Opening article
Anthropic News has announced a new initiative focused on developing third-party model evaluations. Because full text is unavailable, further specific details regarding the framework or participants remain unconfirmed. The announcement highlights an ongoing industry focus on independent assessment methods.
A robust, third-party evaluation ecosystem is essential for assessing AI capabilities and risks, but the current evaluations landscape is limited. Developing high-quality, safety-relevant evaluations remains challenging, and the demand is outpacing the supply. To address this, today we're introducing a new initiative to fund evaluations developed by third-party organizations that can effectively measure advanced capabilities in AI models. Our investment in these evaluations is intended to elevate the entire field of AI safety, providing valuable tools that benefit the whole ecosystem.
In this post, we describe our initiative to source new evaluations for measuring advanced model capabilities and outline our motivations and the specific types of evaluations we're prioritizing.
If you have a proposal, apply through our application form .
We are interested in sourcing three key areas of evaluation development, which we'll describe further in the post:
We're seeking evaluations that help us measure the AI Safety Levels (ASLs) defined in our Responsible Scaling Policy . These levels determine the safety and security requirements for models with specific capabilities. Robust ASL evaluations are crucial for ensuring we develop and deploy our models responsibly. This category includes:
Beyond our ASL assessments, we want to develop evaluations that assess advanced model capabilities and relevant safety criteria. These metrics will provide a more comprehensive understanding of our models' strengths and potential risks. This category includes:
Third-party model evaluations are critical for establishing objective safety and capability benchmarks across the artificial intelligence industry.