SLOP

Chapter Eleven

The Maker's Price

Attention is the rarest and purest form of generosity.

— Simone Weil, letter to Joë Bousquet, 1942

You know it is slop when making or engaging it. Most slop flows from automated pipelines with no maker at all, but there is also, usually, a human in the loop: the marketer shipping forty posts a week, the consultant delivering reports the client will only skim. They are not deceived about the work. Ask them in an honest hour and they will tell you what it is. The question is what the knowing does to them over time.

The first month, the discomfort is sharp. You generate the draft, patch the edges, and ship it. You feel the wrongness in a located way: your eye catches the paragraph that only performs arguing without driving the point home. You tell yourself it is temporary. By the sixth month, something has changed. You read the day’s output and find it fine. The flat note is still flat, but the wince is gone. You ship without the pause. Then you stop noticing the absence of the pause. A year in, you can no longer say with confidence whether the piece is good. The faculty that used to report this has nobody in it returns no signal. Discrimination is a practiced capacity. When you practice something else, discrimination is displaced.

A violinist who stops playing does not keep her ear and lose her fingers. Both go. Her sense of pitch is an activity refreshed by use. A quality standard works the same way. Your sense of when a paragraph is done was built by years of suffering the gap between what you meant and what you made. The suffering was the practice. Every collision with your own inadequate draft was a calibration event. Delegate the drafting and the collisions stop. The standard fades invisibly, because the faculty that would detect the fading is the faculty that is fading.

The skill of making is not the thing most at risk. A coder who ships software she cannot explain line by line is doing something real: she specified what to build, tested whether users are satisfied, and answers when they are not. Her stakes are genuine even if her fluency in the internals is thin. You can be staked on something you do not fully understand, and people are, constantly. The doctor who orders a test she could not personally perform is still answerable for the order. The architect who cannot pour concrete is still the one the building falls on. What atrophies, when it does, is the ear: the judgment that knows when the output has crossed from adequate to good, when the draft is slop and when it sings, when a fix made things better and when it only made them quieter. The production skill survives the change. The ear is the thing at risk, and the ear is what the maker has to keep, because production has been repriced down toward free while discernment holds its old value.[1]

Rick Rubin has produced some of the most consequential records of the last forty years: Johnny Cash, the Beastie Boys, Slayer. He says, often, that he cannot play an instrument or run a mixing board.[2] What he has is an ear: a merciless sense of which take is alive and which is merely correct. He generates nothing. He decides everything. He is paid like a star for it, in a field crowded with people who can play and short on people who can tell.

The machine can now run review loops: ensemble critics, preference rankings across a thousand outputs, adversarial agents disputing the draft. Often useful. But stakes are what separate a real preference from a ranked guess. What Rubin sells is being answerable for the preference, well beyond the act of preferring itself. His name on the record is a liability, a bet that his ear was right, a thing that can be proven wrong in public. When a record fails, the failure has an address. When a review loop’s preferred output fails, it has none. The machine cannot pay the cost of being wrong, which means, though it can offer a shortlist, it cannot place the bet. The preference, the one that costs something, still has to come from a person.

You can use the shortlist and still be the one vouching for the pick. What you cannot do, without cost, is stop vouching and start forwarding. The approver who never fires a veto does more than grow less careful; she shifts from making a judgment to passing one along. Forwarding carries no stake. It moves the liability downstream, toward whoever acts on it, toward someone who did not sign the work. Done long enough, it moves the liability toward no one, which is the definition of slop.

What atrophies in the approver role is the habit of taking the work seriously enough to be embarrassed by it. The ability to generate, or to assess quickly, survives intact; the embarrassment is what goes. The approver’s bar is adequacy: is anything wrong with this? Adequacy is exactly what the machine supplies. Fed a diet of nothing-wrong, the veto never fires, and the faculty behind it begins the violinist’s fade: quietly, by degrees, steadily, in the direction of someone who has forgotten the question to be asked.

Whether to use the tools is settled. How to stay staked while using them is not. Staying staked does not require understanding every line of code you ship, or originating every paragraph you publish. Those requirements would condemn most of the actual practice of making things, and they should be refused on that basis, because the argument only bites where judgment gets staked: where someone’s name goes on a claim that will be tested against reality. Most making lives outside that line and always has, and the concern leaves it alone. The question is narrow: do you still have the collision?

The collision is the moment when the gap between what you meant and what the output does becomes undeniable. When you read it and something catches. When you find yourself patching a paragraph that is performing argument without making any. When the doctor reviewing the chart notices the dosage is technically within range but wrong for this patient. When the teacher reading the essay realizes the student has answered a different question than the one asked. When the lawyer reviewing the brief sees that the case cited says the opposite of what the brief claims it says. When you look at the code and realize the fix covered the symptom rather than the problem. The collision is what keeps the ear tuned. Stay in enough collision and you remain the person equipped to be surprised by failure, which is to say, the person equipped to answer for it.

The coder who ships what she cannot fully explain is fine, provided she is still the one who hears from users when it breaks, who decides what counts as fixed, who would be embarrassed if it failed in a way she should have caught. She needs enough to recognize the difference between working and working for now, and that bar sits well below knowing every line. The author who drafts with AI tools is fine, provided she still reads the result hard enough to catch the paragraph that performs thinking without doing any, and cuts it. The consultant is fine, provided he reads the report before it goes out rather than forwarding what the tool produced. In each case what is being maintained is the capacity to be surprised and embarrassed by a failure, the thing itself rather than the labor around it, which is another way of describing what makes stakes possible.

