Steering interpretable language models with concept algebra