JFC. There's exactly one thing this 'course' lists as a use for LLMs that I can actually see them being a reasonable use not loaded totally full of footguns: "synthetic sample data", and even then there are much more efficient ways to generate that... (You really don't need the monster tens-of-GB models to get "sample-shaped data" out of them...)

(Though I would still never actually use one for that, given how easy it is for them to end up giving you real data in the pile... They're just predictive autocomplete, after all, so if you accidentally give it an input that lines up close enough to data it was trained on enough times, it might just give you verbatim output of that training data. See the recent "extracting famous books from LLMs" research that blows a long-lance sized hole in the slopster's "it's not memorizing, it's learning patterns" argument... Not that that held much water to begin with, given that sufficiently overfitted patterns basically are memorization...)