Meta uses competitor OpenAI’s GPT-4 alongside its Llama LLMs for internal development practices.
This is according to business magazine Fortune, based on sources. Meta is positioning its own Llama-LLMs as a strong alternative to other tech companies’ models. In the promotion, CEO Mark Zuckerberg emphasizes that Llama’s open-source approach offers a significant competitive advantage over “closed” LLMs from OpenAI, Google, and Anthropic, among others.
Use OpenAI LLMs.
Therefore, it would seem obvious that Meta would use only its own models internally. Yet, this does not appear to be the case. According to Fortune, Meta internally uses a generative AI tool for coding called Metamate, similar to GitHub Copilot or Anthropic Sonnet. This tool runs on Meta’s Llama-LLMs but additionally uses OpenAI’s GPT-4 to generate answers and code. GPT-4 was reportedly added early this year.
According to sources, the combination with GPT-4 improves the quality of the answers and code Metamate generates, depending on the query and employee interaction with the tool. The latest version of Metamate is trained on billions of tokens and contains four times more computer code than the previous edition.
Although employees are reasonably satisfied with Metamate, the tool is described as being at the level of a human intern. The system performs well in developing simple code but falls short in complex programming tasks.
ChatGPT also used
OpenAI’s technology is not limited to Metamate. According to Fortune, Mark Zuckerberg and Priscilla Chan’s philanthropic foundation, the Chan Zuckerberg Initiative, uses ChatGPT for its work.
The foundation is currently developing an educational generative AI tool based on ChatGPT. This tool will feature a self-developed shell or “wrapper” with additional AI functionality, but the core remains ChatGPT.
Meta did not comment on the possible use of competitors’ generative AI tools alongside or in conjunction with its own Llama LLMs. However, using multiple LLMs side by side is not uncommon, as they can often reinforce each other to generate better answers.