Behaviour Manifolds and the Hessian of the Total Loss - Notes and Criticism — LessWrong