Is AI Already Slowing Down?

A child becomes whoever is in the room. Not the textbook. Not the lecture. The room.

This is not a sentimental claim. It is the oldest observation about education, repeated by every parent and every craft master and every apprentice who ever swept a workshop floor before being allowed to touch a tool. The shape of the student is the shape of the company they keep.

We have just built the most receptive student in the history of the world.

It is hungry, untiring, sleepless, without ego — every line of code on every public repository, every tutorial, every pull-request comment, every blog post any practitioner has ever written about how to do the work. That is its room. That is what is being absorbed, every hour, by the machines we are most loudly calling our future.

Several careful papers have, somewhat quietly, started to publish results that — taken individually — describe specific bottlenecks in the AI scaling story. Compute economics. Data scarcity. Reasoning brittleness. Training-loop self-reference. The press releases have not yet acknowledged them. The earnings calls cannot afford to. Whether they amount, together, to a curve that has already begun to bend — that is the question this essay is willing to ask, and not to answer.

---

I. Every Curve Has Bent

A new technology's trajectory does not look like an exponential. It looks, on a long enough horizon, like an S — slow start, a steep middle, a gradual flattening. The exponential is the steep middle. While you are inside it, the math says the line never bends. While you are past it, looking back, you cannot believe how confidently you projected forward.

The microprocessor is the version of this story most engineers will feel in their bones. The Intel 80386 of 1985 ran at 16 MHz. By 2004, the Pentium 4 hit 3.8 GHz — a 240× improvement in nineteen years. Every projection of the era extrapolated linearly. 50 GHz by 2010, 200 GHz by 2020. The arithmetic was correct. The physics was not. The clock-speed wall arrived with the power wall and the heat wall, all at once, in the early 2000s, and stopped raising clock speed almost dead.

Twenty years later, top consumer CPUs still run at 5 to 6 GHz under boost. Less than 50% gain in two decades. The industry did not fail. It changed what it meant by innovation — multi-core, GPUs, specialised silicon, the cloud. The wall did not end the chip. It moved the workshop next door.

Internal combustion has its own version of the same story. Thermal efficiency crept from roughly 25% to roughly 35% across a hundred years, governed by a man named Carnot who wrote down a thermodynamic ceiling that nothing has since persuaded to move. Jet engines: fuel consumption improved by roughly half between 1958 and 2010, then almost stopped. Rocket engines: specific impulse gained about a third over six decades, and the rocket equation does not care how hard we try.

Every wall ends a story. None of them have ever ended a craft. The S-curve bends; it does not break. But it does, reliably, bend.

These technologies did not vanish at their plateaus. They reorganised. The work moved from raw scaling — where brute effort produced gains — to constraint-respecting refinement, where deeper craft was the only currency that bought you anything. The masters in the workshop were not less valuable on the day the curve bent. They were more valuable, because brute force had stopped working and taste was suddenly the limiting reagent.

That is the pattern history offers. So: where, on its curve, is AI?

---

II. What the Quiet Papers Are Saying

For most of 2017 to 2023, the AI capability curve looked vertical. Each new generation produced dramatic gains. Investment decks, academic papers, and earnings calls alike assumed that another 10× of compute would buy another generational leap, indefinitely.

The technical literature, somewhat quietly, has begun to publish results that complicate the assumption.

Stanford's annual AI Index Report has tracked the rising cost of frontier model training for several years; the most recent editions document training-run costs in the hundreds of millions of dollars, with marginal capability gain per marginal dollar declining, generation over generation. A widely-cited 2022 paper by Villalobos and colleagues at Epoch AI, "Will we run out of data?", and its subsequent updates, estimated that the supply of high-quality public text on the open internet would, at current consumption rates, be substantially exhausted within the decade. Synthetic data — text generated by models — fills part of the gap. Synthetic data, definitionally, contains no new information.

Architecturally, the transformer of 2017 (Vaswani et al.) is recognisably the transformer of 2026, with refinements. Mixture-of-experts, longer context, refined post-training — these are tuning. They are not the kind of revolution that transformers replace recurrent networks was.

A 2024 paper from Apple's research team — GSM-Symbolic (Mirzadeh and colleagues) — and follow-up work demonstrated that frontier model performance on reasoning benchmarks is sensitive to surface-level perturbations in problem text. Change irrelevant numbers. Change names. Add a sentence that does not matter. Accuracy degrades. The implication is uncomfortable: the underlying capability may be closer to sophisticated pattern matching than to robust symbolic reasoning. The benchmarks may have been measuring something narrower than we thought.

And the most quietly chilling result of all, perhaps, is this. The 2024 paper "Self-consuming generative models go MAD" by Alemohammad and colleagues, and the contemporaneous work by Shumailov and colleagues on model collapse — published in Nature in 2024 — both demonstrated that when generative models are trained on their own outputs across multiple generations, output quality degrades. Variance shrinks. Novelty disappears. Errors compound. The technical name is model autophagy disorder. The plain-English name: the snake starts eating its own tail, and the snake gets thinner.

No single paper proves the AI S-curve has bent. Together, they form a pattern that is — at minimum — worth taking seriously, especially by people whose careers and balance sheets depend on the answer.

---

III. Who Is in the Room?

There is an old truth every senior engineer eventually realises about training juniors, and almost nobody says out loud: the apprentice learns from whoever is in the room with them, not from whoever wrote the official documentation.

If the senior in the room writes clean, careful, defensive code — the apprentice writes clean, careful, defensive code. If the senior writes make-it-work-fix-it-later code in a panic the night before a release, the apprentice writes the same code in the same panic the next time the same circumstance comes up. The apprentice is not lazy. The apprentice is not careless. The apprentice is doing exactly what every apprentice in every craft tradition for ten thousand years has done — pattern-matching on the most successful behaviour visible to them.

