It is also a much less interpretable and auditable path towards recursive self-improvement. At least with training code, human engineers can inspect and reason about the output.