How-to Transformer Mechanistic Interpretability—in 50 lines of code or less! — LessWrong