AI Technologies
Anit Vallurupalli
August 23, 2024
How Meta delivers the most innovative Generative AI tool yet.
When Artificial Intelligence comes into mind, an overwhelming influence of forms of generative A.I. through chat models are prevalent. The endless applications of such a revolutionary system of A.I. Truly an innovation that is the closest to the human mind and thus able to learn from every single interaction executed. With ChatGPT being a prominent and the most popular form of some of the most applicable and accessible technologies as of today, it is both a technological and industrial imperative that we ask: What is Llama 2, Facebook’s answer to ChatGPT? Llama, or the Large Language Model Meta Artificial Intelligence, is a second-generation open-source LLM (large language model) from Meta. Similarly to OpenAI, it can be used to build chatbots and many more A.I. applications for the future.
A reality we have to accept is that the future lies within corporations. Corporations that focus on new innovations to make revenue but also create outstanding technologies. Llama 2 is reflective of this factor as it has been co-developed by Meta and Microsoft. This gives away to several more applications that everyone uses. Llama 2 will be available but not just restricted to Windows PCs, phones, and laptops powered by popular semiconductor manufacturers. This new business model of mega-corporations working together to develop such models powers innovation and facilitates an ever-increasing technologically influenced environment. This has further been proved by its performances as this second-generation model has been trained on 40% more data than the original Llama, which interprets into a more precise and powerful LLM that gets ever so close to providing human-like responses.
Not only is Llama 2 a powerful open-source LLM, it is a stark competition to ChatGPT, or perhaps the most influential innovation of the decade. Because of the sheer popularity of OpenAI’s model, it is hard to notice any other competition. But this skewed perspective is ignoring the stronger alternatives that are being developed, with Llama 2 being one of them. When put into perspective, Llama 2 strongly compares to both open-source and closed-source chat models.
Llama 2 Chat is a fine-tuned version of Llama 2 that is used primarily for dialogue use cases. It is trained in a systemized framework. The process for training begins with pretraining in which publicly available online sources are used. Then, an initial version of Llama 2 Chat is created through supervised fine-tuning. Following this initial model, a precept of generative A.I. is executed. The chat model is iteratively improved and refined using RLHF (Reinforcement Learning using Human Feedback. This methodology specifically uses rejection sampling (basic sampling technique to generate observations from distributions) and Proximal Policy Optimization (PPO - policy gradient method, or methods that push up probabilities of actions that lead to higher return, optimization model for reinforcement learning). The model below best helps visualize the training of Llama 2 Chat:
To recap, Llama 2 is a strong contender not only for future technological innovation as it acts as a competitive open-source platform to normalize accessibility to powerful generative A.I., but it is also an accomplishment and a sign of what is to come as tech companies keep funding the ever-changing industry of innovation. And to give context to our first question, it is a response and a firm one. A response to ignite a new competition for generative A.I. and to create further potential to this technology.
Works Cited:
Touvron, Hugo, et al. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv, 19 July 2023. arXiv.org, https://doi.org/10.48550/arXiv.2307.09288.
“What Is Llama 2: Meta’s AI Explained.” Dexerto,
https://www.dexerto.com/tech/what-is-llama-2-2224223/. Accessed 4 Nov. 2023.
