"What is it you want, old man?"
The Broken Trough §
The old woman in Pushkin's tale had a perfect executive agent: instant, tireless, unlimited budget. A trough, a cottage, noble rank, a whole kingdom, all delivered on time and bug-free. It ended, as we remember, at a broken trough.
The fish never once erred. The intention did. Incoherent: each granted wish bred the next. Ungrounded: she had no idea that ruling the sea would leave her cold. And not her own: what she wanted was never a cottage or a crown but the bowing.
Genie stories were never about genies.
Now the fish comes by subscription, twenty dollars a month. Execution (code, prose, analysis) grows cheaper with every release, now that it has stopped being scarce. Which raises the common question: what in me stays valuable?
The answer is smaller and sturdier than you'd like. What stays is whatever can't be delegated in principle. There is exactly one such thing, and it can be trained.
no gentle slope
Every six months, another wave. Models write the code we wrote yesterday, the prose that was a profession yesterday, the analysis that was expertise yesterday. Everything we took pride in and got paid for is cheapening before our eyes. Not from any drop in quality; it simply stopped being rare. And everyone asks the same question, with more anxiety in it than curiosity: what should I learn? what do I hold on to? what in me will still be worth something in five years?
The Ladder of Delegation §
Managing the agents will remain valuable, true, but only for a while: agents already manage agents, and every rung you retreat to ("fine, I'll just supervise") is automated a year later.
The ladder of delegation doesn't end in a cleverer activity. It ends at a point: the source of preferences. No one can want on your behalf, not because it's forbidden but by definition.
Ronald Coase and the prosthetic will
Outside, in the economy, falling coordination costs shrink the firm. Ronald Coase explained this back in 1937, in "The Nature of the Firm": a firm exists as long as coordinating inside it stays cheaper than the market, so the optimal organization contracts toward one person with a tree of agents beneath them. Inside, in the psyche, the same thing happens: an agent that knows your state, your calendar and your priorities, and settles the small decisions for you, is doing the work of executive control, the very thing we usually call the will.
In both places the bottleneck is identical: not computation, not skill, but clarity about your own wanting.
Where do I end, if the agent becomes a prosthetic will? The line runs by role, not by task. To the adviser you can hand almost everything: counting, reminding, showing you the odds. The principal's seat, the act of deciding itself, you keep for the cases where the decision shapes you, or trains a faculty you'd hate to lose, or plays out on hostile ground.
the principal's burden
The line runs not by task ("the alarm clock, fine; my salary, no") but by seat. The principal's seat, the act of deciding, is worth holding wherever at least one of three holds:
- the decision is constitutive: a conversation about your salary isn't "computed"; it forms your sense of your own worth, which is the whole point of having it;
- the decision trains a faculty you'd hate to lose: your feel for the right moment for a hard conversation atrophies if the agent always decides;
- the field is hostile, and the data the advice rests on can be captured: your employer's HR bot will be delighted to tell you when to ask for a raise.
Judgment you can rent; the will you can't. That is exactly why people don't stop mattering to one another. When execution is free, what we buy from each other is calibrated judgment: a critic's taste, a mentor's eye, a doctor's discernment, plus help in growing our own. The economy of doing turns into an economy of trust and taste.
So the final human skill of the AI age is not prompt engineering, and not "managing agents." It is the ability to want at high resolution. Sounds nice. But what does it actually mean?
The Anatomy of Wanting §
We're used to treating a desire as a point: "I want X," and there's nothing more to discuss. It's more useful to treat it as an object with internal structure, one that has measurable properties. I count seven, grouped in three families. The number is a working grid, not scripture; which of the seven carry the load, and why, comes right after the list.
Richness — how much structure is inside the desire.
- Resolution. How many gradations you tell apart. A designer sees twelve steps between "good enough" and "right"; the client sees two. A sixteen-color palette against truecolor: both are pictures, but you live in them differently.
- Dimensionality. How many properties you even have an opinion about. A novice picks wine by price. A sommelier holds twenty features in mind and, more to the point, knows which to sacrifice.
- Horizon. Whether the desire reaches the person you'll be in ten years, and reaches far enough to change what you do today. Wanting "now" is universal; wanting on behalf of your future self is a skill.
Integrity — whether you can trust this desire as a bearing: whether it contradicts itself, whether it's right about the world, and whether it's truly yours. Three different questions, three axes.
