If concepts correspond to directions in activation space (see superposition), you can push the model along them. That is the idea behind steering. Unlike fine-tuning, it changes no parameters; unlike a prompt, it leaves the input alone: the intervention happens in the residual stream while the model computes.
The simplest version is Turner et al. (2023)'s activation addition: run a pair of contrasting texts through the model, such as "Love" and "Hate," subtract their activations at some layer, and add the resulting vector while the model generates another text. This shifts topic or tone with no optimization at all. Another version clamps a feature found with a sparse autoencoder: Templeton et al. (2024) clamped Claude 3 Sonnet's "Golden Gate Bridge" feature at ten times its maximum activation, and the model began to present itself as the bridge. In May 2024 Anthropic opened such a version to the public for 24 hours: Golden Gate Claude (Anthropic, 2024b). In the same vein, Lu et al. (2026) found that the leading component of several models' "persona space" is an Assistant Axis: moving away from it leads the model to identify as other entities, and capping activations within a region of the axis stabilizes it against persona-based jailbreaks.
Steering also works as evidence. If a probe finds a direction and pushing along it changes behavior as predicted, the direction is causally relevant rather than a mere correlate; this is how the board representations in Othello-GPT were validated (Millière & Buckner, 2024b, sec. 2.5.3).
Golden Gate Claude is also a lesson about anthropomorphism: a "self" that changes with a dial invites caution about a chatbot's self-reports. What, then, is the assistant persona: a stable trait of the system, or a region of its activation space from which it can be pushed out?