The Problem with Meta’s Llama 3 Launch; JD Vance Likes Open-Source

While sports fans gear up for the Summer Olympics later this month, artificial intelligence enthusiasts are eagerly anticipating another major event: the July 23 release of Meta Platforms’ flagship Llama 3 large language model, which the company offers for free.

Meta already released smaller versions of Llama 3, which AI app developers have raved about, and many of them expect the larger, flagship version—which will have 405 billion parameters (the “settings” that determine how a model responds to questions)—to rival OpenAI’s best models. That might help answer the multibillion-dollar question of whether open-source models will steal thunder, and dollars, from OpenAI and other providers of closed-source models. 

But there’s a looming problem that could turn hope to anguish for Meta and other proponents of open-source AI. App developers up till this point have been able to run open-source models using one set of eight, connected AI server chips, known as a node. The smaller versions of Llama 3 run fine that way, but the larger model likely won’t be able to fit on a single node, requiring multiple nodes instead.