Mark Zuckerberg has consistently championed Meta’s Llama AI model as a leader in generative AI technology, positioning it as a strong competitor to OpenAI and Google. However, behind the scenes, Meta is complementing Llama with a rival AI model to meet its internal needs.
Meta’s internal AI-powered coding assistant, Metamate, uses both Meta’s Llama model and OpenAI’s GPT-4 to help developers and employees with coding tasks, reported The Fortune. The tool, which has been operational since early 2024, dynamically switches between the two models depending on the query, according to a current and a former Meta employee who spoke anonymously to The Fortune.
“It’s one of the better GenAI things they’ve done,” one of the sources told The Fortune about Metamate, which is described as being on par with other AI coding tools like Microsoft’s GitHub Copilot. Despite Zuckerberg’s assertion that Llama is competitive with the most advanced models, Metamate’s reliance on GPT-4 underscores the challenges of creating a one-size-fits-all solution for generative AI.
The sources added that while Metamate is effective for basic tasks, it struggles with more complex engineering work. One person likened its performance to “at least as good as an intern.”
“While Llama marks a significant step forward for open-source AI, Meta’s internal reliance on GPT-4 highlights the enduring edge of proprietary models in specific high-stakes, real-world applications,” said Abhivyakti Sengar, Senior Analyst at Everest Group.
Llama’s promises vs practical realities
The dual reliance on Llama and OpenAI raises questions about Meta’s broader AI ambitions. Zuckerberg has positioned Llama as a key player in what he calls the “model wars,” touting its open-source framework as a competitive advantage.
When Llama’s latest version was released mid-year, Zuckerberg stated it was “competitive with the most advanced models and leading in some areas.” He claimed Llama would be “the most advanced in the industry by next year.”
However, the integration of GPT-4 into key Meta tools suggests that Llama, while powerful, still has limitations, particularly in addressing diverse queries and providing robust support across various use cases.
Meta’s hybrid AI approach
The dual reliance on Llama and GPT-4 reflects Meta’s nuanced strategy in navigating the fast-paced AI landscape. While Llama remains the cornerstone of Meta’s AI vision, the inclusion of GPT-4 underscores the practical realities of building tools that meet the diverse and evolving needs of employees, developers, and end users.
This hybrid approach allows Meta to leverage the strengths of multiple AI models, but it also raises questions about how far the company has come in realizing its ambition to lead the generative AI market. As Meta continues to innovate with Llama, its quiet dependence on GPT-4 reveals the complexities of staying competitive in an industry defined by rapid technological advancements.
“The smartest product managers and founders will be those who keep egos aside to leverage the best of each model,” said Vaibhav Bansal, Vice President at Everest Group. “While both these LLMs have been found to be very good for coding assistance, each has varying strengths and weaknesses in terms of accuracy, resource utilization, reasoning, creativity, and multi-lingual support; which is probably what Meta is cognizant of and is trying to optimize for Metamate.”
Expanding AI’s role beyond Meta
Meta’s hybrid approach and use of OpenAI’s technology isn’t confined to internal operations. The Chan Zuckerberg Initiative (CZI), the philanthropic organization led by Zuckerberg and his wife Priscilla Chan, is developing an AI tool for education, built on OpenAI’s ChatGPT. This effort includes customizations for its specific needs, known in the industry as a “wrapper” around the core AI, The Fortune report added.
Notably, OpenAI CEO Sam Altman joined an AI advisory board at CZI earlier this year, signaling deeper collaboration between the two entities.
Meta’s reliance on a rival model like GPT-4 highlights the complexities of staying competitive in the rapidly evolving AI landscape. While Llama remains a key focus for Meta’s generative AI strategy, the company’s pragmatic use of other models reflects the ongoing challenges of building and deploying AI systems that meet both technical and business needs.
As Meta continues to innovate with Llama, the quiet integration of GPT-4 into its operations serves as a reminder that even industry leaders often require external support to stay ahead in the AI arms race.
“It serves as a clear reminder that no single model dominates all domains, and the future of generative AI lies in a pragmatic, hybrid approach that integrates diverse best-in-class solutions,” noted Sengar.
A request for comment from Meta was not returned.