ai模型改进速度超过安全测试,metr警告
ai模型在编码任务中击败人类专业人士,但安全测试难以跟上,构成灾难性风险。
ai models are improving rapidly, surpassing human professionals in clearly defined coding and computational tasks with direct feedback. metr found that models can handle tasks that take humans longer, with capabilities doubling every 3.6 months. a naive extrapolation puts models at a 50 percent success rate on 200-hour tasks by july 2027. however, systems struggle with open-ended questions and situations hard to recreate for training. metr warns that common tests may overestimate model performance in real-world use.