Our previous interpretability research found that we could artificially dial up or down certain “features” within a model to modify its behavior. Here, we ran larger-scale tests of feature steering, focusing on measuring social biases.

→ View original post on X — @anthropicai