I don't know how many users of LLMs fall into each of these buckets but I have the sense that there these distinct buckets. I don't think it's necessarily a global thing. A single user could fall into all three buckets depending on the domain they are working in. For instance for CI/CD I tend to ship clanker, for tests I interrogate LLM about output, but for code algorithms I am more likely to interrogate to learn.