This pretty closely matches with my experience, and I’m not sure when people show these amazing tools or games built with vibe coding, exactly what percentage is, especially when they talk about not reading the code and just pushing accept all.

Don’t get me wrong, it’s amazing, but it’s also just wrong a lot and develops weird habit, like assuming it knows certain business logic rules that you’ll want, even when you explicitly say you don’t want them, and silently bringing them back during later iterations.

Also, and maybe I either know too much or too little about LLMs, but if you’re training data has lots of bad examples, why necessarily will LLMs keep getting better and better? How does an LLM decide between a good pattern and a bad pattern, particularly for patterns that are not commonly represented in the training data?