Anthropic warns AI may soon begin recursive self-improvement - Scientific American