Meta Unveils LLaMA 2: A New Open-Source AI Model
Meta is giving away it’s AI tech to compete with other major players in the AI chatbot space, such as OpenAI's ChatGPT, Microsoft, and Google.
In a bold move, Meta, the tech giant formerly known as Facebook, has announced the release of LLaMA 2, its latest open-source language model. In short, Meta is giving away it’s AI tech to compete with other major players in the AI chatbot space, such as OpenAI's ChatGPT, Microsoft, and Google.
LLaMA 2, the successor to Meta's earlier research-restricted version, has been designed to cater to a wide range of users, making it freely accessible for anyone to build commercial products. Unlike ChatGPT, available directly through OpenAI's website, LLaMA 2 must be downloaded from Meta's launch partners, including Microsoft Azure, Amazon Web Services, and Hugging Face.
Ahmad Al-Dahle, Vice President at Meta leading the generative AI work, emphasizes that this move benefits the entire AI community, allowing developers to choose between closed-source or open-source approaches that suit their specific applications. However, there are some caveats to consider.
One notable concern is more information about the data set used to train LLaMA 2. A research paper shared exclusively with MIT Technology Review reveals that Meta cannot guarantee the exclusion of copyrighted works or personal data from the model. LLaMA 2, like other large language models, also needs help with issues such as generating falsehoods and offensive language.
Meta, however, sees this open-source initiative as an opportunity to learn and improve. By releasing the model to developers and companies, the company hopes to gain valuable insights into making their models safer, less biased, and more efficient.
Percy Liang, Director of Stanford's Center for Research on Foundation Models, acknowledges that LLaMA 2 is not directly comparable to OpenAI's state-of-the-art GPT-4 but asserts that, for many use cases, the customizable and transparent nature of LLaMA 2 can lead to faster product and service development.
Despite Meta's past gaffes, including the short-lived Galactica language model and a leaked version of the previous LLaMA model, the company has significantly improved LLaMA 2's safety. Meta employed machine learning techniques and trained the model on 40% more data than its predecessor to achieve this. The training data came from online sources and a carefully fine-tuned dataset shaped with feedback from human annotators to ensure more desirable behavior. Meta claims that user data from its platform was not used, and data from sites containing personal information was excluded.
Despite these precautions, LLaMA 2 is not entirely free from harmful language generation. Meta has decided not to remove toxic data from the model, believing that it aids in better detecting hate speech. The company fears that eliminating such data might inadvertently lead to filtering out certain demographic groups.
Meta's commitment to openness has been met with excitement among researchers. The open-source nature of LLaMA 2 allows external researchers and developers to analyze it for biases, ethics, and efficiency, ultimately making it safer than proprietary models.
Have you tried out LLaMA 2 yet? If so, let us know what you think in the comments below!