Consensus Weight
“But then we saw Meta’s head of AI safety use OpenClaw, and it deleted her entire inbox. That’s the head of safety for Meta. So, if she’s having problems, I think we all have to be asking: ‘How do we make sure that these systems are really hardened? How do we make sure that they’re rigorously tested? How do we make sure that we can actually delegate to them in a trusted way?’ And that’s really the hardest problem to face, right?”
。PDF资料对此有专业解读
Photograph: Chris Haslam
进步在电子表格建模上表现得最为具体。GPT-5.4 模拟初级投行分析师完成建模任务，平均得分 87.3%，GPT-5.2 是 68.4%，GPT-5.3-Codex 是 79.3%，差距将近 20 个百分点。