When does an AI-designed experiment become a validated successor-model improvement?
Posts here