Meta's Dual AI Strategy: Llama Promotion and GPT-4 Integration in Internal Tools

Exclusive: Mark Zuckerberg publicly praises Meta's Llama AI, but also uses rival GPT-4 to improve an internal AI coding tool

Despite Mark Zuckerberg hailing Meta's Llama AI model as among the best in tech, his company is happy to also use a rival when needed. Meta's internal coding tool, Metamate, incorporates OpenAI's GPT-4 model, alongside Meta's own Llama model, to make the tool more useful for developers and other Meta employee who use it, two people who have used the tool told Fortune. OpenAI's model has been a part of Metamate since at least early this year. The people, a current and a former Meta employee who requested anonymity for fear of retribution from Meta, noted they find Metamate helpful. It appears to pull answers to coding questions or prompts from Llama or GPT-4, they said, depending on the kind of query or how an employee uses the tool. "It's one of the better genAI things they've done," one of the people who used the tool said of Meta's efforts in generative AI. Metamate is similar to other AI coding tools like Microsoft's GitHub CoPilot or Anthropic's Claude Sonnet. Separately, the Chan Zuckerberg Initiative, the large philanthropic organization run by Zuckerberg and his wife, Priscilla Chan, is also an OpenAI customer, two people familiar with that entity told Fortune. CZI is developing an "educational" genAI tool based on ChatGPT with some additional customization, or what's often referred to as a "wrapper" of genAI tech. Earlier this year, Sam Altman, CEO of OpenAI, was named to a new AI advisory board at CZI. A spokesman for Meta declined to comment. Representatives of CZI did not respond to emails seeking comment. While it's not unusual for a company deploying genAI tools for their workforce to rely on more than one model, Zuckerberg has positioned Meta as a key player in the model wars. It's a fight most companies are waging against OpenAI, which released ChatGPT two years ago and kicked off a race to catch up in order to gain a foothold in what's widely considered to be a new era of tech. Zuckerberg has spent the last year marketing Llama as an "open source" alternative that's as good or better than completely closed models, like those from OpenAI and Google. When Meta released a new version of Llama mid-year, the CEO said it was" competitive with the most advanced models and leading in some areas." By next year, Zuckerberg declared, the Llama model will be "the most advanced in the industry." Nevertheless, a tool like Metamate, and its underlying Llama model, still needs the aid of another like GPT-4 to perform well. And it doesn't matter that Llama is one of the world's largest models, with the most recent version pretrained on trillions of tokens, the individual bits of information fed into an LLM, and four times more computer code than it's predecessor. One of the sources who spoke to Fortune agreed that the current version of Metamate, originally dubbed Code Compose, is helpful in technical work at Meta, noting it's "at least as good as an intern." Translation: It's good at basic coding but not for more complex engineering work. However, a side effect of AI's helpfulness is that top executives think they need fewer workers. "It's part of what spurs layoffs," the person added. Indeed, Meta has incrementally laid off scores of workers this year after cutting more than 20,000 employees between 2022 and 2023.

Meta quietly leans on rival GPT-4 despite Zuckerberg's bold Llama claims

Even as Meta touts its Llama model, the company is incorporating OpenAI's GPT-4 to enhance internal tools and philanthropic ventures. 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.