Anthropic is launching an initiative to develop better standards for evaluating the performance and impact of AI models. The AI developer is offering independent third-party consultants or researchers funds to develop these standards.
The new benchmark standards initiative should ensure new standards for better evaluating AI models such as LLMs. According to the AI developer, this involves new benchmarks for assessing their safety and the ability to manage advanced outputs such as hypothesis generation, as well as standards for infrastructure, tools, and methods for developing these evaluations.
Shaking up existing AI benchmarks
With this initiative, Anthropic is calling for a complete change to all existing benchmark methods for evaluating AI models, especially when it comes to the security of AI, by developing tools that benefit the entire ecosystem.
One example is a benchmark that evaluates whether an AI model is capable of malicious actions, such as cyberattacks, manipulating or deceiving people, enhancing weapons of mass destruction, and more. This should eventually provide an early warning system for potentially dangerous models that could threaten security.
Through this initiative, Anthropic also wants to see more benchmarks specifically focused on scientific research, where it is important to eliminate bias and be able to interact in multiple languages.
Ultimately, this in turn should create new tooling and infrastructure that allows experts to create their own evaluations for specific tasks, followed by large test runs with hundreds or thousands of users.
Setting up the initiative
The initiative, led by a specially appointed coordinator, also offers research fellowships. In addition, Anthropic indicates that the most promising projects in the field of benchmarks may receive investment.
In doing so, the AI developer indicates that it does not want to favour its own AI models. The initiative should catalyse the broad AI industry and pave the way for future developments where understandable evaluations are the norm.
Anthropic is not the only AI developer setting up an initiative to develop new benchmark standards. Sierra Technologies also recently announced a new self-developed benchmark standard for evaluating the performance of so-called AI agents.
Also read: Claude 3.5 Sonnet: Anthropic once again beats OpenAI and Google