Muse Wonder: Zuckerberg and Wang's $100B AI

Open original

Mark Zuckerberg and Alexandr Wang discuss Meta's new Muse Wonder AI agent, detailing how they trained it for multimodal perception and agentic capabilities while committing $100B+ to frontier model development without guaranteed success. The conversation covers real use cases—Zuckerberg's MMA coaching agent and Wang's tea-sourcing agent—and Meta's bet that personal AI agents will become the dominant product form factor.

  • Meta built Muse Wonder specifically for personal agent use cases, prioritizing multimodal capabilities (video, audio, vision) over pure text/coding, which diverged from industry consensus that coding was all that mattered
  • Zuckerberg uses Muse as a real-time MMA coach that processes gym camera feeds and generates tactical feedback within 60-second breaks between rounds, demonstrating the model's video processing and domain knowledge
  • Meta committed $100-200B in CapEx before validating whether the approach would work, betting that superintelligence and agentic commerce will drive future product categories beyond smartphones

Mark Zuckerberg and Alexandr Wang discussed Meta's new Muse Wonder AI agent during a podcast with John Collison. Meta is betting $100+ billion on AI infrastructure scaling past five gigawatts, and they made this massive commitment before even knowing if the models would work. Muse is apparently growing faster than any previous AI agent or app. Wang's personal favorite use case is having his Muse agent source obscure teas from farms in Taiwan and Japan on a weekly basis. Zuckerberg's preferred application is more demanding: he uses Muse to coach him during MMA training by processing video feeds from cameras in his gym and delivering feedback between rounds—a tight constraint since he only gets 60 seconds between rounds. The agent had to independently figure out how to extract still frames from video files and build an entire video processing pipeline without explicit instruction, then optimize itself when Zuckerberg complained about latency. The conversation revealed how Muse works as a product. Zuckerberg explained that the agent essentially writes code behind the scenes without users seeing it—people will end up writing tons of code through natural interaction without realizing it. The MMA coaching use case doubles as a stress test for the model's capabilities, particularly its frontier coding ability and multimodal perception. Zuckerberg noted that Muse has solid native understanding of MMA from its training data, suggesting the model absorbed knowledge about fighting posture, balance, and technique from its corpus. When asked how the model learned to distinguish proper versus improper MMA positioning, Zuckerberg pointed to their deliberate training approach: Meta trained the models end-to-end specifically to be good personal agents, with heavy emphasis on visual perception data because multimodal skills matter most for how people actually use these services. Wang pushed back against the industry consensus that coding ability is all that matters in AI development. He argued that because most people experience the world multimodally—talking, listening, watching videos, filming things—consumer products need multimodal excellence. Meta had an advantage here because of its massive reservoir of image and video data, but Wang emphasized that focus and intentional design mattered as much as raw data. The broader vision involves agentic commerce, hardware like smart glasses, and distributing AI capabilities to billions of people rather than concentrating them. Wang also noted that superintelligence won't be winner-take-all, suggesting room for multiple players in the AI space.

Ask about this article