The Death of Syntax Was the Wrong Funeral

The first time I noticed it, I was writing a query I had written a hundred times before — pull the active accounts, join them to their most recent invoice, filter out the ones already flagged. There is a shape this query has. Everyone who has written it arrives at roughly the same shape, because the shape is dictated by the data and the engine, not by the person. You do not get points for an unusual SELECT. If you reach for something unusual, you are usually about to make the database scan a table it should have indexed, and the query will be slow, and someone will quietly rewrite it. SQL felt, for a long time, like the one corner of programming where the romance had already been wrung out. There was the correct query and the incorrect query, the fast one and the slow one, and no third axis along which a person could be interesting.

The argument I want to examine takes that observation and lets it metastasize. It goes like this: most programming is just SQL that hasn't admitted it yet. The clever algorithm, the elegant formatting, the idiosyncratic way you structure a module — these were always decoration on top of a function that had a correct form. Now that a machine can produce the correct form on demand, the decoration is exposed as what it always was, which is decoration. And — here the argument makes its real move — the same is true of prose. The machine has perfected the essay the way the optimizer perfected the query. There is no longer a creative way to write inside the medium, because the medium has a correct output and the machine produces it. To be creative, then, you cannot write a better essay. You have to refuse the essay entirely and do something that does not yet have a perfected form, something the machine has not yet learned to converge on.

It is a clean argument. It has the particular seductiveness of arguments that flatter your sense of loss — it tells you that the thing you mourn was real and that its death was inevitable and structural, which is more comfortable than the possibility that you simply got worse at it or that it was never quite what you thought. I have held this argument myself. I want to take it apart carefully, because I think it is wrong in an interesting way, and the way it is wrong tells you something about what is actually happening.

Start with the premise, because the premise is load-bearing and it is only half true — and the false half is the half doing the work. Take the narrow claim first and grant it. For a single query against a fixed schema, there usually is a most-performant form, and reaching for an unusual one usually means you are about to be wrong. At that level the original observation holds. There is no expressive freedom in a SELECT for the same reason there is none in long division. Conceding this is not a problem for me; it is a problem for the argument, which needs the concession to mean more than it does.

What the premise quietly does is slide from there is a best form to there is no craft, and those are different claims. Finding the best form is not the absence of skill — it is the entire content of the skill. The same rows can be returned by a query that lets the engine use its indexes and one that makes it abandon them and scan the whole table: an OR across two columns, against a UNION of two indexed branches, will often produce the same result set at wildly different cost. Which you reach for is a judgment about how this engine, with this data distribution and these indexes, will actually behave, and the judgment is wrong often enough that an entire discipline exists to house the people who are good at it. The optimizer that makes SQL feel automatic is not the absence of choice. It is a layer that hides the choices and makes its own, and it is beatable, and beating it has no other name but craft.

There is a distinction the original argument elides, and eliding it back would be cheating. There is craft, which is convergent: many people working hard arrive at roughly the same good answer, because the problem has a direction. And there is art, which is divergent: the value is in the spread, in two excellent people producing different things where the difference is the point. Single-query optimization is craft, not art. So proving that SQL demands skill does not, by itself, prove that SQL permits expression — the most-performant query is in fact the least expressive one. On this the author is right: the surface of SQL has had its art wrung out, and no amount of affection brings it back. He is wrong about what that implies, because the art in SQL was never on the surface. It lives one layer up — in schema design, in what to index and what to denormalize, in whether to precompute, in how to model a messy domain as tables at all. Those decisions optimize against an unknown future workload and several objectives at once: read speed against write speed against storage against the next engineer's ability to understand any of it. They genuinely have no single right answer. That is divergent. That is where two good database people part ways. The surface converged; the design never did.

Now the second move, the one that smuggles in the conclusion. The argument says the machine has perfected the essay. It has not. It has done something quite different, and the difference is the entire question. The machine produces, very cheaply, the median essay — the statistically central response to a prompt, the most probable next sentence given the previous one. "Perfected" and "regressed to the mean" are not two descriptions of the same event. They are opposite claims. A perfected medium would be one with no ceiling above it, where the best possible output had been reached and further effort was wasted. A homogenized medium is the reverse: the floor has been raised to the middle, and the ceiling is not only still there but now stands in a much emptier room. When you say the machine perfected the essay, you are quietly agreeing to measure the essay by its floor. The moment you measure it by its ceiling, the claim inverts. The machine did not climb to the top of the medium. It made the middle free, which is a different and in some ways more disruptive thing, but it is not the death of the top.

You can watch this confusion produce a falsifiable prediction and then watch the prediction fail. If the machine had perfected the essay, then essays converging on the machine's output should do at least as well as essays that don't. The opposite is what the data shows. A study out of Cornell and Carnegie Mellon, published in 2026, looked at roughly eighty-one thousand college admissions essays from 2020 through 2024 and found that after ChatGPT's release the essays began converging in style — and that higher estimated AI use was associated with worse predicted admissions outcomes, an effect concentrated among lower-income applicants who had the fewest other writing supports to fall back on. If the median were the perfected form, converging on it would be safe. Instead, converging on it is penalized. The market for these essays is already repricing the mean downward. That is not what the death of a medium looks like. That is what it looks like when a medium's floor becomes worthless and its ceiling becomes the only thing anyone is paying for.

