"Inside the Black Box": interactive 11-chapter survey of mechanistic interpretability
cneuralnetwork · x · 2026-11-09
An interactive survey, Inside the Black Box (2026 edition) by @neuralavb, is making the rounds: 11 chapters + reference section, 8.5 hours of reading, interactive diagrams throughout, in plain English. It traces four eras of mech interp — 2013–2019 vision-model probing (saliency maps, DeepDream), the 2020–2021 Circuits program, 2022–2023 language-model circuits (induction heads, activation patching, first SAEs), and 2024–2026 scale-up: sparse autoencoders unpacking millions of readable features and attribution graphs tracing how Claude plans rhymes and refuses requests, plus white-box audits entering system cards and production.
More from Research
- SimpleHC cuts Hyper-Connections training overhead from 59–100% to just 2.4% — SonglinYang4 · 2026-11-09
- LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis — Kwai-Kolors · 2026-11-09
- NVIDIA's AutoCompress Beats Minitron by Up to 6.57 Points With Hardware-Guided LLM Pruning — nvidia · 2026-11-09
- BILBY: New Benchmarks Test Protein Scaffold Designs for Dynamic Stability, Not Just Static Success — HFMI-PROFOUND · 2026-11-09
- Stanford: Safety Alignment Only Cuts AI Agent Impersonation from 65% to 44% — AI Engineer · 2026-11-09
- Modal: RL Training of Your Own Agent Models Is Now Practical — AI Engineer · 2026-11-09