On April 19th, Meta, the parent company of Facebook, unveiled the latest iteration of its open-source language model—Llama 3. This highlights the escalating battle among Silicon Valley's artificial intelligence contenders.
Meta introduced two versions of the Llama 3 model: one with 80 billion parameters and another with 700 billion parameters. Parameter count is a crucial metric for measuring the size and performance of AI systems, directly reflecting the model's complexity and computational power. Meta noted that even the version of Llama 3 with fewer parameters is on par with the largest version of Llama 2, which had 700 billion parameters, released last year.
Mark Zuckerberg, CEO of Meta, revealed that the upcoming large-scale version of Llama 3 will surpass 400 billion parameters, far exceeding previous versions. While Llama 2 utilized 20 trillion tokens during training, the large version of Llama 3 will employ over 15 trillion tokens, indicating broader data processing capabilities. Although OpenAI has not publicly confirmed the parameters or token count for GPT-4, industry estimates suggest parameters could reach 1.5 trillion. Meta's Chief Product Officer, Chris Cox, stated that this large-scale model will be officially released once training is completed. He added, "Based on benchmark test results we've seen so far, Llama 3 is undoubtedly the smartest and highest-performing large model available for free in the market."
Meta also plans to roll out more feature-rich versions of Llama 3 in the coming months, including versions supporting additional languages. This move demonstrates Meta's continuous investment in competing with rivals like OpenAI in the fiercely competitive AI market.
Meta intends to utilize the Llama 3 model to enhance its Meta AI assistant chatbots, which will play a more prominent role across Meta's various applications. Additionally, Llama 3 will be deeply integrated into the search functionality of Meta's primary platforms, including Facebook, Instagram, WhatsApp, and Messenger. The upgraded Meta AI will also be capable of providing simple summaries of linked content.
Looking back, OpenAI launched the ChatGPT AI chatbot in late 2022, kicking off the Silicon Valley AI race. Since then, an increasing number of companies, whether startups or established tech giants, have joined the competition.
In July last year, Meta publicly introduced the Llama 2 model and deployed it to support Meta AI chatbots on Instagram, WhatsApp, and Messenger in September. Unlike other companies in the industry, Meta chose to release its models as open-source software, allowing developers worldwide to use and enhance Llama 2 for free.
Debra Aho Williamson, Chief Analyst at Sonata Insights, pointed out that if Meta can attract billions of users to its AI products, the company may embed advertisements in the models and potentially expand into other paid services. She believes, "Most consumer-facing AI experiences may eventually incorporate some form of paid advertising, such as targeted ad results within chatbot conversations."
With the release of Llama 3, the functionalities of Meta AI will further expand, covering 13 English-speaking markets, including Australia and Canada. However, it's worth noting that the service will not initially launch in the European Union.
Cox added, "Meta AI will become even more powerful and intelligent, particularly for enterprise users, with the link summarization feature being highly valuable when handling lengthy content."
Furthermore, the image generation feature of Meta AI, Emu, will also receive upgrades, not only improving the speed and quality of generated results but also providing instant preview results when users input queries.
Cox stated, "In the realm of generative AI, Meta is committed to making these advanced technologies freely accessible to a wide range of users." He also emphasized that with the launch of larger versions of the Llama model, the company may consider adopting a monthly subscription model similar to OpenAI's ChatGPT Plus and Google One AI Premium.
It's worth noting that Meta used synthetic data during the training of Llama 3, leveraging text generated by other AI models. As tech companies face a growing scarcity of data for training AI models, they must seek new solutions to acquire or create more data.
