llama.cpp adds activations, control-vector steering

This is a genuinely useful bridge between mechanistic interpretability and a production-adjacent local serving stack, not just another notebook experiment.

• –The live capture API makes activation analysis practical inside the server you already run, instead of forcing a separate instrumentation stack.
• –The binary collection format and top-K mean view keep the data path simple enough to automate and scale.
• –The inter-cluster differential scoring is the smartest part here: it targets behavior-specific features, not just whatever lights up on a single phrase set.
• –The MoE scale caveat matters a lot; control vectors are powerful, but they are also model- and embedding-dimension-sensitive enough to require calibration.
• –For local model users, this opens a clean path from observability to intervention: collect, train, probe, export, steer.