What the source published

The author opens with a simple claim: he loves building things, and writing software is his easiest way to conjure something useful from nothing. Two dumb little Python scrapers he hand-wrote this year gave him the most fun he had programming in years — then AI arrived. He won't pretend AI is dumb or unproductive (both claims have expired), but he retreats to a more honest position: AI coding is less fun. Vibecoding is absolutely thrilling at first; refactoring vibecoded software is not.

His core argument is that vibecoding frontloads the fun: an idea becomes a working prototype in minutes, with instant dopamine — even for people who have never written code. He splits builder satisfaction into six sources: the thrill of the idea, of making it real, of hard-won accomplishment, of a job well done, of learning, and of solving a problem you actually have. Vibecoding naturally delivers sources 1, 2 and 6 — but 3, 4 and 5 don't follow on their own.

His tooling spectrum is wide: local models via llama-server, copy-pasting from free no-signup services like Phind and Lumo for one-shot scripts, all the way to a remote “Clod Masterpiece 200™” all-you-can-eat plan with custom auto-approval hooks worryingly close to --dangerously-skip-permissions, used for algorithmic trading models. The verdict was always the same: productive, but it made him feel dumb and superfluous, and he disliked the output quality.

In fairness, he lists five things he would never have built otherwise: a one-shot has-this-been-submitted-to-HN bookmarklet (inspired by Taylor Troesh) that he uses daily on his Wallabag page; “Party Market”, a prediction-market party game that turned a dry lecture into grade-A entertainment despite garbage code; a Pyronear-based wildfire early-warning system vibecoded in two afternoons on free Lumo, detecting the first frame of smoke on a live feed from his laptop; a browser-based Kobo database migration assistant for people who don't know SQLite; and the realtime pitch-training app he had planned for a decade, finally built the day his singing lesson fell through.

Sources: [1] [2]

Try it yourself: six steps to put AI back in its place

  1. Start from something you genuinely need — like the author's two scrapers. The fuel of joy is usefulness, not coolness: pick a small problem that has annoyed you for a long time.
  2. Set a “hand-write first” rule for yourself: write v1 by hand, however clumsy or slow, and savour the whole journey from zero to working. It is the only source of joys 3, 4 and 5.
  3. Confine AI to the “not worth weeks, worth minutes” quadrant: one-shot utilities, party games, emergency prototypes, decade-old wishlists. These are “do it with AI or not at all” choices, not “AI versus hand-writing” choices.
  4. Never merge AI-generated code you can't explain. As the author puts it, vibecoding is like running a barely-controlled company — whatever grows inside the black box will bite back unless you refactor it yourself.
  5. Keep a weekly block of “no-AI time”: switch off completions, hand-write something, and revisit Type II fun — miserable in the moment, glorious afterwards, and easier the next time.
  6. Score your projects against the six joys: if a project only gave you 1, 2 and 6, AI took the nourishing parts — take the keyboard back.

Sources: [1]

Deep dive: the cost of frontloaded fun

The author drew a thrill/fun/learning curve: the vibecoding curve spikes hard at the start — near-zero friction from idea to prototype — then flattens or sinks, because maintenance and refactoring bring no learning, no conquest and no accomplishment. The hand-coding curve starts flat or even painful, but every “I did it!” pushes it higher; over the long run it sits higher and steadier. The graph is the whole essay in one picture: fun consumed upfront leaves only debt behind.

The author's hand-drawn thrill/fun/learning curve: vibecoding spikes early then flattens, hand-coding starts flat but runs higher and steadier long-term
The author's “thrill curve”: vibecoding frontloads the fun while hand-coding runs higher and steadier over time (image: The Autodidacts)

Of the six joys, the three vibecoding misses are exactly the nourishing ones. Hard-won accomplishment came from fighting the Rust compiler over twenty lines of code and grinding through a first Go project against a deadline — certified Not Fun™ at the time, glorious “I did it!” afterwards. A job well done requires owning every line; AI hands you things “fully formed, but not quite how you imagined” — when they break, all you can do is pour more money into the black box and yell louder. And the thrill of learning? There is now a layer between you and reality.

The five defense cases actually draw a clean dividing line: whatever was worth weeks of hand-writing, he hand-wrote; whatever was only worth minutes, he vibed. The HN bookmarklet, Party Market, the wildfire system, the Kobo migrator and the pitch trainer share one trait: without AI they simply wouldn't exist. This isn't an efficiency question but an existence question — AI dropped the cost of “doing” to zero, so a class of otherwise-never-built things got built.

The two killer metaphors deserve remembering: vibecoding is “a series of first dates” — dopamine in bulk, but you never develop oxytocin-style attachment to the project, and thrills go dull with repetition; and it is “being an executive of a barely-controlled company” — never feeling you really know what's going on, always suspecting sabotage and loafing. His closing jab: this essay was outlined in blue ink by candlelight, and AI wasn't even allowed to fix the typos. Savour them.

Sources: [1] [2]

How to read it

For AI subscribers, this isn't an anti-AI sermon — it's a smarter way to spend a subscription: burn quota in AI's burst zone (one-shot prototypes, party tools, emergency widgets, decade-old wishlists) and keep the nourishing zone for yourself (long-term projects, core code, refactoring). A subscription buys optionality, not a license to stop typing.

Deeper still, it's a reminder that joy is productivity fuel. The irrational pleasure of two dumb scrapers was the peak of the author's programming years — and it came from hand-won conquest. Vibecoding gives you joys 1, 2 and 6 and takes 3, 4 and 5; live on 1, 2 and 6 alone and dopamine keeps raising the bar until even the opening thrill feels dull. Keep a little clumsy hand-typing around — it's maintenance for your career.

Sources: [1]

Key takeaways

  • The claim: AI coding isn't dumb or unproductive, but less fun — it frontloads joy, delivers a prototype in minutes, and quietly takes the hard-won, well-crafted, learned-something joys.
  • Six builder joys: the thrill of the idea, of making it real, of hard-won accomplishment, of a job well done, of learning, and of solving your own problem. Vibecoding only naturally delivers 1, 2 and 6.
  • The dividing line: hand-write what's worth weeks; vibe what's only worth minutes and “wouldn't exist otherwise” — the HN bookmarklet, Party Market, wildfire warning, Kobo migrator and pitch trainer all fall in the latter camp.
  • The warning: vibecoding is like running a barely-controlled company, like a string of first dates — plenty of dopamine, no attachment; the bar keeps rising until the thrill goes dull.
  • The move: keep weekly no-AI hand-typing time, score projects against the six joys, spend subscription quota in AI's burst zone and keep the nourishing parts for yourself.

Sources: [1]