Edward Yang: LLM-Annotated Terminals Make One-Off Dynamic Analyses Trivial

ezyang · x · 2026-09-19

PyTorch core developer Edward Yang (ezyang) shares a "shower thought": it has never been easier to create one-off dynamic analyses. The recipe is to design the abstract domain, use an LLM to annotate all the terminals in your program, and then run the analyses to check non-local problems.

The idea casts the LLM as a program annotator, replacing annotations that traditionally required manual effort or heavy static inference, making ad-hoc, problem-specific program analysis cheap and fast to build.

Original post →

More from coding & agent

coding & agent channel →