Between the Layers– Interpreting Large Language Models
summary
Survey of recent interpretability work argues that mechanistic features replace single-neuron stories: Anthropic’s 2023–2024 papers use sparse autoencoders to unmix superposed activations, extract mostly monosemantic features, and apply ablation/steering as causal tests. Scaling to Claude 3 Sonnet and Claude 3.5 Haiku raises coverage and yields prompt-level attribution graphs and a Neuronedia circuit tracer that reproduce many predicted interventions—coverage improves substantially but remains incomplete, so monitors and causal probes catch recurring failure modes rather than proving global explanations.
AI-written summary · about
chapters (20)
- 0:00Introduction to interpretability in large language models
- 1:15Speaker and context
- 2:38Interpretability vs explainability
- 4:44Trust in AI systems
- 8:23Measuring trust in LLMs
- 11:17Why interpretability matters
- 16:54Classical explainability limits
- 23:19Realistic interpretability limits
- 25:35Visualizing neural networks
- 35:40Why classic methods fail
- 36:44Transformer architecture
- 39:16Modern interpretability tools
- 43:05Anthropic’s 2023 paper
- 45:35Superposition and autoencoders
- 48:45Scaling monosemanticity
- 50:54Biology of LLMs paper
- 53:51Mechanistic stories with graphs
- 57:06Talk summary and takeaways
- 58:34Choosing interpretability focus
- 59:37Transform Conf closing
related talks
59mBetween the Layers– Interpreting Large Language Models
52mGeneralization in the representations and computations of frontier language models.
20mMechanistic Interpretability
35mReading the Mind of an LLM
44mAI Intuition: Exploring Language Model Latent Space
40mIs my Large Language Model a Strange Loop?
21mWhy you should care about AI interpretability
41mSub-Linear Algorithms meets Large Language Models