Vibe coding only seems to work, insofar as it does when the training data includes multiple exemplars of solutions to a given problem.

As noted elsethread, there's only one geometric kernel which is decently far along and opensource and it's over 1 million LOC --- I doubt it's being included in any training data, and I doubt that an LLM could regurgitate such a large project which would then compile w/o errors and then work as expected --- the number of tokens required to get such a project to an initial state is a marked hurdle as well.