OpenAI 拉了四家外部机构进公司内部查模型行为和训练数据,比通常的红队测试深一层。帖子还猜这就是 GPT-6.2 跳票的原因。关注模型发布节奏和安全监管怎么落地的人值得看看。
#third-party evaluations
8 stories · 1 in 7 days · 1 in 30 days
Stories, products and papers tagged "third-party evaluations", newest first.
Nov 17
Chinese contentGPT-6.2 或因安全审查延期,OpenAI 签约四家独立机构做内部评估
Oct 16
Chinese contentAnthropic 向第三方评估机构 METR 开放员工级访问权限
Anthropic 给了 METR 类员工权限来查自家模型安全,还写了篇博客讲怎么和第三方评估方合作,做法挺少见。
Sep 23
OpenAI Commits to Independent Safety Assessments Across Training, Evaluation and Deployment
The depth of access offered — reaching into training and deployment processes, not just final model releases — goes beyond the audit scope most labs currently grant external reviewers. Safety researchers, policymakers and enterprise buyers evaluating frontier-model risk can use the four priority areas as a reference point when comparing how OpenAI and rivals such as Anthropic and Google DeepMind structure external oversight.
Sep 18
Chinese contentAI 教父辛顿参与联署公开信:科技巨头应保证第三方评估人员工作无干预、受保护
“AI 教父”辛顿都参与联署了,这封信很重要,是给科技巨头的,要求他们让第三方评估人员能真正独立工作,不受公司干预,这个事很关键。
IT之家Source
Sep 15
Chinese contentAEF-1标准发布,Xai、OpenAI、Anthropic共同签署
Xai、OpenAI和Anthropic这些大公司一起搞了个新标准,以后第三方评估AI的时候会更统一、更透明。
Aug 5
Chinese contentOpenAI披露两起外部测评事故,Gary Marcus称乐观解决不了对齐
OpenAI自曝两次外部测评事故细节,Gary Marcus说乐观没用。快速了解AI安全测试的风险。
Jul 31
May 30