Meta, the parent company of Facebook, Instagram and Threads, has launched Llama 2, an advanced collection of AI models designed to enhance the capabilities of chatbots such as OpenAI’s ChatGPT and Bing Chat. This new generation of Llama models exhibits considerable improvements in performance compared to its predecessor, and unlike the previous version, Llama 2 is now readily available for both research and commercial purposes.
Users can now fine-tune and utilize Llama 2 on various platforms including AWS, Azure, and Hugging Face’s AI model hosting platform. Meta has optimized Llama 2 for Windows, smartphones, and PCs powered by Qualcomm’s Snapdragon system-on-chip, making it accessible to a wider audience.
Llama 2 comes in two versions: Llama 2 and Llama 2-Chat. The latter is tailored specifically for two-way conversations. These versions are further categorized into three levels of sophistication based on their parameters: 7 billion, 13 billion, and 70 billion. Those parameters refer to the learned components of a model from training data, which determine the model’s ability to generate text. Llama 2 was trained on an extensive dataset of two trillion tokens, nearly twice the size of Llama’s training data. Generally, having more tokens results in better performance for generative AI applications.
While Meta remains secretive about the specific sources of their training data, they emphasize that it mainly consists of publicly available web content in English, excluding Meta’s own products or services. This approach aims to prioritize text of a factual nature. Such discretion may stem from competitive and legal factors surrounding generative AI.
To evaluate Llama 2’s performance, Meta conducted a range of benchmarks. In comparison to its closed-source competitors, GPT-4 and PaLM 2, Llama 2 falls slightly behind but is deemed similarly “helpful” as ChatGPT by human evaluators. Meta claims that Llama 2 consistently performed well across approximately 4,000 assessments of “helpfulness” and “safety.” However, it is important to approach these results with caution as real-world scenarios may present challenges not captured by the benchmarks, such as coding and human reasoning.
Furthermore, like all generative AI models, Llama 2 exhibits biases. For instance, it tends to generate “he” pronouns more frequently than “she” pronouns due to imbalances in the training data. The model also displays a Western-centric bias, as evidenced by an overrepresentation of terms like “Christian,” “Catholic,” and “Jewish” in the training data.
In terms of toxicity benchmarks, the Llama 2-Chat models outperform the regular Llama 2 models based on Meta’s internal assessments. However, caution is advised as these models err on the side of excessive caution, often declining certain requests or providing an abundance of safety details in their responses.
Notably, the benchmarks do not encompass additional safety measures that can be implemented when using hosted Llama 2 models. Meta collaborates with Microsoft to utilize Azure AI Content Safety, a service that helps detect inappropriate content in AI-generated text and images, thus reducing toxic outputs from Llama 2 on the Azure platform.
Despite these safety precautions, Meta emphasizes the importance of responsible usage and compliance with their license and acceptable use policy in the whitepaper. They believe that openly sharing large language models like Llama 2 will contribute to the development of helpful and safer generative AI. They eagerly anticipate the creative applications that will emerge with the use of Llama 2.
Given the rapid pace of the internet, it won’t be long before we witness the impact of Llama 2 in various domains. Meta’s Llama 2 offer promising prospects for the future of AI-powered chatbots and generative AI as a whole.
The whytry.ai article you just read is a brief synopsis; the original article can be found here: Read the Full Article…
