As the market rapidly embraces AI, there is a need for organizations to foster responsible development and use of generative AI.
Llama models now have over 100 million downloads of Llama models, spurring Meta to consider its future impact. To instill trust in developers spearheading innovation, Meta has introduced Purple Llama. This umbrella project aims to foster responsible development with open generative AI models.
Meta has adopted a purple teaming approach for Purple Llama in a nod to cybersecurity strategies. Combining attack (red team) and defensive (blue team) postures, this collaborative method evaluates and mitigates potential risks. Purple Llama has focused on cybersecurity and input/output safeguards, with a commitment to expand its offerings in the near future.
Meta has also unveiled an industry-wide set of cybersecurity safety evaluations for Large Language Models (LLMs). Aligned with industry guidance and standards, these benchmarks– developed in collaboration with security experts–aim to address risks outlined in White House commitments. The tools provided include metrics for quantifying LLM cybersecurity risk, evaluating insecure code suggestions, and making it harder for LLMs to generate malicious code or aid in cyber attacks. Meta envisions these tools reducing the occurrence of insecure AI-generated code and diminishing the utility of LLMs to cyber adversaries.
Building on Llama 2’s Responsible Use Guide, Meta recommends thorough checks and filters for inputs and outputs to LLMs. To support this, Meta has released Llama Guard, a foundational model openly available to help developers avoid generating potentially risky outputs. With transparency in mind, Meta shares the methodology and results in a paper. Llama Guard, trained on publicly available datasets, enables the detection of potentially risky or violating content. The ultimate goal is to empower developers to customize future versions based on their requirements, facilitating the adoption of best practices and enhancing the open ecosystem.
Meta’s commitment to an open approach to AI is exemplified by its collaborative mindset. Partnerships with over 100 organizations, including AI Alliance, AWS, Google Cloud, IBM, Microsoft, and many more, signify a shared vision for an open ecosystem of responsibly-developed generative AI. Meta looks forward to continued collaboration in shaping the future of open and responsible AI development.