- Coherence. Whether your preferences run in circles (A over B, B over C, C over A), and whether they flip with your mood or the way a choice is framed.
- Groundedness. Whether the desire is tied to true facts about what will actually satisfy you. The classic case: wanting a promotion for years that, it turns out, won't make you happy.
- Authorship. How much of the desire is yours, versus quietly imported by advertising, status, someone else's feed. Girard was right: desire is mimetic; we want what others want, and most often we don't notice the swap.
Bandwidth — whether you can get it out.
- Articulability. Inside it may be rich; on the way out it becomes "make it, you know, hit hard." The loss in compression between the latent desire and the spec you hand the executor. Until recently this was the most expensive axis. Now, see below.
And the promised note on the number. There are five load-bearing axes: groundedness, coherence, authorship, dimensionality, horizon. Resolution and articulability are there for completeness: the first you can underpay but can't harm, the second AI has already collapsed to near zero. So "seven" reads more honestly as "five, plus two," and the number itself is a hypothesis about how many columns there are, one the same journal is free to refute, not a fact about the soul.
why these five
By an axis I mean a property that (1) can be read off a single journal on its own and (2) changes what you delegate. The five load-bearing ones are picked out by two independent tests, both pointing at the same spot: they are the axes where a miss is most dangerous, amplified by execution rather than merely underfed, and the axes that read separately from a blind reconciliation. The other two fail both: resolution is bounded above, not below (you can underpay, you can't do harm), and articulability AI has zeroed out. Both tests unfold below, in the distinction between "ordinary" and "dangerous" poverty, and in the chapter on measurement.
Why this taxonomy? Because AI acts on these axes very differently, and that difference is the crux.
AI multiplies execution by the spec you gave it. So it boosts articulability instantly and almost for free: anything stated clearly is now buildable. But a shortfall in integrity it exposes without mercy. An incoherent or ungrounded goal it optimizes perfectly, then delivers, fast and polite and on time, exactly the thing you didn't really want.
Hence the central distinction of this essay. Ordinary poverty (low resolution, small dimensionality) merely underserves: you get the mediocre. Dangerous poverty, the poverty of integrity, actively harms, because execution amplifies it. Multiply by muddy and you get a great deal of muddy. The old woman at her broken trough is dangerous poverty with an unlimited execution budget.
The diagnostic sign of dangerous poverty: the better the execution, the worse the outcome. A strong agent optimizes a broken goal flawlessly, and so does more damage than a weak one.
one example per integrity axis
Groundedness broken. You tell the agent, "make me as productive as possible." It flawlessly clears your calendar: aimless walks, long lunches with friends, "empty" slots. All optimized to zero. And those were exactly what the life you were doing this for was made of. The grounded goal was "a life I won't regret"; the ungrounded proxy, "productivity," is executed brilliantly and leads in precisely the opposite direction. A weak executor would have been safer here: it couldn't have reached the walks.
Coherence broken. Your preferences flip with your mood: Monday you want to drop everything and leave, Tuesday you want stability. This used to cancel itself out; the impulse passed before you could act on it, and slow execution filtered your inconstancy. A fast agent quits your job on Monday, while the impulse is still warm, and freezes the fleeting into an irreversible fact. With money it's harsher still: inconsistent preferences are the money-pump argument (A>B>C>A can be milked in a loop), and a tireless executor will find that loop and drain it.
Authorship broken. You want what your feed wants: the body, the house, the career from the picture. The agent delivers it, beautifully and on schedule. The trouble is that you've spent real years and real resources reaching someone else's terminal goal, and you'll notice the swap only at the finish line. The amplification here is the most insidious: the same infrastructure that shapes the desire (recommendation feeds) now also fulfills it. The loop closes with no human inside; desire is imported and satisfied through one and the same captured channel.
The through-line of all three: multiply by muddy, get a lot of muddy. "Ordinary poverty" (low resolution, small dimensionality) merely underfeeds; you get the mediocre, but harmless. "Dangerous" poverty is amplified by the very tool that was meant to help, which is why the old woman, with an unlimited fish budget, ends up worse than she'd have been without it.
AI doesn't so much make us more productive as expose that, for the most part, we don't know what we want. This ignorance used to be cheap: years of labor stood between a wish and its result, so the wish had time to ripen or fall away. Now only minutes stand between them. For the first time in history, not knowing yourself has become expensive.
