Wanism’s Newsletter
What happened in tech that actually mattered, and what did it mean?
Meta Releases Llama 3.1
Share your love
Meta released the Llama 3.1 large language model (LLM) on July 23. This new model is undoubtedly a game-changer in open-source AI.
The Tug-of-War Between Open Source and Closed Source
This model comes in three versions: 8 billion, 70 billion, and 405 billion parameters, with the debut of the 405 billion parameter version being particularly noteworthy. Llama 3.1’s performance rivals closed-source models and even outperform them in certain aspects, signaling a new phase in the competition between open-source and closed-source AI.
Performance
To gauge Llama 3.1’s prowess, a series of benchmark tests are necessary for comparison. In this face-off, GPT-4 and Claude 3.5 Sonnet, as current frontrunners, naturally become the leading contenders. Here are the results of several critical tests:
| Test Name | Test Content | Llama 3.1 | GPT-4 | Claude 3.5 Sonnet |
|---|---|---|---|---|
| MMLU | College-level knowledge test | 88.6 | 88.7 | 88.3 |
| HumanEval | Programming test | 89.0 | 90.0 | 92.0 |
| GSMK | Mathematics test | 96.8 | 96.1 | 96.4 |
| ARC Challenge | Abstraction and Reasoning | 96.9 | 96.7 | 96.7 |
Among the 15 benchmark tests listed by Meta, the 405 billion parameter version of Llama 3.1 emerged victorious in 7 tests, Claude 3.5 Sonnet in 6, and GPT-4 in 3. These figures demonstrate that Llama 3.1 has reached a level on par with top-tier closed-source models, even gaining a slight edge in specific domains.
The emergence of Llama 3.1 marks a crucial shift from open-source models being slightly inferior to closed-source models to standing shoulder-to-shoulder with them. This development undoubtedly puts immense pressure on companies like OpenAI and Anthropic. LLMs could become a commodity if closed-source models fail to maintain a consistent lead. This pressure might compel OpenAI to accelerate the development of GPT-5 to retain its leading position in the AI field.
Meta’s Advantages: The User Base
Meta plans to integrate Llama 3.1 into its Meta AI Chatbot and has announced its goal to surpass ChatGPT as the most widely used chatbot by year-end. Achieving this objective relies on Llama 3.1’s technical capabilities and Meta’s massive user base.
Meta’s rise in the AI realm is reflected in its LLM’s technical prowess and, more importantly, in its ownership of the world’s largest user base. Facebook Messenger and WhatsApp boast over 1 billion active users, while Facebook’s global user count soars to 3 billion. This enormous user base provides Meta with an unparalleled advantage in AI applications.
The Importance of Platform Advantage
Just as Apple holds a significant position in the AI field thanks to its iPhone ecosystem, Meta’s social media platforms support its AI strategy. Even if Meta’s AI technology may lag behind competitors in some aspects, its control over the daily interactions of billions of users makes it an indispensable player in the AI arena.
With the dual strength of AI technology and user base, Meta is gradually establishing its key position in the AI field. As Llama 3.1 is launched and continuously improved, it’s foreseeable that Meta will play an increasingly important role in future AI competition, bringing more innovation and transformation to the entire industry.
