Vibe coding feels like productivity for the same reason slot machines feel like winning

Rachel Thomas of fast.ai published an essay titled Breaking the spell of vibe coding that lays out the strongest empirical case against the dominant practice. The core argument: generating massive quantities of AI code intended to remain unread mimics gambling’s addictive dark flow state rather than genuine creative engagement.

This sounds like rhetoric until you read the supporting data. Thomas cites a METR study where developers self-estimated 20% speed gains from AI coding tools — and actually worked 19% slower when measured against control. A 40-percentage-point gap between perception and reality. The pattern is structurally identical to slot machines: misleading signals (code appearing productive while containing hidden bugs), false agency, and reward delivered on a schedule that mimics actual achievement.

Why the perception gap is so large

Thomas applies Mihaly Csikszentmihalyi’s flow framework: true flow requires matched skill-challenge and clear performance feedback. Vibe coding fails both criteria. The challenge floats freely (the AI does the hard part, so what is the developer’s challenge?). The feedback is misleading (the code compiles, looks reasonable, but contains bugs the developer didn’t write and can’t easily spot).

What the developer experiences subjectively is the dopaminergic reward of I made progress. What’s actually happening, sometimes, is the AI generated something that looks like progress and I didn’t catch the parts that weren’t. Armin Ronacher, cited in the essay, described building tools with AI that he never used, only realizing later they did not end up working as I thought.

The deeper warning Thomas makes

The essay’s most aggressive section is about what developers should NOT do based on AI predictions. Thomas points to a series of failed forecasts: Geoffrey Hinton in 2021 said AI would replace radiologists within five years. Five years later, radiologists are still employed, AI is a tool they use. Dario Amodei claimed AI would write 90% of code by late 2025. The actual fraction is much lower and the human role remains substantive.

Thomas’s question: What if these forecasts don’t materialize? If you’ve stopped investing in your own skills based on Hinton- or Amodei-style predictions, and the predictions miss, you’ve voluntarily depreciated yourself. The downside of being wrong is being unemployable; the upside is being early to a future that didn’t arrive on schedule. Asymmetric in the wrong direction.

The actionable read

Thomas quotes Jeremy Howard: People who go all in on AI agents now are guaranteeing their obsolescence through outsourced thinking. The phrasing is harsh but the underlying claim is testable. If you’ve spent the last six months vibe-coding and not building durable skills (reading hard codebases, understanding system architecture at depth, writing without an AI in the loop), your bet on the AI takeover being complete is the thing securing your obsolescence — not the AI itself.

The essay isn’t anti-AI. It’s anti-religious-faith-in-AI. The empirical 19% slowdown finding is the part that should make every developer pause: subjective experience of acceleration is unreliable; only measured output is. If you haven’t measured your own output before and after going all-in on agentic tools, you’re operating on vibes about your vibes.