Facebook’s parent company, Meta, has made a groundbreaking announcement at the Microsoft Inspire conference by unveiling its new AI model called LLaMA 2 (Large Language Model Meta AI). Unlike its predecessor, LLaMA 2 is not only open source but also freely available for commercial use. This development has significant implications for the world of generative AI, as enterprises now have another option to explore in their AI endeavors, at no cost.
Meta’s decision to release LLaMA 2 as open source has garnered strong support from Microsoft, further intensifying competition in the rapidly evolving field of large language models (LLMs). While other players like OpenAI and Cohere offer proprietary solutions, LLaMA 2 stands out as a free and accessible alternative.
The anticipation surrounding LLaMA 2 had been building for weeks, with US senators questioning Meta about its availability. The first iteration of LLaMA was licensed exclusively for research purposes, but leaked model weights prompted controversy and a government inquiry. With LLaMA 2, Meta aims to leave the past behind and unleash a more powerful model that has broader usability, potentially causing a stir in the LLM landscape.
One notable aspect of LLaMA 2’s release is its availability on Microsoft Azure, which is also home to OpenAI’s GPT-3/GPT-4 LLM family. Microsoft’s investments in both Meta’s former company, Facebook, and OpenAI indicate the company’s commitment to advancing AI technology.
Meta’s founder and CEO, Mark Zuckerberg, expressed his excitement about LLaMA being open source, highlighting Meta’s long-standing history of contributions to the open-source community, particularly in the field of AI through the PyTorch machine learning framework. Zuckerberg emphasized that open source not only drives innovation but also enhances safety and security by enabling more scrutiny and collaboration.
Yann LeCun, Meta’s VP and chief AI scientist, took to Twitter to celebrate the open source release of LLaMA 2, predicting that it will revolutionize the LLM market. LeCun revealed that LLaMA 2 will be available on Microsoft Azure, AWS, Hugging Face, and other providers.
LLaMA, which is based on the transformer architecture, is an auto-regressive language model. The first version, LLaMA 1, was unveiled by Meta in February and boasted 65 billion parameters, enabling it to tackle various generative AI tasks.
In contrast, LLaMA 2 offers different model sizes, including 7, 13, and 70 billion parameters. Meta claims that LLaMA 2 has been trained on a significantly larger dataset than its predecessor, with a context length expanded to 2 trillion tokens, twice that of LLaMA 1.
Notably, LLaMA 2 prioritizes both power and safety. The model undergoes a multi-stage process of supervised fine-tuning (SFT) after pretraining with publicly available data. Furthermore, it benefits from a Reinforcement Learning from Human Feedback (RLHF) cycle, adding an extra layer of safety and responsibility.
Meta’s research paper on LLaMA 2 provides detailed insights into the safety measures implemented and addresses concerns regarding transparency and potential bias. The paper emphasizes the importance of understanding the pretraining data to improve transparency, mitigate potential issues, and ensure appropriate model use.
With the release of LLaMA 2, Meta has positioned itself as a major player in the open-source LLM space. Its availability on Microsoft Azure, paired with the strong endorsements from CEO Mark Zuckerberg and AI scientist Yann LeCun, makes LLaMA 2 a formidable competitor for other commercially licensed LLMs. As the AI landscape continues to evolve, enterprises now have access to a powerful and freely available AI model that has the potential to reshape the industry.
