Also, if you want to replace RA classification work via an LLM API, going with an Open-Source LLM (nowadays quite capable) and a low model temperature might improve replicability. In this case you should also try different model parameters for robustness, especially now that the APIs are so cheap.

I think it’s worth emphasizing how good Claude Code is in VSCode nowadays. Both for quick prototyping (e.g., building zero-shot streamlit GUIs for classification tasks) but also for important code/ debugging. For the latter, it’s important to iterate slowly in order to be able to verify the code.

Bonn is (close to) free regarding tuition if that helps

Crazy that this is in such a stark contrast to manufactoring labor productivity in Germany (not quite the same statistic though: different timeline & output per worker, from here: www.wirtschaftsdienst.eu/inhalt/jahr/... )

Related reporting: www.ft.com/content/7052...

You can order these online through various online shops, here's an overview PDF with some shops on the third page: assets.adac.de/image/upload...