New Apple study shows LLMs can tell what you’re doing from audio and motion data