Anthropic Launches AI Benchmark Improvement Program

Anthropic, an AI research and development company, has revealed its new program to sponsor humanity to produce better and more efficient AI benchmarks. Recognizing the current disjointed state of AI evaluations, this initiative aims to support third parties in developing new tools, structures, and approaches to accurately assess advanced AI performance.

Anthropic stated, “Our investment in these evaluations is intended to elevate the entire field of AI safety, providing valuable tools that benefit the whole ecosystem. Developing high-quality, safety-relevant evaluations remains challenging, and the demand is outpacing the supply.”

The situation has pointed to the necessity of improving AI performance benchmarks in the past few years. Today’s measures tend to miss the application of Artificial Intelligence systems and projects and can be misleading of what they are designed to measure in the first place. This is especially true about many of the modern generative AI models, which often leave many of the evaluated metrics behind.

Introducing Anthropic’s Claude 3 iOS App & Premium Plan for Businesses

Instead, the current state of affairs of benchmarks and metrics is seen as a problem by Anthropic, for which the proposed solution is setting new high standards that are difficult but oriented towards the security of AI as well as society. The company wants tests that would challenge a model’s prowess in conducting cyber attacks, improving on weapons, applying psychological control on people and many other aspects. For national security and defence risks, Anthropic is obliged to work on an early warning system, however, more details on this system have not been disclosed.

“We offer a range of funding options tailored to the needs and stage of each project. Teams will have the opportunity to interact directly with Anthropic’s domain experts from the frontier red team, fine-tuning, trust and safety and other relevant teams.” – Anthropic

Apart from security objectives, Anthropic’s program will enable research on prospective benchmarks of AI utilization for scientific analysis, ML translations, reduction of bias, and self-censorship of toxicity. To this end, Anthropic aims to build new platforms for subject matter specialists to build their models for evaluation and perform large-scale tests with thousands of subjects. Currently, the company has employed a full-time coordinator for the program; the company might buy or expand some of the attractive projects.

Opus vs Sonnet vs Haiku: Check Key Differences Between Models Of Anthropic Claude 3

Concerns

Nevertheless, Anthropic’s attempts to improve the benchmark of AI are honourable; however, there have been some concerns about its implicitly commercial approach and undue focus on what are called catastrophic risks. Some human resources challenges cited include a claim that the company’s safety classifications may precisely apply pressure on applicants, while the discourse on world-ending problems diverts attention from tangible AI governance concerns.

Thus, Anthropic’s goal of turning proper AI assessment into a regular process is a good one to have. It can be said that a need for a better AI benchmark is required to be set and Anthropic’s program is a step in the right direction.

Claude 3.5 Sonnet by Anthropic AI: Faster, Smarter, and Now Available