Measuring What Lies About Itself §
How do you find out where you stand on these axes? Asking yourself won't do: self-report doesn't measure the trait, it measures the self-portrait, and the self-portrait flatters. Systematically. Judge yourself on evidence, not testimony: on your choices, and on your forecasts checked against outcomes. Anyone who has honestly reread a diary from a year ago knows this better than the psychologists do.
You don't need seven instruments. One decision journal is enough: every wanting that matters, with its options, the choice, the motives, and a forecast of feeling. After that, time does the work; the axes are just different readings of the same journal. Reread it blind → coherence. Live to the appointed date → groundedness. Ask "do I still sign off on this?" → authorship.
And here is where discipline hits its ceiling; two things are past it. The first is blinding. You can't honestly re-ask yourself six weeks on: you remember your old answer, and memory obligingly fits the new one to it. The second is return. You won't come back to the entry on the appointed day; life gets in the way. Neither gap is closed by willpower, only by a simple tool: it keeps your forecast sealed until you answer afresh, and it comes back for you on its own, right on time.
And Goodhart: the moment a measure becomes a target, it goes bad. No streaks, no points. A compass, not a thermometer.
One last objection remains, the deepest one. Whose compass am I measuring with? Isn't choosing the method itself an act of will? Maybe I should trust only my own sensations? Half of this is right: the decision to measure yourself, the choice of axes, and the final "yes, that's me" are the principal's seat; they truly can't be handed to a methodology or a tool. But the other half inverts the real order of reliability. Raw sensation is the most capturable layer of all: it's molded by lack of sleep, blood sugar, status anxiety, and yesterday's feed. Method is not the will's rival but its lever: a ruler doesn't take away your sense of proportion, it hands you a number to judge. The true picture is Neurath's boat: we rebuild the ship on the open sea, plank by plank, never putting in to dry dock. What's sovereign is neither feeling nor rule, but their mutual correction against outcome. Feeling is the last judge, but a poor witness. And the fear "have I imported someone else's method?" dissolves under that method's own instrument: if you can name why you took it and under what condition you'd drop it, the method is yours, wherever it came from. Everything worthwhile came from outside: language, mathematics, values. Authorship isn't about the source but about adoption after scrutiny.
what psychoanalysis would say
That last line, "authorship isn't about the source but about adoption," Lacan would sign: his ethics is precisely to assume your desire, not to trace its pedigree. But then he'd object to the whole frame. For psychoanalysis (Freud, Lacan) desire is unconscious, structurally insatiable, and foreign from the start: "desire is the desire of the Other." In that light, the three things the journal treats as faults are the very nature of desire. The gap "forecast ≠ fact" is not a calibration error to be driven to zero but its constitution: you always miss, because you never wanted the object, only to keep wanting. Incoherence is not noise but the voice of a split subject: Monday's "drop it all" and Tuesday's "stability" aren't a glitch, they're the conflict that is you. And a blind reconciliation is still the ego rereading its own journal, whereas on this line the truth surfaces not in the evidence but in the slip and the transference, in the relation to another.
I don't dismiss the objection; I acknowledge it as a boundary. intlog doesn't reach into the unconscious and decodes nothing; it works one floor up, at the layer of revealed preference, and promises not the "real you," only a compass. And where the needle swings into the red (recall the "relationships" row) the tool itself tells you to rent someone else's resolution: a mentor, a therapist. Psychoanalysis here doesn't refute the journal; it marks where the journal ends and the couch begins.
intlog: an intention journal §
I put all of this into intlog, a small local utility: one SQLite file, no cloud. Each intention has exactly two touches: an entry with a forecast, and a blind reconciliation on the appointed day, where you answer before you see your own prediction.
The first touch, the entry, takes two or three minutes. First the tool asks: would it upset you to be proven wrong in a month? If not, nothing is written; the journal is worth keeping only while it holds what matters. Then come the domain, what you're deciding, the options you actually weighed, what you chose, one to five tags for what tipped the choice, and above all the affective forecast: what feeling you expect, how strong, how long, and why. Reconciliation dates snap to Sunday, so everything ripe waits for you on one morning, inside a single ritual.
