The Picture I Am Holding. What a Year of Building and Writing Adds Up To

In one of the essays this summer I wrote that the product of a hard week was never the code — "the corrected picture was the product, and the code was its receipt." This blog has now issued a lot of receipts. Two articles in three years, then twenty-six in a season, each in six languages, narrated in a clone of my voice, illustrated by scripts, several of them built on experiments that ran for real and kept their failures in. It is time to do for the whole body of work what I would demand of any system I ship: read the entire trace and state the model it implies. Not a recap. The picture.

Here it is, drawn once.

One mechanism, twenty-six essays

Every essay, whatever its subject, found the same split. A thing I cared about turned out to be two things wearing one name — a cheap half and an expensive half — and AI had made the cheap half free.

Knowing a language split into typing and depth, and typing went to the machine. Learning to code split into producing syntax and telling right output from wrong, and producing went. Being stuck split into tax and tuition. Reviewing code split into seeing and understanding, which feel identical from the inside and are not. Stack Overflow's archive split into facts and judgment, and the facts — the half the site's constitution optimized for — were automated first, by models trained on the archive itself. The questions between colleagues got fewer and heavier, because the light ones stopped needing a person. The polished artifact split from the making of it, and stopped being a signal the day anyone could generate it. Even the word "agent" split into the shape that decides at runtime and the shapes that merely contain an LLM, and I had to build the same system three ways and run it a hundred and fifty times to make the difference undeniable.

One mechanism, every time: when production gets cheap, the value does not evaporate — it moves. Scarcity changes address. The interesting question was never whether AI ends scarcity. It is where scarcity now lives, because whoever owns the new address owns the abundance.

The claim, stated plainly

Whatever a machine can make gets cheap. What no machine can do is tell if it's right, answer when it fails, add hours to your day, or love anyone. Those four are the job now.

That is the whole model in three sentences. The first is the mechanism. The second names the four addresses — judgment, recourse, attention, the person. The third is the conclusion, and it prices your career, your product, and your evening all at once. Everything I have built, written, and measured sorts under one of the four, so the rest of this essay defends them the way an argument deserves to be defended: a thesis, the reasons, and the strongest example I have.

First address: judgment

The thesis: rightness is not a property of the output. It is the fit between the output and an intention that lives in a person.

The machine produces plausible everything. Plausible is precisely the problem, because a language model is a machine tuned to produce exactly the signals a reviewer uses as proxies for correctness. So the first thing that repriced upward is the capacity to tell right from wrong when both look the same: the alarm that goes off in a deeply known language before you can say why, the reviewer who never merges what they cannot explain, the person whose answer to "what would you do?" is testimony weighted by consequences they personally paid, which is the one answer that cannot be generated on demand.

Judgment has a supply chain, and this is the uncomfortable part: it is built out of exactly the friction the machine now removes. The reps that made depth used to come free with the job. Now getting stuck is a budget you spend on purpose, at walls inside the problem rather than beside it, and depth stopped being something that happens to you and became a decision. This is also why the crowd still gathers to watch handmade code — over a thousand people around a hand-computed checksum nobody needs — because watching skilled making is how understanding propagates, and it is nearly the last class of content whose audience the machines have not absorbed.

The cleanest example I know is domestic. One day the machine will make my tea — kettle, steep, no sugar — and it will know my preferences to the decimal, because preferences are history and history is data. And the tea will still be wrong on the evening I want sparkling water instead. No amount of execution quality closes that gap. Rightness is indexed to a present intention, and the present intention is the one thing that is never in the record. My measurement experiment hit the same wall from the other side: the agent's "wrong" answer turned out to be a reasonable reading of an ambiguous world, and the fix belonged in the world, not the model. Right and wrong were never inside the output. They were in the fit.

Second address: recourse

The thesis: a decision governs only when an architecture can enforce it and a person answers for it.

Judgment is a person's property. The second address is what an organization has to build around it: someone who carries the cost of being wrong, and machinery that can make a no stick.

