Work video meetings in the traditional grid interface have inclusion, effectiveness, and fatigue problems, due in part to the difficulty of directing or communicating attention. Virtual 3D meeting spaces have value, but representing people in them is a challenge. Avatars face resistance, and 2D video is limited to near-frontal views, constraining the spatial layout. We present a novel experimental system for virtual meeting rooms that predicts 3D video of users in real-time from a standard webcam, positions them in a shared 3D space, and renders a controllable first-person view. We report study results comparing this system to a traditional grid, and to 2D video of people in the same 3D space. While spatial layouts fared better in terms of attention and co-presence, the traditional grid was more comfortable and professional. This is likely due to unsettled 3D design, the need for manual control, and a preference for the familiar.