This is the point at which the original argument tries to escape forward. Fine, it says — if you can't win inside the essay, then creativity moves to the space that isn't perfected yet. Don't write the essay; do the thing that doesn't fit the genre. And there is something true in this. Genres do exhaust, forms do calcify, and there is real work in finding the shape that hasn't yet hardened. But examine what has happened to the definition of creativity in the course of that sentence. Creativity has been redefined as novelty of form — as occupying whatever territory the machine has not yet mapped. Under that definition creativity is by construction a fugitive. It is always running one step ahead of convergence, and the moment the machine catches up to any form, that form is dead and the creative person must flee again. This is a regress. If creativity is only ever the not-yet-perfected space, then it has no home, no content of its own, nothing it is about — it is purely defined by what the machine cannot yet do. That is the same category error as before, wearing different clothes. The first version mistook the simplicity of SQL's surface for the absence of craft. This version mistakes the familiarity of the essay's form for the exhaustion of the thinking the essay can carry. But the essay is not the thing being made. The thought is the thing being made; the essay is its instrument. A machine that produces the median arrangement of sentences has not exhausted the space of things worth thinking any more than a machine that produces the median chord progression has exhausted music. It has made the median cheap. The ceiling is untouched because the ceiling was never about the form.

I want to be careful here, because the argument has an empirical wing and the empirical wing is its strongest part, and I do not want to win against the weak version. The strongest version points at the industry and says: look, it is already over. In June 2026 Anthropic reported that more than eighty percent of the code merged into its own production codebase was written by its model rather than by humans. The franchise has already won; the artisans are already gone; I am not predicting a future, I am describing a present.

That figure is real, and it is also a good case study in how a real figure gets conscripted into an argument it does not support. Anthropic's own report says lines of code are an imperfect measure — they count quantity, not quality — and that the accompanying claim of an eightfold productivity increase is "almost certainly an overstatement," with the company's own researchers putting their median self-estimated uplift closer to fourfold. More tellingly: the same CEO who now reports eighty percent had predicted, in March 2025, that AI would be writing ninety percent of all code within three to six months and essentially all of it within a year. Six months later that had plainly not happened across the industry, and the prediction quietly aged into a cautionary example. So the number that gets cited as proof the era is over is (a) a measure the people reporting it tell you not to trust, (b) drawn from the one company on earth most selected for this outcome, and (c) the surviving fragment of a forecast that was, on its own terms, wrong about timing by a wide margin. None of this means the number is fake. It means the number measures volume of generated text, not the transfer of judgment, and the argument needs the second thing while citing the first.

And then there is the finding that should trouble the thesis most, because it comes from the side the thesis assumes is winning. A randomized controlled trial run by METR in early 2025 took experienced open-source developers working in codebases they knew well and measured how fast they completed real tasks with and without AI tools. The developers expected the tools to make them about twenty-four percent faster. Afterward, they believed the tools had made them about twenty percent faster. In fact the tools made them nineteen percent slower. Hold the two numbers together: the felt experience was a large speedup; the measured reality was a slowdown. This is the most important thing in the whole discussion and it is almost never said plainly. The sensation of frictionlessness is not the same as the fact of progress. The machine removes the feeling of effort, and we have been trained to read the absence of effort as the presence of velocity, and in at least one careful measurement that reading was not merely imprecise but inverted. The study is small, it used early-2025 tools, it looked at mature repositories rather than greenfield work, and METR itself has since revised its experimental design — I am not claiming it proves AI slows everyone everywhere. I am claiming it punctures the confident half of the thesis, the half that says the senior is being exponentially amplified. In the one place someone measured instead of asked, the amplification ran backward, and the people inside it could not feel that it had.

What survives all this is narrower than the thesis and, I think, truer. The apprenticeship claim survives: it is probably correct that the junior developer's traditional grunt work was load-bearing, that struggling through the boilerplate was how the map of the system got built, and that automating the struggle breaks the path without anyone deciding to break it. The labor data is consistent with this — Stanford's "Canaries in the Coal Mine?" study found employment for software developers aged twenty-two to twenty-five down nearly twenty percent from its late-2022 peak, even as employment for developers over thirty-five in the same field grew. But notice that the mechanism Stanford proposes is that AI is good at exactly the textbook part — the syntax and the standard algorithms taught in school — and weaker at the tacit, situational knowledge that only accrues with years. That mechanism does not say the craft was decoration. It says the opposite. It says the part the machine took is the part that was always the floor, the memorizable surface, the SQL-that-looks-like-there's-only-one-way — and the part it cannot take is the judgment that the original argument dismissed as decoration in the first place.

Which returns the SQL analogy to its author, pointed the other way. What died was never the craft — it was the visibility of the floor as a place where effort was required. Typing the query felt like work, and the feeling of work felt like craft, and now the typing is free and the feeling is gone, and we have mistaken the disappearance of the feeling for the disappearance of the thing. But the craft was never the typing or the sentences, the layers with one right answer. It was always the judgment above them — what to store and what to index, which of the true things to say and in what order and what to leave out — the layer where more than one answer is defensible and the choosing between them is the work. The machine produces the median arrangement and calls it done. Choosing against the median — knowing that the probable next sentence is the wrong one and being able to say why — is not a fugitive hiding in the not-yet-perfected margins. It is the divergent judgment that sat one level above the surface all along, and it has not become less valuable. It has become most of what is left.