x
Malign initializations are more robust when the model can think better in the reasoning language than in the output language — LessWrong