In the global panorama of artificial intelligence, a new major competitor is looming on the horizon. Mark Zuckerberg, CEO of Meta, has indeed announced his intention to develop advanced models of artificial intelligence, with the aim of achieving artificial general intelligence (AGI). The move could upend the global competition for AI supremacy, putting Meta in direct competition with OpenAI, whose offering of advanced language models is currently proprietary and not open to the public.
Mark Zuckerberg speaks
“It’s clear that future services need comprehensive general intelligence: to develop superior AI assistants, AI for creators and businesses, and much more,” Zuckerberg said in a Facebook post. “Achieving this requires advancements in all areas of AI: from reasoning to planning, from coding to memory and other cognitive abilities. The importance and opportunities offered by this technology are such that we must make it as accessible as possible.
AGI has long been a coveted achievement in the industry. This is the ability of machines to perform general tasks at the same level, or even better, than humans. Among the main players interested in this goal are OpenAI and Google DeepMind . However, the precise definition of AGI still remains a matter of debate. “I don’t have a concise, unambiguous definition,” Zuckerberg admitted in an interview with The Verge. “It is questionable whether general intelligence is comparable to human intelligence, or whether it is some sort of superintelligence of the future. For me, the key aspect is the breadth of the definition: intelligence includes many different abilities, including reasoning and intuition. »
The focus on AGI also attracts top talent in the industry, because "many of the brightest researchers want to work on the hardest problems," Zuckerberg added.
Meta's AI advances
In line with this strategy, Meta is intensifying its efforts. LLaMA and the subsequent Llama 2 family of large language models have formed the basis of many open source AI models. Meta is currently working on Llama 3, its next advanced language model. However, if Meta's models reach AGI, Zuckerberg has not committed to making them open source. “As long as it's safe and responsible, I think we'll tend to move toward open source,” he told The Verge.
As part of this new strategic direction, Meta is reorganizing the FAIR (Fundamental Research in AI) laboratory in collaboration with the AI products division, called GénAI . Joelle Pineau, who leads FAIR, and Yann LeCun, Meta's chief AI scientist, will report to Chris Cox, Meta's chief product officer.
Zuckerberg also announced that Meta was building a “massive IT infrastructure” to support its AI projects. The goal is to buy 350,000 H100 chips from Nvidia by the end of 2024, bringing Meta's total computing power to nearly the equivalent of 600,000 H100s, excluding other GPUs.
Research subsidiary Omdia expects Meta to continue to be a major buyer of Nvidia's AI chips. In fact, it is estimated that by 2023 Meta will have ordered at least 150,000 H100 chips to power its AI products.
Comments
Post a Comment