“Anthropic can identify tens of millions of 'features' inside Claude’s neural network that correspond to human-understandable ideas and concepts.”