Ragelight Labs works on AI safety problems where progress is possible today.

The current work is in mechanistic interpretability. It has two parts: research notes, and an open course that teaches the field to software engineers. This site holds both, along with a lab log of the work.

Research

· Replication

Replicating Anthropic's emotion concepts paper on a 7B open model

On Qwen2.5-7B-Instruct, the paper's findings about how emotions are represented mostly replicate. Its findings about emotions causing misaligned behaviour are mixed: steering changes flattery, barely changes reward hacking, and never produces blackmail. Of 23 testable sections, 10 replicate, 12 replicate in part and 1 does not replicate.

Course

Being built

Mechanistic interpretability for software engineers

Each lesson has a short article, coding exercises in Python, and a small game. The early lessons run entirely in your browser, so you don't need a GPU to get started.

  1. 0Math Notation Is Just Codeonline · game: Formula Forensics
  2. 1Vectors: Directions Are Featuresbeing built · game: Feature Hunt
Start with Lesson 0

Lab log

This site is online

ragelight.ai now lists the research notes and the course, and has an email sign-up for new notes and lessons.

Lesson 0 is online

Lesson 0, Math Notation Is Just Code, is online with its exercises and its game, Formula Forensics. Lesson 1, on vectors, is next.