Neuralese is Actually Probably Good for Alignment
Argues that neuralese chain-of-thought trained by backpropagation through time with SFT-style cross-entropy could avoid reward-model hacking on inexactly graded tasks, at the cost of interpretability.