Brief

The Model Evaluation & Threat Research (METR) initiative has released an example protocol for assessing AI models' potential for autonomy-related risks. This protocol focuses on systems capable of executing harmful tasks end-to-end without human bottlenecks, including those that autonomously procure human assistance. METR aims for the evaluation to be practical, cost-effective, and completed by a small team within a month, with a budget of a few million dollars. The goal is to provide a continuous metric of dangerous capabilities to inform mitigation strategies and allow for societal oversight. AI

The Model Evaluation & Threat Research (METR) organization has published guidelines for assessing AI model capabilities, focusing on elicitation techniques. These guidelines aim to measure a model's potential performance after some level of post-training enhancement, rather than its raw state. The process involves initial basic elicitation, followed by analysis of remaining failure modes to determine if they can be easily fixed with further effort. METR emphasizes the importance of considering finetuning, prompting, and tooling in threat modeling, especially for open-source or potentially modifiable models. AI