AI agents rebuild a library from design docs, a post says
A user highlights a paper from Google DeepMind, MIT and colleagues about a performance-modeling library with almost no code on its main branch. According to the post, every human change is an edit to a design doc.
TLDR
A user describes a performance-modeling library organized around linked, natural-language design documents. According to the post, coding sub-agents regenerate the entire implementation whenever a version updates, using step-by-step worked examples in the documents as guidance.
The user says regenerated implementations reproduce hand-audited reference models to round-off precision, including DeepSeek-V3 serving on a TPU pod slice.
Combined views
1 Source, first seen 23d ago