I built an LLM regression gate in 60 lines of Python that catches 80% of prompt bugs in 4 minutes per PR

Forschungstagebuch Nr. 1 – Rekursion, Persistenz und Attraktorbildung

CMU research study on spec-driven development — looking for open-source devs to interview (45-60 min, Zoom)

Why are ai tools becoming worse and more of an issue than an asset

**"What's actually useful in a master prompt and what's just placebo?"**

this prompt turns a pile of sources into a fully structured essay argument you just need to copy and paste it

minimax m3 hit 83.5 on browsecomp vs opus 4.7 at 79.3. ran 5 of my actual deep research prompts side by side this week