This is where my products live, and I only saw the pattern of my own portfolio while writing this. OrbitSpec refuses to answer before the question deserves an answer — a clarification engine whose deterministic rules are the floor and whose model is "a guest, never the host." plan-to-pr turns delegation into artifacts and puts a hard human gate in the middle, because a pipeline stage is code that cannot be skipped and "after enough approvals you stop reading" a gate that is merely a question. Smithproof is the no at fleet scale — registry, telemetry, policy, kill switch — built on the lesson that detection without response governs nothing and that a policy is a decision while enforcement is an architecture. Superforecaster replaces the feeling wearing a percentage with a 35% you can argue with, and scores itself afterward. Even the measurement piece is a recourse argument read backwards: you can tell whether something is an agent by what it needs — tracing, a budget that binds mid-run, a way to say no to it. Nobody puts a kill switch on a mail merge.

Specify the question, structure the delegation, govern the agent, forecast honestly. Four products, one conviction: the parts of an AI system you can trust are the parts that were never intelligent, and their job is to make the intelligent parts safe to use. The trustworthy half doesn't photograph well. All of it is the product — and it is also, not coincidentally, what my consulting sells. When a client asks me to scope AI music generation, the deliverable is the adapter, the storage, and the legal terms, not the API call. When the work is routing models across a GPU fleet, it is the boring endpoint that cannot fail users who don't know it exists. The models are nobody's moat. The value collects around them, and "around them" means: at this address.

Third address: attention

The thesis: the machines multiply what asks for your attention and cannot add one minute to its supply.

Turn from the supply side to the demand side and the same mechanism runs on the reader. Production now scales like software; attention still scales like sleep. A model can make more video. No model can make more of your Tuesday. So markets under generative pressure split rather than sink: the floor goes to generation, the top — where the buyer can feel the difference or someone must answer when it fails — reprices upward, and the anonymous competent middle dies first.

What gets paid, on this side, is choosing. Curation is the profession the decade reopens, and I have started practicing it rather than predicting it: a shelf, not a feed — forty-five films the Church listed in 1995, extended by the same three-lens discipline to books and music, every entry carrying a fact rather than an adjective. A shelf is finite, it commits, it can be argued with, and that is the entire difference: a feed optimizes for what you will click, which is a fact about your reflexes; a shelf claims what is worth your attention, which is a claim about the good, and claims about the good can only be made by someone standing somewhere.

Attention is also where my own flood lands. I am a tributary of the thing I describe — the six languages, the cloned voice, the covers. The 2023 version of me asked books questions and marveled; the 2026 version knows that answering got cheap and that the scarce thing was on the other side of the exchange all along: the evening of the person reading.

Fourth address: the person

The thesis: some goods consist of persons sharing them, and generation cannot touch those at all.

The last address is the one the other three keep pointing at. The machine can generate a better episode than the one everyone watched. It cannot generate the fact that everyone watched it. A perfectly personalised artifact is a room with one chair. So the goods that survive generation whole are the ones made of persons in the plural: the crowd at the potter's window, the master who actually answers, the couple whose product test I wrote for RotinaTop — does it end with two people looking at each other, or looking at it.

And because persons stand somewhere, this address is where ownership lives. The guild I built exists because the people whose skilled hands are becoming training data currently hand the library over for a wage or for nothing, and a royalty on what sells — revocable consent, signed licences, settlement on the rail machines will actually pay on — is what participation looks like when the asset is small and the sales are many. An allowance is what you give someone the system no longer needs. Being needed is most of what work gives a person, and no cheque replaces that.

The Catholic reading, offered rather than argued

I have kept the habit of marking this section clearly: what follows I hold as a Catholic, and I offer it rather than argue it. But I no longer think it is an appendix to the model. I think it is the reason the model was findable.

