love this part

”This language-independent approach represents a different paradigm for AI safety and interpretability.”