Generalization in the representations and computations of frontier language models.
summary
Walks through geometry of LLM activations and shows a concrete decoding method using sparse autoencoders applied to the residual stream of Claude 3 Sonnet, recovering roughly 34 million features. Finds interpretable, causally manipulable directions: a Golden Gate Bridge detector that fires across languages and images, a cherry‑blossom vector that steers tone when amplified, and a code‑error vector whose ablation repairs tracebacks. Tests include histograms, forward‑pass stimulation and ablation, a 1-in-30 hit rate for bias steering, and open questions on compositionality and hierarchy.
AI-written summary · about
chapters (10)
- 0:00Introduction and context
- 1:22Generalization challenges
- 6:01Linear representation hypothesis
- 12:54Sparse autoencoders
- 17:48Golden Gate Bridge feature
- 23:11Interactive feature analysis
- 26:44Interpreting model features
- 36:07Steering and interventions
- 40:46Steering case studies
- 45:36Future directions and closing
related talks
01h 01mBetween the Layers– Interpreting Large Language Models
35mReading the Mind of an LLM
59mBetween the Layers– Interpreting Large Language Models
37mLinear representations of concepts in modern AI models
49mHow grounded are ungrounded LLM models
46mUnder the Hood of LLMs — all those questions you never dared ask about them
44mAI Intuition: Exploring Language Model Latent Space
12mLanguage-Model Probabilistic Programming for Improving Autoformalization via Cycle(…)