Start with a kettle. The mathematician John Lennox likes to ask why the water in it is boiling. Science has a complete answer — the element transfers heat, the molecules accelerate, the vapour pressure rises — and there is a second answer, just as true and unreachable from inside the first: because I want tea. His other version is his aunt Matilda's cake. Set it before all the scientists in the world and they can tell you everything about it — the nutritionists its calories, the biochemists its proteins, the physicists its particles — and not one of them can tell you why it exists, because the why is a birthday surprise, and only Matilda can reveal it. The two levels of explanation do not compete; that is Lennox's standing point about science and faith, and it is also, I now see, the exact shape of this essay. The machines are the most complete how-explainers ever built. And all four addresses are why-level facts. Whether it is right depends on what someone intends — the tea is wrong on the night I wanted sparkling water. Who answers depends on who promised. What deserves attention depends on what is good. And love is a why with a face. No accumulation of how ever adds up to a single why, which is the deepest reason the four keep not being automated — and why none of this makes me anti-technology, any more than Lennox's kettle makes him anti-physics.

Because at every address, the tradition got there first. On work: Laborem Exercens distinguishes the objective dimension of work from the subjective one and insists that man "is the subject of work, never merely its instrument" — which is, almost word for word, the split every craft essay found: the machines take over the instrumental half and leave the subject standing, more visible than before. On ownership: the guild is not a blockchain idea, it is Rerum Novarum's lament for the abolished guilds plus Magnifica Humanitas counting algorithms and data among the goods destined for everyone — and its participation clause is the one with teeth, because surviving is not participating. On attention: Augustine wrote amor meus pondus meum — my love is my weight — sixteen centuries before I wrote that where your attention goes is the inventory of what you love, kept whether you meant to keep it or not. Where your treasure is, there your heart will be also; that stopped being a metaphor the day attention became the treasure. On curation: an institution that spent fifteen centuries deciding which books belong in the canon and which music belongs at the altar handed me a working method in a 45-film list with its own humility clause. And on the day the machines have automated every necessity, Aquinas's acedia — sorrow at the good — will still be sitting there, undissolved by abundance, exactly as advertised.

The pattern I cannot unsee: the technology is a sorting force. It automates whatever was never the point, and everything my faith always said was the point — understanding, responsibility, attention, love — is precisely what comes out the other side repriced. The tradition keeps turning out to have already answered questions the technology only just made urgent. I do not think that is a coincidence. I offer it; you may file it as one.

What I am holding loosely

A picture held too tightly stops being corrigible, and the receipts this year include corrections to my own theses. The METR study found experienced developers who believed themselves 20% faster with AI were measurably slower — a standing warning that my own sense of acceleration needs receipts, which is half the reason I measure instead of testify. In that same experiment, my agent lost to a plain function twice before the data outgrew the prompt — the honest sequence was function wins, function wins, function vanishes — and the grader I wrote turned out to be more ambiguous than the model reasoning against it. The tool question I dismiss still matters at exactly one layer, the terms of the deal, and one changed deal cost me a four-figure week before I learned it. And the open questions from the attention essay stay open: how long institutional inertia protects the middle, whether generated companionship is a witness of a sort, what franchise concentration does to shared culture.

So the model above is held the way Superforecaster holds a probability: publicly, with reasons attached, scored later, revisable on evidence. That is not a disclaimer. It is the method.

The test

Strip everything above to one instrument and it is a single question asked four ways. Can you explain it — or did you merely see it? Will you answer for it — or does the cost of being wrong land on nobody? Did you attend to it — or did it merely pass through your feed? Does it end with persons — two people looking at each other — or with each of us alone in a perfectly furnished room with one chair?

Whatever a machine can make gets cheap. What no machine can do is tell if it's right, answer when it fails, add hours to your day, or love anyone. Those four are the job now. What I am holding, after a year of building the gates and writing the essays and running the numbers, is the old picture, corrected in my own handwriting — and everything I ship is just scaffolding around somebody doing those four things.

The scaffolding is good business. The four things are the point.