John Collison interviews Mark Zuckerberg and Alexandr Wang about Muse Wonder and Meta's personal-agent strategy. The discussion ranges from shopping and video coaching to active memory and long-running tasks, with an emphasis on training useful behavior into the model rather than relying on a complicated interface. Performance and adoption claims are the speakers' own account, not independently verified results.
Mark Zuckerberg and Alexandr Wang describe how pretraining, supervised fine-tuning and reinforcement learning contribute to a product's capabilities and behavior. They distinguish benchmark progress from internal product evaluation and argue that generalization, user experience and alignment need attention alongside measured scores.
Mark Zuckerberg and Alexandr Wang discuss the long lead times and substantial investment required to build frontier models and computing capacity. They connect that infrastructure strategy with delegated commerce, voice interaction and a wider ecosystem of applications, then describe the technical leadership and concentrated teams behind Meta's AI reboot.
Watch on YouTube
