown by time and consistent behavior.
Why almost every list stops at Layer 4
If the four-layer logic is this straightforward, why does almost every “super individual gear” discussion talk only about Layer 4?
I’ve thought of three reasons, and none of them is “the authors don’t get it.”
First, Layer 4 is the only layer that can be turned into a list. Tools have names, links, prices, and alternatives — they’re natively table-shaped. “Judgment,” “audience,” and “reputation” are not. You can’t write “Recommended: three years of consistent public output, ⭐⭐⭐⭐⭐.” The medium determines the content: as long as the format is a list, the content can only be Layer 4.
Second, Layer 4 has the fastest feedback. Switch tools and you feel faster the same day. Improved judgment takes months to show; an audience takes a year or two; reputation takes longer. Human behavior is naturally captured by the shortest feedback loop — that’s not a willpower problem, it’s a structural one.
Third, and most crucial: Layer 4 is the only layer money can solve. The other three run only on time and consistency, and neither can be added to cart. So when someone is anxious about falling behind, buying tools is the only action that relieves the anxiety immediately — even though it doesn’t solve the problem.
Once I saw these three points, my attitude toward those lists actually softened: they aren’t wrong, they’re constrained by their form. The real problem is on the reader’s side — taking a list that is structurally limited to Layer 4 and treating it as the complete answer.
The four layers have a direction
If this essay stopped at the layering, it would be just another scoring table. So let me stress the part of the framework that actually matters: the layers are not parallel — they have a direction.
The direction runs bottom-up: every unit of output from Layer 4 should settle upward into Layer 2 and Layer 1 assets.
Layer 4 produces a thing
│
├──► Direct value: the product's own revenue (linear, catchable)
│
└──► Settled value: turn "how it was built" into content
│
├──► Layer 2: the content brings an audience
│
└──► Layer 1: consistent output builds reputation
│
└──► Feedback: the next product's cold start
gets cheaper
If this direction isn’t wired up, every delivery you make is one-off: ship the product, earn a little, go back to zero, start the next product from scratch. Wire it up, and product N starts from the audience and reputation accumulated by products 1 through N-1.
This is the real difference between a “super individual” and “a person who’s very good with AI.” It’s not that the latter uses tools worse — it’s that none of the latter’s effort leaves anything behind.
I’ve seen too many technically excellent people stuck in the exact same place: absurdly strong at Layer 4, sitting on a Layer 1 foundation they’ve never deliberately worked, Layer 2 left entirely to organic growth. No pipes between the three. That’s not a talent problem. It’s an architecture problem.
How to diagnose which layer you’re stuck at
The value of a framework isn’t in classification; it’s in diagnosis. So before the next four essays, here’s a crude but useful self-check.
Find your bottleneck, and you’ve found your layer.
Q: If you were handed 5x execution capacity right now, would your output improve?
Yes ─────────────────────────► You're stuck at Layer 4 (production)
Keep provisioning gear — it's the right move
No, because I don't know what
to build ────────────────────► You're stuck at Layer 3 (judgment)
More speed just means building the wrong
thing faster
No, because nobody sees what
I build ─────────────────────► You're stuck at Layer 2 (distribution)
Excess capacity, blocked outlet
No, because nobody trusts
what I build ────────────────► You're stuck at Layer 1 (reputation)
The slowest layer, and the most worth
investing in
This self-check has one very practical use: it blocks the most common bad decision — the bottleneck is clearly higher up, and you keep investing in Layer 4 anyway.
I’ve made this mistake myself, and in textbook form: whenever output was disappointing for a stretch, my first reaction was always “maybe my tools aren’t good enough,” so off I went researching new workflows and new orchestration setups. Looking back, in at least two of those episodes the real problem was that I hadn’t figured out what I was building at all — a Layer 3 problem, treated with Layer 4 medicine.
The reason this happens is that Layer 4 problems are the only ones with clear solutions. Bad tool, swap the tool; rough process, fix the process — these actions come with a definite sense of completion. “I don’t know what to build” offers no sense of completion; you don’t even know when you count as having figured it out. People instinctively flee toward the side that has solutions.
There’s an even more alarming signal: if your bottleneck is in the upper three layers, every speedup at Layer 4 makes the problem worse. Speed up before your judgment is clear and you produce wrong things faster; speed up before distribution is wired and you pile up more output nobody sees. This isn’t a metaphor — it’s real wasted resources, and because the feeling of “I’m busy, I’m producing a lot” is so intense, it’s very hard to catch in yourself.
