scott cunningham

I’ve been doing a series on Claude code for “quantitative social scientists” on my substack. It’s not much a series I guess as much as I am regularly documenting what I’m learning. It started mid Dec 2025 and will continue. causalinf.substack.com/s/claude-code

HTML is the new markdown. I've stopped writing markdown files for almost everything and switched to using Claude Code to generate HTML for me. This is why.

@alexolegimas @GoogleDeepMind @ShaneLegg That's great bud! Congratulations! They're lucky to have you.

Some news: This week I am starting at @GoogleDeepMind as Director of AGI Economics on @shanelegg’s team. I will be joining the other amazing cross-disciplinary scientists researching AGI there. My team will study how frontier AI could reshape the economy: what happens to work and labor, how wealth and power are distributed, how institutions adapt, how AI agents shape markets, and what kinds of models can help us reason clearly about futures that may look very different from the past. I’m incredibly excited to help build this research agenda. If AGI changes how society operates, economics is going to be critical for shaping our shared future. Many more announcements soon.

Main result: Agents recover most published findings. ✅ Sign of coefficients correct: ~80–90% ✅ Within 95% CI: >70% (best models: >80%) That’s real progress toward automated reproducibility.

Can AI agents read a social science paper and write the code from scratch to reproduce its results? No access to original code. Just text + data. New paper with Ben Kohler, @david_rzs, @__jae_1, and @miserlis_ 👇

@ahall_research @alexolegimas @emollick @causalinf @cblatts @YanagizawaD @soumitrashukla9 @BrianNosek @MetaScience @stuartbuck1 @andrewgelman @karpathy @xuyiqing @ddkang @sayashk @random_walker Arxiv version now up: arxiv.org/abs/2604.21965

I am sharing a set of slides from my lecture on agentic AI tools for Applied Economists at @penn_state. Materials are available here: github.com/ballesterogh/A… This builds on this thread and great work by Paul Goldsmith-Pinkham, @aniketapanjwani, and @VelikovMihail

Somehow this substack is twice as long as my normal substacks which are twice as long as any sane persons substack should be. But it’s on AI agents and the minimum wage causalinf.substack.com/p/claude-code-…

Okay, it's been 3 months since I took the plunge into Claude Code and 2 months since I created Claude Blattman to help other non-coding, non-technical, forever-behind people like myself learn how to use these tools. This is a Here's what I've developed and created since then. 🧵

🚨New preprint! We find evidence of LLMs enabling people to file lawsuits without lawyers (filing "pro se") at historically unprecedented rates in federal courts.👇 1/n

Legend

Another entry at my substack on the Callaway, Goodman-bacon and Sant’Anna (CBS) continuous diff in diff paper (conditional accept at AER). I continue to talk their Table 1 TWFE decomposition weights. causalinf.substack.com/p/twfe-continu…

Journal submissions are increasing, while reviewer time is not. At @RefineDotInk, we want to integrate the best technical verification AI can provide to improve academic peer review. To that end, we're inviting journals to partner with Refine, with usage heavily subsidized. 1/

Dear followers, please see this conversation with Jon Stewart and David Autor on AI, work, inequality and learning. I personally had a great time.

More TWFE decomposition for continuous treatment diff in diff at my substack today. I update the shiny app too. Short video at the top explains the changes to the shiny app. causalinf.substack.com/p/twfe-continu…

More TWFE decomposition for continuous treatment diff in diff at my substack today. I update the shiny app too. Short video at the top explains the changes to the shiny app. open.substack.com/pub/causalinf/…

Both worth understanding 😀

Pow! Returning champ John Korir from Kenya sets a Boston record at 2:01:52, shaving 30 seconds off his PR!

Great example of always keeping in mind baselines!

@rajdeep_c28 Check out today’s post. I have a shiny app to help reason through the FWL decomposition of the TWFE regression coefficient. It can help answer your question. This is for the continuous case, not multi-valued. causalinf.substack.com/p/making-a-shi…

@causalinf But for a 3*3 scenario, there are 4 independent basis vectors for the treatment indicator, and their linear combinations as well, these combinations could have the "bad" comparisons. How do we extend this framework to the continuous case, if we were to think along these lines?

I’m going slowly through this continuous did paper by Callaway, bacon and Sant’Anna on my substack. I usually teach the parameters and the estimator but really downplay the TWFE decomposition. It sort of dawned on me that bacons decomp had made me interested in did tho.