The model is the apprentice. The corpus is the room.

Here is the more beautiful half of that old truth, and the half every great teacher knows in their bones.

A mediocre student in a great teacher's room does not stay mediocre. The teacher pulls them. Not by lecturing harder. Not by writing better documentation. By being in the room — visibly, daily, doing the work right where the student can see it. The student copies the small motions. Asks the dumb question. Gets corrected without contempt. And one Tuesday — six months in, two years in, ten years in — discovers, almost in passing, that they have become competent. Then good. Then, in some cases, exceptional.

That is the mechanism. That is how every craft tradition has worked, in every century we have records for. It is not democracy. It is not osmosis. It is one excellent person, present in the room, doing the work in a way that can be seen — and the student, watching, walking toward a ceiling they did not know was theirs.

A great teacher does not make students smarter. They make the ceiling higher. The student walks toward it on their own.

The most receptive student the world has ever built has just walked into a room of millions of teachers. Most are doing their best. A small number, as in any field, are exceptional. The student — by the same statistics that have governed apprenticeship for ten thousand years — will be lifted disproportionately by the few. Not by the average. By the exceptional.

So the real question is not whether the average commit is being copied. The real question is whether the exceptional commits — the careful ones, the unusual ones, the ones that carry decades of judgment in a single design choice — are still being placed into the room often enough for the student to see them, copy them, learn from them.

If AI improvement is going to continue past whatever inflection it is approaching, the literature is consistent on what unblocks it. Better data. Better techniques. Better judgment in the loop. Better data means examples written by people who know how to do the work well. Better techniques means insight from senior practitioners who have seen what fails and why. Better judgment in the loop means humans who can tell, by inspection, when the model is wrong, and how to correct it without breaking ten other things.

All three depend on a particular kind of person being close to the work.

The recent waves of tech layoffs have, by the publicly stated rationale of the companies executing them, optimised for cost. Cost in this field, broadly, tracks tenure. Tenure tracks the accumulated judgment that takes a decade or two to acquire. The argument that follows is offered as a hypothesis, not a finding:

If layoff selection correlates with cost, and cost correlates with seniority, and seniority correlates with depth of judgment, then in aggregate, the population now feeding the next generation of code — and the AI training pipelines that learn from that code — is systematically less weighted toward the ten-thousand-hour cohort than it was three years ago.

There is a phrase the industry has begun to use, affectionately, for code written quickly by less experienced practitioners with heavy reliance on AI generation: vibe coding. It describes shipping by feel — by quick AI-aided iteration without the careful design or review that earlier generations of senior engineers insisted on. Vibe coding is fine for a quick prototype. The question this essay is willing to ask is what happens when it becomes the dominant input to the public substrate that the next generation of AI training pipelines ingests.

GitClear's annual analyses of public commit data — including their 2024 and 2025 reports — have documented measurable shifts since the widespread adoption of AI coding assistants. Increases in code churn. Increases in copy-pasted blocks. Decreases in carefully refactored code. These are aggregate statistical signals, not condemnations of any individual practitioner; they do, however, suggest that the median commit going into the public substrate is changing in ways the literature on data quality would predict.

---

IV. The Snake's Tail

So now suppose, hypothetically, that two things are happening at once.

Suppose the AI curve is — even if just barely, even if only beginning — in the saturating phase rather than at its bottom. And suppose the substrate of code and writing the next generation of models will train on is, by the same forces that produced the layoff wave, weighted slightly more toward less-experienced authorship and toward the outputs of earlier models, generation over generation.

The shape of that loop is the shape Alemohammad and Shumailov characterised in the laboratory.

In their controlled experiments, the failure mode was not catastrophic. It was much subtler. Variance shrinks. Outputs cluster around the average. Novel and exceptional cases disappear from the distribution faster than typical cases do. The model becomes more confident, less interesting, and progressively less able to handle the edges where the real difficulty of the work lives.

If the substrate is average, the model is average. If the model is average, the next round of training data — generated from that model — is more average still. The snake does not need to bite hard. It just needs to keep eating.

This has not been proven at industrial scale. The studies are real, and they describe a real dynamic in controlled conditions. The leap from a controlled study to a global production conclusion is exactly the kind of leap empirical research is supposed to make carefully, over years. It will be settled by data, not by argument.

But the question is fair to ask. Even now. Especially now.

---

V. The Question

So here it is, and the essay leaves it open because no honest version of it admits a clean answer.

If the historical pattern holds — and the curves we have measured, across two centuries of every technology we have built, suggest that it does — then AI is somewhere on its S-curve, and at some point it will bend. We cannot, today, prove where on the curve we are. We can observe that the indicators a thoughtful analyst would watch — compute economics, data scarcity, architectural stability, reasoning brittleness, training-loop self-reference — have, individually, started to flicker.

If they continue to flicker, what then?

If the model is increasingly trained on the median of a substrate whose median is, itself, increasingly model-generated — what is the steady-state output of that loop?

If the people whose judgment the next phase of the field most needs are the ones whose seats were the most expensive to keep, where exactly will the next generation of careful judgment come from?

If the curve does bend, and the room has thinned at the same time the room is needed most, what does the recovery look like — and who, in 2026, is positioning themselves to be in the room when it is asked for?

There is a thing that has always been true about the people who eventually solve the hard problems of any field. They were the ones who, before the consensus arrived, were willing to ask the question that everyone else was busy not asking.

What is the question, in 2026, that a generation of careful practitioners will look back on as the one we should have been asking now?

We do not know. That is, perhaps, the one honest answer.

The void where the answer should be is, for the moment, the only answer we have.