Meta Platforms, the parent company of Facebook, has recently announced Llama 2, a groundbreaking source-available family of AI language models. A notable feature of Llama 2 is its commercial license, allowing integration into commercial products, unlike its predecessor. Ranging from 7 to 70 billion parameters, Llama 2 has reportedly outperformed open-source chat models on various benchmarks. It is now available on Microsoft Azure and AWS.
The Llama 2 “pretrained” models, trained on an impressive 2 trillion tokens with a context window of 4,096 tokens, offer a solid foundation for fine-tuning and chat applications. Though it may not match the performance of OpenAI’s GPT-4, Llama 2 is highly regarded as a powerful and source-available model. The release of Llama 2 follows the leak and widespread use of its precursor, LLaMA, leading to a burgeoning underground LLM development scene.
While open-source AI models like Llama 2 have received significant support, they have also faced criticism due to potential risks associated with misuse, spam, and disinformation. Meta’s commitment to transparency and access has been praised, but critics emphasize concerns about the lack of data transparency in training these models, as unauthorized scraping of the internet may impact privacy and commercial interests.
The decision to release Llama 2 under a commercial license allows for broader use, but its license restrictions have led to debates about whether it qualifies as “open-source” according to the Open Source Initiative’s definition. Industry observers argue that the license limitations, such as restrictions for users with over 700 million daily active users and prohibitions on using outputs to improve other LLMs, deviate from traditional open-source principles.
Despite the controversy, Meta remains committed to an open approach to AI development, believing that it fosters transparency, encourages economic competition, and democratizes access to AI technologies. The company’s Statement of Support for Meta’s Open Approach to Today’s AI has garnered endorsements from industry executives and educators, emphasizing the value of responsible and open innovation in AI.
With Llama 2’s release, the AI community awaits the potential benefits of an accessible, commercially usable language model. However, concerns about data transparency and license restrictions highlight the need for ongoing discussions about the responsible development and deployment of large language models.