Anthropic's S-1 filing defines recursive self-improvement as the capability of an artificial intelligence model to improve subsequent generations of models