Between the Layers– Interpreting Large Language Models
summary
Distinguishes explainability (outside, human accounts) from interpretability (inside, causal wiring) and reports concrete methods: unmixing activations with sparse autoencoders, probing and causal ablation to test features, and tracing prompt-level flows with attribution graphs. Cites Anthropic papers showing monosemantic features increase with scale and a Claude 3 Sonnet experiment that raises feature coverage but not complete. Points to the open Neuronipedia circuit tracer as working tooling for auditing jailbreaks, bias flips, and targeted monitoring while warning this is local, testable evidence, not a global proof.
AI-written summary · about
chapters (9)
- 0:00Introduction and framing
- 3:26Understanding AI trust
- 9:36Defining model understanding
- 16:50Ethics and infrastructure
- 22:55Capability boundaries
- 24:20Classical explainability
- 29:07Limitations of classical tools
- 39:42Modern research landscape
- 50:24Scaling and future outlook
related talks
01h 01mBetween the Layers– Interpreting Large Language Models
35mReading the Mind of an LLM
52mGeneralization in the representations and computations of frontier language models.
44mAI Intuition: Exploring Language Model Latent Space
20mMechanistic Interpretability
24mOpen and Understandable: Designing Explainable AI for Everyone
21mWhy you should care about AI interpretability
40mIs my Large Language Model a Strange Loop?