A method to inspect what a language model 'knows' at each layer by examining the vocabulary predictions at intermediate depths.