No papers needed: a 4-step recipe to start mech-interp with a local SLM and Claude
burny_tech · x · 2026-10-29
@neuralavb lays out an ultra-low-barrier path into mech-interp: all you need is a locally runnable model and a good coding agent. Steps: (1) forward a prompt or contrastive dataset through a small dense model like Qwen3.5 0.8B/2B/4B; (2) extract all neuron activations at one layer; (3) start with MLP neurons, which have a privileged basis; (4) feed activation data to an LLM agent like Claude to analyze specific neurons, optionally with standard topic-detection techniques. He calls it one of the coolest, most accessible fields in DL/LLM research.
More from coding & agent
- Awesome AI Apps repo: 132 open-source agent and RAG projects, 16.7k stars — Arindam_1729 · 2026-10-29
- My MCP Server Runs 100% of My SaaS, Yet I Still Open the UI Daily — Storm39276 · 2026-10-29
- Dev uses Claude to turn clunky old patent drawings into an interactive browsing site — jwt0625 · 2026-10-29
- Unofficial REA Guide: Use doctor to Tell MCP Registration From Provider Readiness — tian521057 · 2026-10-29
- OpenAI Log Preservation Order Shows Agent Builders Don't Control Data Retention — srivas_ashi · 2026-10-29
- Open-Source Orinyx Launches MCP Tool to Independently Verify Healthcare AI Agent Outputs — Practical_Green5775 · 2026-10-29