Meta uses OpenAI’s GPT-4 alongside Llama to augment internal AI coding tool

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.

Are you a Meta employee or someone with insight or a tip to share? Contact Kali Hays securely through Signal at +1-949-280-0267 or at kali.hays@fortune.com.

How many degrees of separation are you from the globe’s most powerful business leaders? Explore who made our brand-new list of the 100 Most Powerful People in Business. Plus, learn about the metrics we used to make it.