Vibe coding: the gate was syntax, and it is gone

For seventy years, the price of admission to software was one of two things: capital or syntax. If you had money you hired a team. If you had syntax you could build alone at two in the morning. Everyone else had ideas about tools they would never own.

That arrangement ended somewhere between February 2025, when Andrej Karpathy described "vibe coding" as giving in to the vibes and forgetting the code exists, and November 2025, when Collins made it Word of the Year. It is worth being precise about what actually opened, because the usual story — AI made coding easier — is the least interesting version of it.

Two gates, both of them gone

The first gate was money. Custom software was a capital purchase, so it belonged to organizations large enough to absorb a six-figure build and a twelve-month wait. The second gate was syntax: a few hundred thousand people on earth could make a machine do something new without permission.

Vibe coding did not lower the second gate. It deleted the requirement. The interface is no longer a language you must learn in order to state intent; the interface is the statement of intent.

The demographic is the story, not the volume. Roughly four in five people building on consumer prompt-to-app platforms are not developers. Almost half are founders. That is a new population of software producers, and it did not exist as a market 24 months ago.

The volume claim needs a caveat, because the honest data contradicts itself. Adaptavist surveyed 240 US and UK engineers and found 83.9% vibe code in some form. Stack Overflow surveyed over 49,000 developers and found 77% say it is not part of their professional workflow. Both numbers are probably true; they asked different populations different questions. Any adoption figure quoted without its definition is a marketing slide.

What actually changed is the bottleneck

When syntax stops being the constraint, the constraint moves to the two things syntax used to hide: specifying what you want, and knowing whether you got it.

That is why the output of this era looks the way it does. Creation has scaled. Verification has not.

Read the last line twice. As AI-assisted output rose, the proportion of work spent reorganising what already exists collapsed. Teams are now adding to codebases without maintaining them. Nothing breaks this quarter, and everything is slow next year.

An infinite supply of technical outcomes does not mean an infinite supply of good ones. It means the scarce input moved, from the ability to produce software to the ability to judge it.

Will we get smarter? Not automatically

This is the part of the conversation the industry avoids, and the research is not ambiguous about the shape of the risk.

The perception gap is the first problem. METR ran a randomised controlled trial with experienced open-source developers on their own codebases in early 2025. With AI tools they took 19% longer to finish issues. Before the trial they forecast a 24% speed-up; after it, they still believed AI had made them 20% faster. A 39-point gap between what happened and what they felt. The tool did not make them worse developers. It made them misinformed about themselves, which is worse, because miscalibration is what you build the next decision on.

The cognitive bill is the second. MIT Media Lab's "Your Brain on ChatGPT" study put 54 participants through three essay sessions wired to EEG, split across brain-only, search-engine and LLM groups. Brain-only showed the strongest and most distributed connectivity; LLM users showed the weakest. In a fourth session, participants who had used the LLM could not reliably quote essays they had written minutes earlier, and showed low ownership of their own text. The researchers called the accumulation "cognitive debt" — effort deferred, not avoided.

The variable is you, not the model. A Microsoft Research and Carnegie Mellon survey of 319 knowledge workers across 936 real task examples at CHI 2025 found the pattern plainly: higher confidence in AI predicted less critical thinking, even though it felt like less effort; higher confidence in yourself predicted more critical thinking, even though it felt like more effort. The tool does not decide whether you think. Your confidence level does.

So the answer to "AI will get smarter with time — will we?" is: not by default, and not as a species. A model is retrained every few months by someone else's decision. A person is retrained only by their own. That is the asymmetry nobody puts in the launch copy.

What actually compounds in a person now

Four capabilities were previously gated behind syntax and are now exposed, naked, in every prompt:

1. Specification. Stating what "done" means in terms a machine can be tested against, before you know how to build it.
2. Verification. Being able to tell whether the thing in front of you is wrong. This is why 96% of developers still distrust AI code, and why the 68% who call their own output flawed are the honest ones.
3. Systems thinking. Knowing what breaks at ten times the volume, six months after launch, when the database has thirty thousand rows and eleven people are editing at once.
4. Taste. Choosing correctly between nine answers that all pass the test.

Notice that none of these are programming languages. They always mattered; syntax was simply the only way to demonstrate them. The mid-level developer who had taste but wrote slowly now looks the same as the operator who ran a service department for fifteen years and can describe a workflow down to the minute. Both can produce software this quarter. The ones who cannot explain what they want precisely will keep buying it.

The 2027 read

The practical consequence for a dealer group, distributor or aftermarket brand is not "everyone becomes a developer." It is that the threshold at which a specific internal tool becomes worth building has collapsed. A service-lane scheduling sheet, a parts-lookup shortcut, a missed-call tracker, a compliance checklist your legal team actually accepts — these were never worth a six-figure project and are now worth a Tuesday afternoon and a review.

Which means three things change, and one does not:

• Buy versus build flips for anything narrow, internal and cheap to change.
• Review capacity becomes the limiting factor, in every organization, at every scale. Two people who can audit what a hundred people generate is not a bottleneck you can prompt your way out of.
• Ownership gets named. Every vibe-coded tool in production needs a human whose name is on it, because an agent will not answer for it.
• You cannot outsource judgment. The model generates; the organization still has to know what it wants and recognize when it got something else.

The honest answer

Human history is marked by tools that expanded what a person could do alone: writing, printing, the spreadsheet, the compiler. Vibe coding belongs on that list, not because it removes skill but because it relocates it. It has moved the test from the syntax you memorized to the clarity with which you can say what you mean, and the honesty with which you check that you got it.

The machines will keep getting smarter. Whether that compounds through us or around us is not a model decision. It is a cultural one, made in how much friction we are willing to keep.