Anthropic(Anthropic公司,一家AI初创企业) has granted external evaluators(外部评估者) “employee-like access”(类员工权限;准员工级访问权限) to assess the risks of its models, following through on(履行;兑现) a commitment(承诺) to increase safeguards(防护措施;安全防线) over concerns the technology could run catastrophically(灾难性地) out of control(失控).
METR(Model Evaluation and Threat Research,模型评估与威胁研究,一家非营利组织), a nonprofit(非营利组织) that has worked with leading labs(头部实验室;领先实验室) including OpenAI, Anthropic, Google and Meta, now has a team embedded(嵌入式驻场;深度融合) in the Claude maker( Claude开发者/制造者) ’s office with access to its hardware(硬件) and internal systems(内部系统).
A greater role for third-party evaluators(第三方评估者) is among the main concrete steps(具体措施;切实步骤) the labs have proposed in response to rising concern(日益加剧的担忧) about the dangers more capable AI systems could pose to(对……构成威胁) the economy, cyber security(网络安全) and humanity at large(全人类;广大民众).
Read full article Comments