Anthropic Funds New AI Benchmarks to Enhance Model Evaluation

Anthropic, an AI company, has introduced a program to improve how AI systems are assessed. Announced recently, the program will provide funding to external groups that develop new methods for evaluating AI performance.

 

These methods aim to measure the “advanced capabilities” of AI models more effectively. Interested organizations can submit applications at any time. According to Anthropic’s official blog, the goal is to elevate AI safety and provide valuable tools for the entire ecosystem. Developing high-quality, safety-relevant evaluations is challenging, and the current demand exceeds the supply.

 

Current AI testing methods often fail to reflect real-world usage and might not measure relevant aspects, especially those designed before advanced generative AI emerged. Anthropic plans to create new, more challenging tests focusing on AI security and societal impacts.

 

The initiative will address potential risks, such as the misuse of AI for cyberattacks or spreading misinformation. It also aims to develop an “early warning system” to detect potential AI threats.

 

Additionally, the program will fund research into beneficial AI applications, like scientific research, language translation, and reducing bias in decision-making. The goal is to develop AI that can identify and avoid harmful or offensive content.