The Next Frontier of AI Is Spatial Intelligence | Fei-Fei Li on a16z

World Labs' acquisition of SceniX aims to resolve the primary bottleneck in robotics: the scarcity of high-quality data. By building digital versions of the physical world, the combined team creates a foundation for training and evaluating robot control software at scale, enabling robots to learn in simulated environments before they operate in reality.

• Spatial intelligence — Developing AI that perceives, reasons with, and interacts within physical or virtual environments [01:40].

• Data bottleneck — Current robotics development is hindered by a shortage of high-quality training and evaluation data [06:47].

• Real-to-sim-to-real — Mapping physical surroundings into digital models allows developers to bypass the constraints of physical data collection [04:35].

• Counterfactual reasoning — Simulation enables robots to learn from hypothetical scenarios that are too dangerous or difficult to replicate in testing [00:25].

• Infrastructure vs hardware — The goal is to provide an agnostic platform for robot brains rather than building the mechanical machinery [27:47].

• Reliability and speed — Digital platforms allow for training cycles that operate faster than human-led teleoperation [22:29].

• World consistency — Maintaining physical rules across space and time prevents errors where objects inexplicably disappear [00:33].

• Phased integration — The teams will retain independent tech stacks while collaborating on shared projects like the Marble base model [37:20].

• Market focus — Initial efforts prioritize semi-structured settings like warehouses over fully unstructured homes, which present higher complexity [30:44].

• How does simulation allow for systematic randomization of physical parameters like friction and lighting?

• What distinguishes the World Labs approach to spatial intelligence from standard video prediction models?