Most hard problems are hard because of huge uncertainty around what's possible and how to get there. It's true for LLMs as much as it is for humans (and for the same reasons). Here, you gave solid answers to both, all but spelling out the solution.
ETA:
> Is that how you think data gets fed back into models during training?
No, one comment chain on a niche site is not enough.
It is, however, how the data gets fed into prompt, whether by user or autonomously (e.g. RAG).
