Thoughts on self-inspecting neural networks. — LessWrong