←
Brett A. McCully14 posts
Brett A. McCully avatar

Brett A. McCully

@BrettMcCully

Assistant Professor of Economics @CollegioCA

◷ Joined June 2012▧ 2 Photos and videos
Brett A. McCully avatar
Brett A. McCully@BrettMcCully·Aug 17

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.

Brett A. McCully avatar
Brett A. McCully@BrettMcCully·Aug 17

In particular, I found AI agents tended to overwhelmingly choose certain options like weighting regressions, choice of control variables, and sample selection, while human analysts exhibited a much more spread distribution. E.g.:

Post attachment
Brett A. McCully avatar
Brett A. McCully@BrettMcCully·Jul 12

* I’ve only tested it on my Windows machines, but Linux and macOS installers are available! Please let me know if you encounter any errors!

Brett A. McCully avatar
Brett A. McCully@BrettMcCully·Jul 12

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.

Brett A. McCully avatar
Brett A. McCully@BrettMcCully·Jul 12

So I fired up Fable and have vibe coded this fork of Antony Courtney’s Tad over the last month. Added the Stata command bar; updated the backend to neoDuckDB, among other things; overhauled the design; and made some other quality-of-life improvements.

Brett A. McCully avatar
Brett A. McCully@BrettMcCully·Jul 12

I was getting frustrated that Stata’s import parquet feature is only supported on StataNow, which I don’t have (it ain’t cheap!). This means I can’t casually browse parquet data files and spot check how the code cleaning is going, the way I’m used to doing with Stata dta’s.