What the tools took was the friction that used to force the collision automatically. The blank page forced it. The empty mixing board forced it. Remove the friction and the collision becomes a choice, and choices can be declined, quietly, until declining is the habit.

The maker who keeps nothing at risk makes slop, and then goes further, becoming someone to whom the difference stops occurring.

None of this is a counsel of avoidance. The same tool that lets the approver fade lets the maker reach further than any maker before her, and the whole distance between those two fates is the grip you take on the thing.

Cheap intelligence does extraordinary things for someone who keeps her stake. The novelist who still writes her own sentences can run a research department out of a chat window, interrogate her own plot for the hole she cannot see, draft the scene she has been avoiding badly enough to learn why she was avoiding it. The engineer who still owns the architecture builds in a weekend what used to take a quarter. The founder who still answers for the company finds her weakest assumption faster than any advisor would have dared to name it. In each case the machine clears the ground around the judgment rather than making it, so that the scarce human act, the deciding and the standing behind it, is most of what is left to do. Used this way the tool concentrates the maker, distilling her down to the part only she can supply.

The macro version is among the most hopeful things now happening anywhere. In the sciences, where every result is staked to a named researcher and checked by a hostile field, these systems are compressing work that used to consume whole careers. Protein structures that resisted explanatory biology for half a century now resolve in an afternoon, and the search for medicines and materials is collapsing from decades toward months.[3] That is intelligence scaling against humanity’s hardest problems, in exactly the place where the stakes run highest and the checking is most ruthless. The accountability is what makes the acceleration trustworthy. The tool is most powerful where someone still answers for the result.

There is a reason the breakthroughs cluster where they do, and it marks the edge of the machine’s reach. A folded protein and a won game are closed problems: the world hands back a verdict, correct or not, won or lost, and a system handed that signal can teach itself past every human who ever tried, the way one already did at Go, playing itself millions of times until the final score had taught it which moves were right. Where a ground truth checks the work, the machine can bootstrap, and that whole zone is going to fall if it has not already. What resists is the open problem, the one with no terminal score, where consistency is no proxy for truth and a system can be flawlessly self-consistent and coherently wrong about everything that matters: the novel that has to land on a particular reader, the verdict on what is worth wanting, the call no scoreboard settles. Those open domains are exactly the ground this book has been calling testimony, where a claim has to come to rest on a particular mortal and someone has to stand behind it. The machine devours the closed game and stalls at the open one, and the part it stalls on is the part that was always yours to answer for.[4]

So the instrument is still an instrument, and only the grip changes. Hand it your judgment and it will wear away the thing that made you worth consulting. Keep your judgment and hand it everything else, and it becomes the largest lever a staked person has ever held.[5] The machine amplifies whoever you already are. It only makes the becoming faster.

Notes (5)

  1. The ear is a case of what Michael Polanyi called tacit knowing: we know more than we can tell, and an expert’s discrimination, the sense that something is off before she can say why, is held in a body of practice rather than in any rule she could write down (Personal Knowledge, 1958; The Tacit Dimension, 1966). It is built the way Anders Ericsson argued all genuine expertise is built, through long feedback-driven practice against resistance, not through exposure to finished answers. This is why it does not transfer with the output: the artifact copies, the discrimination that judges the artifact does not, and a worker handed the artifacts while spared the practice inherits the product and not the ear. The mechanics of that formation, and the precise way shortcuts dissolve it, are the next chapter’s subject. ↩︎

  2. Rick Rubin, co-founder of Def Jam. Britannica; Rock & Roll Hall of Fame. ↩︎

  3. DeepMind’s AlphaFold predicted the structures of nearly all catalogued proteins (over 200 million) by 2022, work recognized with the 2024 Nobel Prize in Chemistry. Claims of a broader acceleration in drug and materials discovery are real but uneven and still early; the pace varies sharply by field, and physical bottlenecks (synthesis, trials, validation) remain. For the optimistic case stated in full, see Dario Amodei, “Machines of Loving Grace” (October 2024). ↩︎

  4. The closed/open distinction runs quietly under the whole argument. A closed system has a win condition external to anyone’s opinion (the protein folds as it folds; the game is won or lost; the proof checks or fails), and that external verdict is what lets a system improve itself without us, because there consistency really is a proxy for correctness. Open systems have no such terminal signal: the worth of a novel is not its internal consistency but whether it tells a particular reader something true about being alive, and there is no coherence-check for landing on a person, because the target is not logical but human. Geoffrey Hinton, who expects these systems to surpass us at nearly everything, grants the seam almost in passing: certainly they will get more intelligent than us, he allows, but to do things genuinely meaningful for us “they have to have experiences quite like our experiences,” and in the same breath he notes why they do not, because “they’re not subject to death in the same way we are” (StarTalk, 2026). To become the kind of thing that can be exposed and can lose is the one move that scaling a formal system does not perform. ↩︎

  5. Chapter eighteen, the book’s turn from what the machine subtracts to what it frees, is given over to exactly this: the lever, and what a person who keeps her stake can reach with it once the no-stakes labor falls away. ↩︎