Mark Zuckerberg has long touted Meta’s Llama AI model as a top contender in the generative AI space, positioning it as a competitor to industry giants like OpenAI and Google. While Meta has marketed Llama as a cutting-edge solution for AI-driven tasks, it appears the company is leaning on a different, more established model behind the scenes to fulfill its internal needs.
Meta’s internal coding assistant, Metamate, integrates both the Llama model and OpenAI’s GPT-4 to assist developers with programming tasks. The tool, which has been in use since early 2024, switches dynamically between the two models depending on the complexity of the query. According to sources familiar with the tool, Metamate is considered highly effective, rivaling other coding assistants like Microsoft’s GitHub Copilot.
Despite Zuckerberg’s strong claims that Llama is competitive with the best generative AI models available, Metamate’s reliance on GPT-4 reveals the limitations of Llama in certain scenarios. The AI tool is praised for its ability to handle basic coding tasks but falls short when faced with more complex engineering challenges. One source described its performance as “at least as good as an intern,” highlighting the gap in capabilities between Llama and more refined proprietary models like GPT-4.
The dual use of Llama and GPT-4 within Meta underscores a broader reality in the AI industry: while open-source models like Llama are making significant strides, proprietary models still hold a significant edge in specialized applications. Abhivyakti Sengar, a Senior Analyst at Everest Group, pointed out that Meta’s reliance on GPT-4 for certain tasks highlights the ongoing superiority of proprietary AI models in high-stakes, real-world use cases.