Please explain how these cause a "convergence to the mean."

I assume they’re saying that the most common and popular solutions propagate, power-law style. LLMs just amplify that loop.