Meta has introduced the “Purple Llama” initiative in connection with its Llama artificial intelligence (AI) technology. Revealed on Thursday (Dec. 6), Purple Llama focuses on safety mechanisms and assessments aimed at assisting developers in the responsible construction of their AI models. The choice of the color purple is inspired by the cybersecurity realm, where the concept of “purple teaming” involves a collaborative approach combining offensive (red team) and defensive (blue team) strategies to effectively address the challenges posed by generative AI, as stated on the company’s blog.
In July, Meta unveiled Llama 2 in collaboration with Microsoft, with over 100 million downloads of Llama models to date. The project’s components will be licensed permissively, enabling both research and commercial utilization to encourage cooperation and standardize trust and safety protocols. Meta also disclosed its development of what may be the initial industry-wide set of cybersecurity safety assessments for large language models (LLMs). These benchmarks, formulated in conjunction with security experts, adhere to industry guidelines and standards to tackle risks highlighted in the White House’s AI safety commitments. They offer metrics for assessing LLM cybersecurity vulnerabilities, methods to evaluate the prevalence of insecure code suggestions, and tools to impede LLMs from generating malicious code or aiding in cyber attacks, aiming to diminish the occurrence of insecure AI-generated code and limit the utility of LLMs to malicious actors.
This announcement follows reports that Meta is experimenting with over 20 generative AI functionalities across its platforms, including Facebook, Instagram, Messenger, and WhatsApp. These features encompass various aspects of social media such as search, advertising, and business communication, according to Bloomberg News. Ahmad Al-Dahle, Meta’s vice president of generative AI, emphasized the company’s dedication to fostering a more positive community, enabling users to express themselves, and developing more beneficial products. Al-Dahle noted that these efforts are reflected in key performance indicators, leading to increased engagement, positive responses, and enhanced user experiences.
