METR uses undisclosed source to evaluate Anthropic's AI research
Summary
METR has utilized "an additional source of information" to gain insights into AI research and development at Anthropic, though details of this source remain undisclosed. This comes amid a broader trend of organizations conducting independent evaluations of advanced AI systems to uncover capabilities and risks that developers may not make public. Additionally, the release of system cards by AI labs aims to enhance transparency by detailing model training, evaluation methods, and safety considerations.
Analysis
METR: METR evaluates frontier AI models for capabilities, safety risks, and potential threats. The organization contributed to Anthropic's Opus 5.5 system card by drawing on an undisclosed additional source of information to analyze the company's AI research and development processes. This method enables deeper assessment when standard public data proves insufficient. Anthropic: Anthropic builds large language models with an emphasis on safety, reliability, and responsible development. The company's Opus 5.5 system card disclosed METR's reliance on confidential information to better understand internal AI research and development activities. This detail emerged as part of standard transparency documentation for the model release. AI Safety Evaluation: Organizations conduct independent assessments of advanced AI systems to identify capabilities and risks beyond what developers publicly share. System Card Practices: AI labs release system cards that outline model training, evaluation methods, and safety considerations to increase transparency around development.
Categories
aimachine_learningtech