Meta announced in London that they’re gearing up to roll out Llama 3 soon, the upgraded version of their big language model that fuels AI assistants. It’s set to hit the scene in about a month.
Llama is Meta’s open-source large language models, available for free for research and commercial use.
Nick Clegg, Meta’s president of global affairs, mentioned that they’re aiming to launch Llama 3, their latest set of foundation models, in less than a month, or even sooner. He hinted at releasing various versions with different features and abilities throughout the year, starting shortly.
Chris Cox, Meta’s Chief Product Officer, said they’re going to use Llama 3 to fuel many products across Meta.
Meta has been racing to keep up with OpenAI. About a year ago, OpenAI released ChatGPT, which caught Meta and other major tech companies like Google off guard. ChatGPT quickly gained popularity, making AI conversations mainstream.
Meta has mostly been careful with AI, but people haven’t liked it much. Previous Llama versions got flak for being too restricted. Llama 2 came out in July 2023, but the first version never officially hit the public, though it still got out online.
Llama 3 is set to fix this issue. It’s broader and can handle more than earlier versions. Not only will it give better answers, but it’s also ready for more controversial questions. Meta hopes this will make users like it more.
Joelle Pineau, vice president of AI Research, said , “They want a Meta AI powered by Llama to be super helpful eventually. But there’s still a lot of work to do. They didn’t say how big Llama 3’s parameters are or show any demos. It’s estimated to have around 140 billion parameters, twice as many as Llama 2’s largest model.
Meta’s Llama series, created as open-source tools, signals a new way of thinking about AI’s growth. By doing this, Meta aims to win over developers who prefer open platforms over closed ones. However, they’re taking a cautious stance, particularly with other generative AI like image creation. Pineau mentioned they’re not ready to release Emu, their image generation tool, just yet.
Chris Cox said, “Meta’s Chief Product Officer, elaborated on the plan, indicating that Llama 3 will empower multiple products across Meta, such as their RayBan Meta glasses.”
Yann LeCun, Meta’s chief AI scientist, criticized the constraints of generative AI and expressed his faith in the next big thing. He believes in joint embedding predicting architecture (JEPA), a new method for training models and getting results. Meta has been applying this to enhance predictive AI, especially in image generation, for better accuracy.
Yann LeCun stated that “JEPA is the future of AI, not generative AI. He joked about needing to rename Chris’s product division to align with this new direction.”