"Zuckerberg Reveals Meta's Development of Open Source AGI"

In a surprise Instagram Reel today, Meta CEO Mark Zuckerberg announced that the company is developing open-source artificial general intelligence (AGI).

Zuckerberg shared that Meta is consolidating its AI research teams—FAIR and GenAI—to pursue the goal of creating full general intelligence while emphasizing open-source accessibility.

“Are you ready for AI agents?”

“Our long-term vision is to build general intelligence, open-source it responsibly, and make it widely available so everyone can benefit,” wrote Zuckerberg in the caption. In the video, he elaborated, “It’s become clear that the next generation of services requires full general intelligence—including top-tier AI assistants, tools for creators, and solutions for businesses. This necessitates advancements in various AI domains, from reasoning and planning to coding and cognitive memory.”

Zuckerberg also highlighted the training of Llama 3 and the extensive compute infrastructure under development, including 350,000 Nvidia H100 GPUs expected by year-end. He discussed the intertwined future of AI and the metaverse, stating, “As people engage with AI throughout the day, new devices like smart glasses will be essential. These glasses will enhance interaction by allowing AI to perceive what you see and hear, ensuring constant availability for assistance.”

This announcement follows recent comments from OpenAI CEO Sam Altman at the World Economic Forum in Davos, Switzerland, where he moderated his stance on the existential risks of AGI. Meanwhile, Meta's chief scientist, Yann LeCun, has voiced doubts about the imminent arrival of AGI, suggesting it may not materialize within the next five years.

The prospect of Meta's open-source AGI comes shortly after reports that Llama and open-source AI 'won' 2023, fueling ongoing debates about the merits of open-source versus closed-source AI. This dialogue is particularly relevant following a recent Anthropic paper that warned of potential risks associated with open models, implying the existence of hidden 'sleeper agents.'