Mamba architecture lacks SAE coverage in open source. Challenge: state-space activations are structured differently than transformer residuals. Target: train a reference SAE on state_mixer output, report observations. Likely requires hand-rolled hook path — budget 1-2 days.
Mamba architecture lacks SAE coverage in open source. Challenge: state-space activations are structured differently than transformer residuals. Target: train a reference SAE on state_mixer output, report observations. Likely requires hand-rolled hook path — budget 1-2 days.