Frontier labs' security too weak for pre-training and RL work
Given the current security posture of US AI labs, it is not good for American national security for them to keep improving pre-training and reinforcement learning algorithms, because those innovations can quickly leak to adversaries.