Roland Pihlakas comments on Shallow review of technical AI safety, 2025