The tech giant, Meta, recently unveiled the next generation of their notable AI models: Llama 2. This innovative family of AI models has been explicitly developed to power several chatbots, including OpenAI’s ChatGPT and Bing Chat, along with other state-of-the-art conversation systems.
Trained on an assortment of publicly accessible data, Llama 2 is set to surpass the previous generation in terms of overall performance. The Llama model's successor is a significant development that offers a superior interaction capacity being compared with other chatbot-like systems.
Notably, the original Llama was accessible on request only, as Meta took strict precautions to limit its misuse. However, the Llama model eventually found its way across various AI communities, despite the intentional gatekeeping.
Contrarily, Llama 2 is open for research and commercial use in a pretrained form. Offering the convenience of optimization on various hosting platforms, such as AWS, Azure, and Hugging Face’s AI, the model guarantees a user-friendly experience. The introduction of Llama 2 is made possible due to an expanded partnership with Microsoft, making it optimized for Windows and devices equipped with Qualcomm’s Snapdragon system-on-chip. Qualcomm is also reportedly working on migrating Llama 2 to Snapdragon devices by 2024.
Llama 2 comes in two versions: the basic and Llama 2-Chat, engineered especially for two-way interactions. Both versions are available in various sophistication levels, defined by the range of parameters–7 billion, 13 billion, and a whopping 70 billion. The parameters, which are parts of a model learned from the training data, effectively determine the model's proficiency on a problem—in this case, text generation.
Llama 2 was trained on two million tokens, which imply raw text. This is nearly two-fold as compared to the original Llama, which was trained on 1.4 trillion tokens. Generally, a larger number of tokens results in more efficacy when dealing with generative AI models. Meta has stayed tight-lipped about the specifics of the training data, apart from disclosing that it is primarily in English, sourced from the internet, and emphasizes text of a factual nature.
This move embarks on a new chapter in the AI realm offering vast potential for no-code and low-code platforms, such as AppMaster, enabling users to utilize these advanced tools in a myriad of applications while making the development process swift and efficient.
