Anyway, it's an interesting approach. Feels more like PI work than DS work at times — everything from "great work!" or "here's some feedback for improvement" to "fine I'll do it myself!" haha.

I also wonder how helpful it may be to ask the model to reason through the task, but then substitute better tokenizers, better algorithms, replace linear models with graphs, replace regex with sentence transformers, etc.

That answer will differ based on the task and relative performance of each model.

The usage question is when that is good enough. It can do something in 5-10 minutes that I can do *WAY* better, but takes me 5-10 hours (or more). As always, it's a trade-off. Is it accurate *enough*? And does the time-to-result outweigh the (in)accuracy?

One thing I keep noticing when watching it "think" is that the process it follows is very remedial (basic tokenization, K-means clusters, regex, etc.). It's doing upper-division undergrad work. (Which is better than the 9th-grade essays it was writing last year! But maybe not professional enough.)

Like you don’t even really need to say why, though if it’s significantly inspired something in your own research or teaching, do say so.

Ditto if you give someone’s work to students and they enjoy it or it sparks lively discussion.

Normalise telling people how awesome they are.