San Francisco-based Anthropic is warning that AI models are improving so fast humans may lose control over its development