Power Steering: Behavior Steering via Layer-to-Layer Jacobian Singular Vectors
Update March 12, 2026: Larger rewrite for improved clarity
Omar Ayyub
My work and interests.
Latest posts
What I Learned (And Didn’t) Steering Qwen3 Models
Recent discussion about the tension between mechanistic understanding and applied safety outcomes got me interested in steering vectors. They offer behavioral control without requiring mechanistic insight: a low-compute way to shift model behavior, even if you don’t fully understand why it works. Probes have been deployed in Anthropic’s systems with remarkable success, and steering vectors have…
Discovering Player Tracking in a Minimal Tic-Tac-Toe Transformer
Interpreting Machine Learned Models Mechanistic interpretability has been gaining prominence as the technical approach to AI safety and alignment. Mapping the algorithms used by models would lead to tools for understanding, analyzing, and manipulating model decision making and risks. The results so far have been promising! Anthropic pushed their LLM, Claude, into constantly discussing the golden…