LLM? Not so much IMO. But neural nets and deep learning are genuinely cool and useful for things like image recognition, speech recognition, etc. As an aide to a doctor reviewing a colonoscopy scope, sure. But instead they will treat it as a replacement that the doctor rubber stamps. Due to the way we will use it, they are dangerous, even if the tech itself can be helpful.

Theres still the issue of electricity, of the human and environmental costs, etc.

A good metric: if correctness is unimportant and verification is easy, then it is not an inappropriate tool (that doesnt mean its necessarily a good tool for a given task though). However if correctness is important and verification is difficult, then it is not appropriate. Basically, dont use it to summarize that many-page report, because its nondeterministic and you'll have to read the whole thing anyway to see if the LLM actually told you all relevant portions and didn't generate things that aren't in the report. If you want a summary, thats literally what the abstract and conclusion parts of a report are for. Dont use it to write fresh code for you, cause you'll have to read through all of it, build a mental model, and understand the problem space inside and out to actually review it well, and at that point you might as well have written the code.