Writing these image summaries has turned out to be a good use of LLMs for me. It isn't fully automatic, I still need to provide a manual outline about what each image shows to ensure the LLM interprets the image correctly. And in the end, there might still be errors in the summary I need to fix. But this is still way faster than doing all of the work manually. Sometimes it even picks up details I would've missed.
It's certainly possible to find valid use-cases for LLMs. The problem is that "if you want to write summaries for all these images faster than before, write a quick one for each of them, then send them to our AI, and then fix any errors" doesn't sound nearly as palatable as "send them to our AI and it will do it for you". Framing the technology as perfect leads to hazardous use, and those critical of the technology are lead to believe it's entirely useless because it doesn't fulfill these grand promises.
Writing the prompt sort of does remind me of programming a little bit. Depending on how you write it, you'll get the model to do its job efficiently or not, correctly or not. It's interesting to figure out.
And because I know this is a risky subject to post about: this is just me doing random shit with a technology to understand it better, same as trying a new eccentric programming language and what not. I still feel the same about the way it's widely used, and none of what I do will ever leave my computer.
