Per Engzell and Nathan Wilmers provided some real insight on experiments with how well LLMs can produce quantitative social science paper production, even held back by using Stata of all things.

This is an interesting project I am tracking. There are so many AI agent harnesses, but harnesses make such a difference in the capabilities and usefulness of the tools. Particularly smaller LLMs really benefit from strong harnesses. Could be super useful!

Fascinating to see people using AI to replicate the work of academia. I saw some interesting critiques asking if this means papers never were and should not be moving forward the primary unit of output. I think that’s an interesting question.

A client-server protocol for DuckDB could really unlock a lot of agentic data science workflows for me - collaborating on the same database without running into conflicts or waiting in line to push updates.