Linear representations of concepts in modern AI models
summary
Claim: many human concepts inside LLMs sit on one-dimensional directions recoverable by linear probes, i.e. single index models, and can be found and used to steer behavior. Method: apply a supervised feature-extraction loop based on Recursive Feature Machines and the average gradient outer product (a supervised PCA analog), bootstrap predictors on ~4,000 examples, iterate, and extract one steering vector per transformer block. Results: vectors reliably flip outputs (SSN leakage, Golden Gate prompts, English↔Chinese and Python↔C++ translation) on open LLaMA models and often preserve or improve correctness.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:22AI model mathematics
- 8:53Dimensionality puzzle
- 12:18Linear representation hypothesis
- 17:22Single-index models
- 20:47Supervised dimensionality reduction
- 26:50Steering LLM outputs
- 31:58Why steering vectors work
- 34:07Conclusion
related talks
35mReading the Mind of an LLM
52mGeneralization in the representations and computations of frontier language models.
46mLanguage Models as World Models?
20mMechanistic Interpretability
28mExplaining Machine Learning Models with LLMs
56mMy AI Learning Journey
36mLLMs and AI agents demystified
18mWhat We Learned from Using LLMs in Pinterest