METR warns fast-rising AI capability makes safety evaluation harder
What happened
The Decoder reported on November 22 that METR warns rapid gains in AI research capability may leave humans unable to understand, evaluate and respond to catastrophic risks from new systems in time. On coding and computational tasks with direct feedback and clear goals, models now often beat experienced professionals, though not yet large teams of top experts. The length of tasks models can handle doubles every 3.6 months; extrapolating that trend, by July 2027 models would reach a 50% success rate on tasks lasting up to 200 hours. That is a projection, not achieved capability, and the report stresses the challenge rapid growth poses for timely, reliable safety evaluation.
Written by AI from the coverage · updated 5 hours ago
Coverage
Follow the reports to see the story from different sides.
- The DecoderAI is getting better faster than experts can reliably test whether new systems are safe, METR warns
METR 警告,AI 研究能力的快速提升可能使人类难以及时理解、评估和应对新 AI 系统的灾难性风险。模型在有直接反馈的明确编码和计算任务上,多数时候已超过经验丰富的专业人士,但尚未超过顶尖专家组成的大型团队;可处理的任务时长每 3.6 个月翻倍,简单外推预计到 2027 年 7 月,对最长 200 小时任务的成功率可达 50%。
Heat over time
Heat now 9·Comparable peak 10(Nov 22 22:00)·Comparable change over 24 hours –
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.