“At a private conference, the consensus view among modelers from the four largest AI companies was that significant AI safety regulation would only be implemented after an accident causes 50 to 100 million deaths.”