I wonder if this behavior emerges when using recursive self improvement to generate new models