Vibe Coding will 10x your best engineers – and your worst ones too
If you haven’t caught on yet, meet ‘vibe coding,’ popularized by AI researcher Andrej Karpathy in a tweet that hit over 4 million views. I won’t try to do it justice in a paraphrase as the term has since taken on a life of its own.
I can only pity the developers who took the advice literally and stopped developing to only ‘vibe.’ I’ve said before that ChatGPT marked the beginning of a broader dumbing-down of how we engage with information; vibe coding is the same pattern arriving for software engineers specifically.
Karpathy himself flags the risk in his own tweet, admitting the generated code ‘grows beyond my usual comprehension’. This, from a presumably rockstar coder! If I were a CIO or CTO today, I’d restrict vibe coding to POCs and demos, full stop, and keep it out of production. I say that after three months of hands-on experimentation with it myself.
The code these models generate is genuinely sloppy, in line with the ‘AI slop’ the media has been reporting on. That’s not a mystery. Most code on the public internet is average at best, so a model trained on it produces average code, favoring a clunky for-loop over a clean list comprehension. To their credit, these models do correct themselves with the right nudge.
Here’s the CxO-relevant framework, in five conclusions I’d stand behind.
One, code generation is a gift to senior engineers with solid fundamentals. A 10x productivity multiplier with the model as sidekick, not a replacement.
Two, the roles of business analyst, systems analyst, and developer are converging into one, with prompt engineering folded in. Technical staff should brush up on language skills, non-technical staff on coding fundamentals.
Three, developers who can only ‘vibe’ get replaced by the models themselves as they improve. AI won’t level the field between mediocre and high performers; it will widen the gap between them.
Four, restrict vibe coding to POCs and small programs, at least for now. The code gets messy fast and debugging it costs more than rewriting it would have.
Five, generate the first draft with the model, then polish it with real engineering judgment. Never ship the generated version as final.
Call it the Amplifier Effect. AI doesn’t level talent; it amplifies whatever talent is already there. A rockstar engineer becomes 10x better. A mediocre one becomes 10x better at being mediocre. The same holds for content creators, researchers, and analysts who are about to start ‘vibing’ in their own domains too.
If your engineering org is adopting vibe coding without a policy on where it’s allowed to touch production, that’s a gap worth closing before the Amplifier Effect works against you instead of for you.