AI that can build itself would be a major development in the history of technology, but "full recursive self-improvement also might increase the risks of humans losing control over AI systems," the AI startup said.