The AI boom has raised the stakes in the tech industry’s long running debate over the merits of open source versus closed source software. But for Facebook parent company Meta, “unlocking” large language models is the only way to go if AI is to reach its full potential.
“It’s ridiculous for me to think of a world where we enclose AI and have an API for intelligence,” Manohar Paluri, a vice president of generative AI engineering at Meta, said Tuesday at Coins2Day’s Brainstorm AI conference in San Francisco. “Progress for AI in the coming decades will only happen if we go the open route.”
Technically, Meta’s Llama large language model does not meet the definition of open source as defined by the Open Source Initiative in a definition updated only a month ago to grapple with a wave of AI models. But Llama is nonetheless among the largest and most powerful AI models that’s freely available for the general public to download, customize and create AI apps or tools with.
Paluri argued that the accessibility of Llama is key to its success so far, because “distribution” is one of four elements among all successful models today. The other three are, talent, compute power, and data.
Critics of open AI models have raised concerns about the technology falling into the wrong hands. Earlier this year, reports surfaced that Chinese researched had used Meta’s Llama technology to create an AI model for military use. Paluri did not address the safety issue when asked about it, emphasizing the benefits of open source instead.
Llama’s accessibility has helped Meta catch up with rivals OpenAI and Alphabet in the AI race, after initially getting caught flat footed by the advent of generative AI at the end of 2022.
Meta has said repeatedly in recent weeks that Meta AI, its generative AI chatbot and search bar, has about 600 million monthly active users. For comparison, OpenAI’s ChatGPT, which operates via a closed model, has 300 million users a week, its CEO Sam Altman said earlier this month.
It’s unclear how many of Meta AI’s reported 600 million monthly users are actively seeking to engage with the AI, versus simply using Meta services such as search, which use AI behind the scenes. Paluri declined to provide a more granular breakdown about the users when asked about it on stage, noting the difficulty of interpreting a user’s specific intent when accessing a feature and saying that technology works best when it’s “seamless.”
While ChatGPT was the chatbot that kicked off the current frenzy for generative AI and had over a year of lead time on Meta AI, Paluri said usage of Meta AI benefits from being put within nearly all of Meta’s apps, including WhatsApp, Instagram and Facebook, which collectively have more than 3 billion users each day.
“Where we started at the beginning of the year, Meta AI was just a fledgling part of what we’re doing,” Paluri said. “The tech is definitely improving day by day.”
Read more coverage from Brainstorm AI:
Microsoft’s biz dev boss on the big bets that could define the future of AI