49:27Qiong Wang and colleagues at Physical Intelligence describe their effort to build a general-purpose model that can control many kinds of robots, and explain how language models, vision-language models, and cross-embodiment robot data have advanced the field. They argue that robotics is becoming cheaper to build as stronger models let teams use less specialized hardware and focus on data collection, evaluation, and fitting robots into real workflows. Demonstrations with Weave and Ultra show laundry folding and warehouse packaging in real-world settings, with human intervention supporting deployment as autonomy improves. The speakers also discuss cloud-hosted inference, the infrastructure gaps facing robotics startups, and their hope that open research and models will enable many new robotics companies.