The paper doesn't compare to human choices, which would be interesting. When I did so, I found a substantial divergence between human and AI research choices:

I think the choice of model may also be important. I tried Haiku 4.5 like them, but dropped it when seeing it generate degenerate results, e.g., SE=0. I ended up using just Sonnet 4.6 and GPT-5.4 mini.

I made a little thing that allows for execution of simple Stata commands while browsing parquet or CSV files. github.com/brettmcc/tads

Incredible that I can now produce functioning* software with a market of ~7 people (those working with parquet who are beholden to Stata’s syntax) just based on vibes and access to frontier AI.