The Leveling

On February 2, 2025, Andrej Karpathy, a founding member of OpenAI and former head of AI at Tesla, wrote about a shift in the nature of work. He called it vibe coding. “You fully give in to the vibes,” he wrote, “embrace exponentials, and forget that the code even exists.”^[1] He meant building software by talking to a machine in English, taking what it gave back, and shipping it without necessarily reading it. Within months, the phrase attached itself to a much larger phenomenon: hundreds of thousands of people shipping apps, songs, and essays by describing what they wanted to a machine.

Watch what happened to the gate.

Learning to program, or paint, or write well once took years. Those years acted as a filter. They selected for patience, for a tolerance of failure, and for the time required to acquire craft. For the newcomers, that decade collapsed into a sentence typed into a prompt. Domain by domain, the distance between having an idea and shipping a working version of it fell to zero.

This is democratizing. For all of history, talent sat stranded on the wrong side of a skill gate it could not afford to cross. The woman with a novel in her but no grammar she could trust; the boy with a film in his eyes but no way onto a screen. The machines lowered the gate for all of them. The supporters who celebrate this are right.

The figure for it is not the robot but the genie. Anything that can be described can now be summoned. For all of prior experience, the constraint on making was execution. Wanting was free; making was ruinous. That constraint inverted in three years. Execution now rounds toward free. The entire weight of the enterprise has landed on the wish.

Folklore was always engineering literature. The stories of genies and monkey’s paws turn on the failure of the wisher; the magic always works. The genie delivers what is asked for, and the catastrophe lives in the gap between the stated want and the actual want. Midas got exactly what he requested. When execution is free, the “wish” (the vision, the taste, the recognition of what is right) becomes the only bottleneck. Wishing well is the whole game, and it is the one part the lamp does not supply.

---

The demos are real, but for a while they hid what Addy Osmani named the “Seventy-Percent Problem.”^[2] The machine takes you to seventy percent of a working product almost magically; the last thirty (the security, the edge cases, the integration with reality) stayed, for a time, as hard as it ever was. It is tempting to build the whole argument on that gap, to say the machine does the easy part and the human keeps the hard part, and rest there. Resist it. The gap is real and the gap is closing, and a thesis pinned to a closing gap closes with it: the edge cases will fall, the security will fall, the integration will fall, and each thing the machine could not do this year it may well do the next. The durable point lives one level down, where the percentages cannot reach it. What the hard thirty percent demanded was never just more difficult labor. It was judgment about the ways a thing can fail, and judgment is the visible form of something the model does not have and cannot be handed: a stake in whether the thing actually works. The model produces what is plausible given the prompt, and the prompt is usually under-specified, with no standing goal that would make the model hunt for the scenario no one named. Agentic systems increasingly generate their own edge cases and tests, so this is not a permanent ceiling on the machine. The gap that does not close is the one about answerability: when the scenario nobody considered finally arrives, no one is answerable for having failed to consider it, because no cost lands on the thing that produced the work. Producing something that satisfies the demo is easy and getting easier. Being the one who answers when the demo meets the world is the part that does not transfer, however much of the work the machine absorbs.

The seventy percent the machine does for free is what the apprenticeship used to spend years drilling. But the apprenticeship was never only a delivery system for hard skills; it was how a person became someone who could be answerable for the work, by failing at tractable problems under conditions where the failure taught something. The gate has been pulled away from the production of the artifact and left standing, invisible, in front of that formation. The newcomer clears the artifact at full speed and hits the closed half of the gate without knowing it is there, holding something that works in the demo, with no formed judgment beneath it and no way to answer when it fails in the world.

You can read that collision in the security logs. In early 2026, Moltbook, a social network built for AI agents, launched to a rush of attention and was breached almost at once: researchers found roughly 1.5 million API keys lying in the open because the most elementary database protection, row-level security, had never been switched on, and they reported it inside a day.^[3] Its founder had boasted that he “didn’t write a single line of code” for it; he had “just had a vision for the technical architecture, and AI made it a reality.” The boast was the diagnosis. He could see exactly as far as he had built and no further, so what the generated code silently failed to do stayed invisible to the one person nominally answerable for it. The machine hung the doors and left the locks unset, and the only human in the building could not read the blueprint.

Moltbook is the texture of the slop era. The app demos perfectly; it is a masterpiece of the first seventy percent. Then a real user or a malicious attacker leans on the part that was never built.

---

The popular story, the field is leveling, runs inside-out. A leveling field would lift the people at the bottom. This one clears them off it.

Between 2022 and 2024, entry-level software postings fell by sixty percent. Junior roles fell faster than senior ones, and employment for older engineers, the ones who can answer for what the machine ships, held up better.^[4] Treat those figures with care: the same window saw the end of the zero-interest-rate era and the 2022 to 2024 tech layoffs, and untangling how much of the junior collapse is AI from how much is cheap money drying up is beyond what the data can settle. The numbers are suggestive, not dispositive. So lean the weight elsewhere. The machine substitutes for the junior and augments the senior, and whatever the precise share, the gate fell and landed on the people standing on the lowest rung, crushing the entrance it was supposed to open. What got leveled was the floor of the profession. The argument does not need the postings figure to hold. It runs on the shape of the pipeline.

Seniors are grown from juniors. A profession that stops hiring its bottom rung is arranging not to have a top rung in fifteen years. The judgment a senior brings to the hard thirty percent was formed by years of failing at tractable problems under conditions where the failure was instructive, the slow way, well past anything coursework installs. The junior who debugs a trivial race condition at two in the morning builds the spatial intuition the senior uses to spot a critical one in three seconds. That is the mechanism of learning. Medicine solved the “apprenticeship is expensive” problem by creating supervised environments where juniors could make real errors at contained cost, keeping the junior work and capping its blast radius. The profession compounds. Cut the entry point and you cut the supply chain. The judges are made out of apprentices, and the apprenticeships are being automated away.

The practitioners who thrive have a finer word for it. Simon Willison distinguishes vibe coding (rolling the dice on a model without reading the output) from vibe engineering, where the model writes the lines but a human reviews, tests, and understands every one.^[5] The vibe coder is the “approver,” who takes what arrives and walks away clean when it breaks. The vibe engineer is the “maker,” who stays answerable for every line.

The gate moved from the writing, which is now free, to the reading—from producing the thing to being able to tell whether the thing is right. This is liberation for the person who always had the eye and was kept out by the hands. It is hard news for the person who only has the hands. And it is a trap for the one who has neither, because the tools sell the most expensive illusion of the age: that you get to skip the becoming.

The question of the era is no longer can you make this? Anyone can make this. What is left is the older question: when it breaks, who is standing there, and what did it cost them to become someone the answer could land on?

Notes (5)

1. Andrej Karpathy, X, February 2, 2025. ↩︎

2. Addy Osmani, “AI’s 70% Problem,” Zed Blog (2025). ↩︎

3. Wiz Research, “Hacking Moltbook: AI Social Network Reveals Millions of API Keys,” wiz.io, 2026; corroborated by Engadget and SiliconANGLE (February 2026). The founder is Matt Schlicht (@mattprd, CEO of Octane AI); the quoted boast is from his own post on the platform. The exposure was reported and patched within roughly a day of launch. ↩︎

4. Stanford analysis, 2025; Stack Overflow (Dec. 2025). ↩︎

5. Simon Willison, “Vibe engineering,” simonwillison.net (Oct. 7, 2025). ↩︎