How do you trust causal inference code that no human has read?
On New Year's Day this year, Isaac Gerber was a little bored. A week later he had shipped diff-diff, a difference-in-differences library that has since crossed 140,000 downloads, built almost entirely by AI agents. In this conversation we get into how he makes causal inference software he can actually stand behind, even when he never reads the code.
In this episode, we cover:
How Isaac built diff-diff, a difference-in-differences library, in a single day (now 140,000+ downloads)
A five-step workflow for building causal inference software you can actually trust
Why he builds with one model family and validates with another
How silent failures, like quietly dropped covariates, slip into AI-written code, and how to catch them
Why verification, not writing code, is becoming the real bottleneck
About The Guest Isaac Gerber is a data science leader focused on causal inference methodology and the open-source tooling around it. Isaac has 20 years of experience at the intersection of data, analytics, and business.
About The Host Aleksander (Alex) Molak is an independent machine learning researcher, educator, entrepreneur and a best-selling author in the area of causality (https://amzn.to/3QhsRz4 ).
Gerber, I. (2026) - "Design-Based Variance Estimation for Modern Heterogeneity-Robust Difference-in-Differences Estimators" (https://arxiv.org/abs/2605.04124)
Podden och tillhörande omslagsbild på den här sidan tillhör
Alex Molak. Innehållet i podden är skapat av Alex Molak och inte av,
eller tillsammans med, Poddtoppen.
