David Roodman

Can AI detect the ten errors in Moretti 2021? I did a test of GPT5.2 vs GPT5.4 vs @RefineInk. Takeaway: current reasoning LLMs are useful, with room for improvement. 1/

I feel this way most weeks tbh. Sometimes I start approaching a problem manually, and have to remind myself “claude can probably do this”. Recently we were debugging a memory leak in Claude Code, and I started approaching it the old fashioned way: connecting a profiler, using the app, pausing the profiler, manually looking through heap allocations. My coworker was looking at the same issue, and just asked Claude to make a heap dump, then read the dump to look for retained objects that probably shouldn’t be there; Claude 1-shotted it and put up a PR. The same thing happens most weeks. In a way, newer coworkers and even new grads that don’t make all sorts of assumptions about what the model can and can’t do — legacy memories formed when using old models — are able to use the model most effectively. It takes significant mental work to re-adjust to what the model can do every month or two, as models continue to become better and better at coding and engineering. The last month was my first month as an engineer that I didn’t open an IDE at all. Opus 4.5 wrote around 200 PRs, every single line. Software engineering is radically changing, and the hardest part even for early adopters and practitioners like us is to continue to re-adjust our expectations. And this is *still* just the beginning.

I wrote a decent paper with AI. It took me about 3 hours from start to finish, including an interactive choose-your-own-border-RD-adventure, it’s a what are we even doing here kind of day.

My dorkiest post yet? Us urbanites with no off-street parking have to sling cords across the sidewalk to charge our cars. In DC, most people do it wrong, violating the ADA at least in spirit. Doing it right is easy if you buy the right stuff. @ggwash ggwash.org/view/102336/th…

@jt_kerwin @Vijayshreeeee Check me, but I have the impression that in the latest Stata version, they made major improvements to commands like collapse. The previous state of affairs was an embarrassment.

@Vijayshreeeee The classic reference nber.org/stata/efficien…

What are some good resources for working with very large datasets in Stata? Is there any resource that compiles relevant packages, like gcollapse, compress, etc.- packages/practices that make data work less computationally intensive and easier on memory? #econtwitter

@davidroodman Just ran 6 regressions on ~24M observations each, absorbing 36,000+ fixed effects (including 700 districts x 24 quarters). Total time: 20 mins with reghdfejl. This is magic!

Thank you. It was a labor of love, done from the privileged perch of a think tank with substantial unrestricted funding.

In 2023, we “red teamed” the research underlying our top charities. We found 4 mistakes and 10 areas we wanted to look into further. How did these end up actually changing our grantmaking a year later? 🧵

It’s been a tumultuous year for global health. On our blog, we share how we’ve used donor funds so far to respond to urgent needs from foreign aid cuts. Join us in doing as much good as possible. Learn about GiveWell’s work this year and ways to give: bit.ly/3XxCzkW

🚨 New working paper! How well do people predict the results of studies? @sdellavi and I leverage data from the first 100 studies to have been posted on the SSPP, containing 1,482 key questions, on which over 50,000 forecasts were placed. Some surprising results below.... 🧵👇

@notanastronomer "Friedrich has chosen to set Wagner's allegory not in the primeval world of German mythology...but in a futuristic "time tunnel," whose general look he confesses was inspired by a picture postcard...of the Washington Metro, the city's subway system." deseret.com/1989/6/25/1881…

this is DC metro ceilings appreciation account

<|:-)

Do you look at this regression chart and think, “ah yes, that’s definitely a negative quadratic relationship”? If you don’t, then you can learn everything else wrong with this paper on my blog today.

Interesting, and maybe very efficient, approach to extracting information about how much to believe an empirical paper.

We've updated a paper on the 3-year, $1000/month U.S. guaranteed income study. New results, in 3 figures: 🧵 1) Subjective well-being significantly improved in the treatment group in year 1, but there were no sig differences between the treatment & control group after that. 1/

David Roodman writes: consumers of economic research are more truth-seeking than the producers. Here is Roodman's recap of his re-analysis of a set of papers on temperature and judge decision-making. The comment process is not working that well! Gory details in 🧵 1/

Who cares when a research claim is found to be in error? Peer-reviewed journals do their best to deflect and dilute legitimate criticism. statmodeling.stat.columbia.edu/2025/09/02/who…

Replication doesn't make you popular, which is one reason it is undersupplied and sorely needed. I'm proud @open_phil is supporting @I4Replication. How the Institute for Replication is making social science more robust and reliable | Open Philanthropy openphilanthropy.org/research/how-t…

Good illustration of how much the Texas grid has changed in just 6 years. yellow = solar; purple = batteries; dark green = wind; blue = gas; brown = coal; light green = nuke via @grid_status

Springer Nature book on machine learning is full of made-up citations. retractionwatch.com/2025/06/30/spr…