Interpretability of SAE Features Representing Check in ChessGPT — LessWrong