Alignment is not free: How model upgrades can silence your confidence signalsvariance.co·121 pts·karinemellata·67
We used sparse autoencoders to explain LLM moderation flags of violent threatsvariance.co·6 pts·karinemellata·0