It's a good start. Also, it's good to use a toy problem for explaining how to do it. It would be great if more people published the results of careful experimen

It's a good start. Also, it's good to use a toy problem for explaining how to do it. It would be great if more people published the results of careful experiments like this, perhaps for things that aren't toy problems? It would be so much better than sharing screenshots! However, when you do have such a simple problem, I wonder if you couldn't ask ChatGPT to write a script to do it? Running a script would be a lot cheaper than calling an LLM in production.