If you use an LLM to generate source code you are vibecoding.
You specify the problem in natural language (the vibes) and the LLM spits out source (the code).
Whether you review it or not, that is vibecoding. You did not go through the rigor of translating the requirements to a programming language, you had a nondeterministic black box generate something in the rough general vicinity of the prompt.
Are people seriously trying to redefine what vibecoding is?
No, that is literally vibecoding. Reviewing vibecoded source is just an extra step. It's like saying "I'm not power toolgardening, I use a pair of gardening scissors afterwards." You still did power tool gardening.
As additional proof, the dictionary definition of vibe coding is "the use of artificial intelligence prompted by natural language to assist with the writing of computer code" [1]
It seems like vibecoders don't like the label and are retconning the term.
Both you and the Collins dictionary (merely one dictionary, not an absolute anuthority) are retconning. “Vibe coding”, as originally coined in this tweet, means something more specific: to generate code with LLMs and not really look at the output. The term itself suggests this too: reviewing code is not exactly a vibes-based activity, is it?
That tweet coins the term, we agree there. The activity it describes is using natural language to generate software. Whether you add a review process or not doesn't substantially change that. Sure, Karpathy says he doesn't "read the diffs anymore". Why does he say "anymore"? Clearly he was reading them at some point. If not reading any diffs was a core part of the activity, that wouldn't be the case, the tweet itself clearly outlines that as optional. He's clearly not talking about a core part of the activity.