While LLM output looks promising and true, there is always something wrong — and you won't find it unless you have experience in the subject area.
For carthography AI is especially bad, because nobody writes proper documentation, so there is nothing to train on.
Here are three examples of Ian's LLM photo-to-tags bridge integrated into Every Door. While opening hours detection works great, everything else... Can you spot what's wrong in each tag list?
@everydoor Why is there nothing to train on? There are millions of POIs in OSM.