Meta accelerates into generative artificial intelligence with Llama 3

Meta unveiled this Thursday the new version of Meta AI, its generative artificial intelligence (AI) assistant, which goes from shadow to light across all its applications: Facebook, Instagram, Messenger and WhatsApp. Until now discreet, the tool which answers users' questions (like ChatGPT) will appear more prominently on the Californian group's networks and messaging services. Meta AI is also more efficient thanks to the new version of the underlying language model, Llama 3, also launched by the company on Thursday.

“We believe that Meta AI is now the smartest artificial intelligence assistant you can use freely,” said Mark Zuckerberg, the boss of the world leader in social networks, in a video on Instagram. The technological giants have been engaged for more than a year in a frantic race to develop and deploy generative AI (production of texts, images and other content, upon simple request in everyday language).

In the lead, OpenAI, which launched this new technological wave with ChatGPT, its main investor Microsoft, and its competitor Google, are competing with conversational robots (“chatbots”) and other virtual assistants to help humans do online research. , to create content or even educate their children. Thanks to Llama 3, Meta AI is supposed to give better answers to user queries and generate images more quickly (“immediately, it’s pretty crazy!” commented Mark Zuckerberg). And it has access to real-time information via Google and Bing, Microsoft's search engine.

The assistant presented in September will also be easier to use on a dedicated website (meta.ai), on Ray-Ban Meta connected glasses and on all applications, in more countries than before. “We integrated it into the search box at the top of Facebook, Instagram, WhatsApp and Messenger, so that whenever you have a question, you can just ask it there,” insisted Mark Zuckerberg. Meta's generative AI tools are less well-known and used by consumers than those of its competitors, according to Sonata Insights analyst Debra Williamson, but the company "still has time to catch up."

“Thanks to its social networks it has a massive user base to test AI experiments,” she explains for AFP. “By making several features available free of charge, it will be able to quickly assess those to which its users gravitate.” The Californian group had to invest massively to join the race for generative AI and adopted a different strategy: its language model is “open source” (free access to programming code) for companies and researchers. Meta thus intends to democratize this powerful technology and reduce the risks associated with its use. “Our goal in the near future is to make Llama 3 multilingual and multimodal, capable of integrating more context and (...) better in reasoning,” detailed the company.

With each release of new language models, companies assure that they are more efficient, more secure, and that they “hallucinate” less than before, according to the term used to designate their sometimes incoherent remarks. “Meta AI claims to have a child in a New York public school and shares her child's experience with teachers... in response to a personal question in a private parents group on Facebook,” it said. indignant example on X Tuesday Aleksandra Korolova, professor of computer science at Princeton, with a screenshot of the Meta AI message in support. At the beginning of March, Anthropic, a competitor to OpenAI, launched its new series of Claude 3 models, including the most advanced, Opus, which according to the start-up shows “the extreme limits of what is possible with generative AI”.

All these companies are trying to achieve so-called “general” AI, that is, having cognitive abilities similar to those of humans. They are therefore investing tens of billions of dollars in sophisticated chips and research and development. But “most AI experiences for consumers will eventually include some form of paid advertising,” says Debra Williamson. “From a business perspective, Meta has a huge advantage when it comes to building ad-supported services,” while OpenAI has no experience in this area, and “depends on paid subscriptions to ChatGPT and sales of business services,” she emphasizes.