At an event in London on Tuesday, Meta confirmed that it plans an initial release of Llama 3, a next-generation large-scale language model used to power generative AI assistants, within the next month. .
This confirms a report published by Information magazine on Monday that Meta is nearing launch.
“Within the next month, in fact within a much shorter period of time, and hopefully within a very short period of time, we hope to begin rolling out Llama 3, our new suite of next-generation infrastructure models,” said Meta's global president. said one Nick Clegg. He described what sounds like the release of several different iterations or versions of the product. “There are a lot of different models coming out with different features and different versatility. [released] It starts really quickly, sometime this year. ”
Meta Chief Product Officer Chris Cox added that the company plans to use Llama 3 to power multiple products across Meta.
Meta surprised other big tech companies like OpenAI and Google when it released ChatGPT over a year ago, and the app went viral, turning generated AI questions and answers into a mainstream everyday experience. We've been working hard to catch up with OpenAI.
Meta has primarily taken a very cautious approach to AI, which has not been well received by the public, and previous versions of Llama have been criticized for being too restrictive. (Llama 2 was released to the public in July 2023. The first version of Llama was not released to the public, but was still leaked online.)
Llama 3, which is broader in scope than previous versions, is expected to address this issue, with the ability to not only answer questions more precisely, but also address a wider range of questions that may include more controversial topics. It has been. We hope that this will make this product a popular choice for users.
“Our long-term goal is to make Meta AI, powered by Llama, the world's most helpful assistant,” said Joelle Pineau, vice president of AI Research. “There remains quite a bit of work to get there.” The company did not say what size parameters it is using in Llama 3, nor did it provide a demo of how it works. I have not. It is expected to have around 140 billion parameters, compared to 70 billion in his largest Llama 2 model.
Most notably, Meta's Llama family, built as open source products, represents a different philosophical approach to how AI should be developed as a broader technology. In doing so, Meta hopes to broadly favor developers rather than a more proprietary model.
However, Meta seems to be taking a more cautious approach, especially when it comes to other generation AIs other than text generation. Pineau said the company has not yet released its image generation tool, Emu.
“No matter the creative context, latency is critical, along with safety and ease of use, to produce images you can be proud of,” said Cox.
Ironically, or perhaps predictably (heh), Meta is working on launching Llama 3, but there are quite a few people within the company who are skeptical of generative AI.
Yann LeCun, a prominent AI scholar and chief AI scientist at Meta, said he is taking a swipe at the limits of generative AI overall and betting on what happens next. He predicts it will be the Joint Embedded Prediction Architecture (JEPA). This is a different approach to both model training and result generation that Meta is using to build more accurate predictive AI in the field of image generation.
“The future of AI is JEPA. It's not generative AI,” he said. “We need to rename Chris' product division.”