Meta CEO Mark Zuckerberg's interview with The Verge's Dwarkesh Patel sheds light on the company's AI plans, including Llama-3 and beyond

Meta CEO Mark Zuckerberg recently sat down for a podcast interview with The Verge's Dwarkesh Patel to discuss the company's plans for AI, particularly regarding its new Llama-3 technology. During the interview, Zuckerberg shared details about the current capabilities and future potential of Meta AI, the company's AI assistant.

One of the key highlights of the interview was Zuckerberg's assertion that Meta AI is currently the most intelligent, freely available assistant that people can use. While this claim is sure to spark debate among experts in the field, Zuckerberg argued that with the release of Llama-3, Meta AI has taken a significant leap forward in terms of intelligence and capabilities.

Zuckerberg also emphasized the importance of creating an AI assistant that can handle a wide range of tasks and provide users with a multi-modal, multilingual experience. To achieve this, Meta is working on developing AI that can understand and generate 3D content, as well as emotional understanding and context.

Another key discussion point was the technical aspects of Llama-3 and how it improves upon previous versions. Zuckerberg noted that Llama-3 is trained in three sizes, including an 8-billion parameter version, a 70-billion parameter version that was released recently, and a massive 405-billion parameter version currently in training. He claimed that the larger models have already demonstrated impressive capabilities, with the 405-billion parameter model already reaching around 85 million multiplicative linguistic universes (MMLU).

The CEO also touched on the challenges Meta faced earlier in its AI journey and how the company is addressing them. Specifically, he referred to the push to recommend unconnected content and the lack of sufficient compute, which resulted in the company's order of double the GPUs it required.

When asked about what Meta AI's future holds, Zuckerberg discussed the potential for an AI agent that can replace engineers at Meta and revolutionize the way people interact with technology. However, he clarified that the goal is not to replace people but to enhance their capabilities, hoping for a 10x increase in productivity.

Zuckerberg also shared his thoughts on the future of AI, arguing that it will become a fundamental and transformative technology, akin to the invention of computing. He believes that AI will enable the creation of new apps and empower people to achieve their goals more efficiently.

On the topic of open sourcing AI models, Zuckerberg maintained his pro-open-source stance but also left room for discretion. He expressed his belief that open sourcing is generally good for the community and Meta, but the company may opt not to release certain models if it believes they are too powerful or unsafe to be shared. This stance aligns with his previous comments regarding AGI (artificial general intelligence).

Overall, the interview offered a comprehensive look at Meta's AI ambitions and the steps it is taking to advance its AI technologies. With the release of Llama-3 and future developments, Meta is poised to make significant contributions to the AI community and shape the future of AI-powered experiences.