The second touch, the reconciliation, is blind. On the appointed day the journal returns your intention but hides the forecast. First you answer: what you actually felt; whether events unfolded as expected (this separates a surprise from the world from a miss in self-knowledge); whether you'd choose the same again; whether the wanting is still yours. Only then does it reveal: forecast beside fact, and one question — did the feeling you were expecting arrive?
An intention journal is about the most intimate data there is: expectations about work, relationships, the body. Hence a local file. There is no reason to lie to it.
the limits of accuracy
At reconciliation you compare the forecast not with the experience itself but with your memory of it; both sides of the comparison are subjective, and memory reconstructs too. The journal measures the consistency of your expectations with your own later assessment. That's more modest than "objective accuracy," but it's the best signal available without sensors on your body.
And it's enough: even if the statistics never become significant, the ritual itself (name the options, name the motives, give a forecast, come back blind) already changes how you decide. Decision journals were recommended by students of the mind long before any AI; all that's been added to the old practice here is the blindness and the inevitability of return.
A Year In §
A year in: thirty to sixty closed entries (the n column in the table).
domain n closed grounded dimens. author no-regret world
money 14 0.93 0.86 0.62 0.90 0.86 0.90
work 16 1.00 0.55 0.60 0.90 0.75 0.80
relationships 9 0.69 0.41 0.30 1.00 0.44 0.90
body 4 0.57 — — — — — · low data
how to read the columns
All the fractions run from 0 to 1; higher is usually better, except n (a count) and world (a reliability context, not a score of you). Everything is computed per domain, because in different areas you predict yourself differently.
- n — how many closed entries accumulated. It's the weight of the row, not a grade: at small n the numbers are noisy and the row dims with a "low data" mark.
- closed — the share of reconciliations completed: closed ÷ (closed + overdue). How many ripe reconciliations you actually went through; we tend to skip precisely the ones we're afraid to answer.
- grounded — the accuracy of the affective forecast: a feeling match × closeness of intensity, averaged over the domain, but only over entries where the world didn't surprise you.
- dimens. — dimensionality: how many motives you even had an opinion about (the count of distinct tags at entry, out of five).
- author — the share of entries where, at blind reconciliation, you answered "yes, this is still my wanting." Reflection, not a check against outcome.
- no-regret — the share of "I'd choose the same again." Also self-report, and tinted by outcome: an unlucky result drags it down even when the decision was right.
- world — the share of entries where events unfolded roughly as you expected. A low "world" is about the unpredictability of the environment, not about you; it tells you how far to trust "grounded" in that domain.
A dash "—" means "nothing to compute": there are still too few closed entries with the needed fields (as in the "body" row).
Money: forecasts accurate, motives many-sided. This judgment you can trust fast; here you can delegate execution freely and spare yourself the double-checks.
Work: a steady bias. Time and again, "the new project will light me up, a five, and for months," and in fact a three, spent within two weeks. One miss is chance; the twelfth is a trait.
Relationships: decisions on a single motive, accuracy below a coin flip. The red zone of delegation: automate nothing here. On the contrary, learn here, rent someone else's resolution (a mentor, a therapist, an old friend), and train dimensionality: force yourself to name the third and fourth motive when your tongue reaches for "on a hunch."
Body: barely more than half the reconciliations closed. This row isn't about the body. It's about avoidance, and it may be the truest needle on the whole compass.
This is the solid ground, and notice what it's made of: not exact numbers (they're noisy) but a map — where to trust your wanting, where to train it, where to guard it from capture.
The models will get smarter every six months, in execution. Your work is no longer there. It's in the one place the ladder of delegation can't reach.
Cautious Optimism §
"Know thyself" was carved at Delphi twenty-six centuries ago. Two things are new: not knowing yourself has become expensive (multiplied now by the power of the executors), and the task has finally acquired a mundane, unembarrassing tool.
A notebook and three rules will do: write down significant decisions with a forecast of feeling; come back on time and answer before you peek; reread once a season.
I don't know where the models will be in five years; no one knows that right now, and anyone who claims otherwise is selling a course. But I do know where I can be in those five years: a little more accurate about what I want.
By today's standards that's an almost indecently reliable investment. Clarity about your own wanting is the one asset that appreciates with every release, because the stronger the executors, the more a clear client is worth.
The old woman had no pause between her wishes; each new one arrived before she could ask herself about the last. We have that pause. And we have a journal, so as not to lose it.
The sea won't calm. But a compass works in a storm too.
In fact, a storm is exactly when you need one.