About “10x productivity” — cold water first
Before the next four essays, I want to get one thing straight, or this whole series would rest on a false premise.
Almost every gear list promises a number: 10x productivity, 5x efficiency, one person equals a team. The source of these numbers is, overwhelmingly, self-report.
And self-report is unreliable on this particular question. There’s one study I think everyone who writes code with AI should know. In July 2025, METR ran a randomized controlled experiment: 16 developers working on mature open-source projects they knew well, averaging five years of experience, 246 real tasks, randomly assigned to “AI allowed” and “AI prohibited.”
The result: with AI allowed, tasks took 19% longer.
But the more devastating figure is the other one: before the experiment, these developers predicted AI would make them 24% faster. After finishing, they still believed AI had made them 20% faster. The measurement said slower; the subjective experience said faster; and the experience never got corrected even after the fact.
I have to add an important update, or this number will get misused. METR itself published in February 2026 that a second round (larger sample, newer tools) had flipped to “speedup” — but they simultaneously judged that batch of data unreliable, because of severe selection bias: many developers refused to participate at all (unwilling to work under the no-AI condition), and thirty to fifty percent of participants admitted to deliberately withholding the tasks AI could accelerate the most. METR’s own words, roughly: we lean toward believing developers in early 2026 really are more accelerated by AI than in early 2025, but the evidence is weak.
So the correct conclusion is not “AI makes people slower.” The correct conclusion is the colder one:
Even the institution that studies this for a living admits it’s now very hard to cleanly measure AI’s effect on productivity.
That dismantles the “10x” pitch better than any specific number could. When something can’t be reliably measured yet reliably produces a strong subjective sense of acceleration, every numerical claim built around it deserves a question mark — including the ones in those gear lists.
This is not an argument against using AI. I use it every day, heavily. It’s an argument for reading the next four essays with one basic suspicion in hand: any claim of “install this and get X times faster” is dodging a more important question — faster, and then what?
What this column covers next
Five essays, one per layer, plus this overview.
The second essay covers the production layer : the current state of AI coding and agent orchestration, what to provision and what counts as over-provisioning, where the real bottleneck of overnight agents lives, and why “first-pass correctness” matters far more than “parallelism.” It will cite empirical data on the quality and security of AI-generated code — the numbers aren’t pretty, but they can’t be skipped.
The third essay covers the judgment layer : when execution trends to zero, judgment is the entire cost. How to guard the demand gate, how to write acceptance criteria, and how to externalize “judgment” — a thing that lives only in your head — into rules an agent can read over and over. That externalization is the only path by which this layer upgrades from “personal ability” to “reusable asset.”
The fourth essay covers the distribution layer : why an audience is the only capacity still compounding, how to turn the process of building products into content itself, and the newly arrived question of the AI-retrieval era — whether your content can be read, understood, and cited by models.
The fifth essay covers the reputation layer , and closes the series: why trust is the last thing AI cannot flatten, how “non-list assets” like open-source contribution get systematically undervalued, and most importantly — how to wire the four layers into one system that feeds itself, instead of four separate chores.
If you’re only going to read one, read the third. If your Layer 4 is already strong, skip straight to the fourth.
One clarification: this column is not a “just follow the steps” manual. Of the four layers, only the bottom one can be covered by a manual — and that’s precisely the least valuable layer. For the upper three, all I can offer is criteria for judgment and some potholes I’ve hit, because their answers are inherently personal. Any claim to a universal answer is, most likely, the certainty of Layer 4 being wrongly borrowed by the layers above it.
One last thing, at the risk of deflating the mood.
Writing this series, I kept circling back to the same thought: our generation may have overestimated how much of this story “tools” actually carry.
Not that tools don’t matter. It’s that when everyone’s tools are equally good, tools stop being the source of difference and become its background. The real dividing lines retreat to some very old things — what you want to build, how good your judgment is, how many people trust you.
In the era of scarce tools, these things were hidden behind the tool gap. Now the tool gap has been flattened, and they’re exposed again.
This may be the most sobering and, at the same time, the most grounding fact of the AI era: you can no longer win on gear. You can only win on you.
Responses