This episode examines a Swedish registry study that achieved seventy three percent area under the curve in predicting melanoma risk using structured health data, enabling selective screening of high-risk cohorts. We cover the convergence of vision-language-action models, world action models, and native embodied foundation models like Generalist’s GEN one, which trained on over five hundred thousand hours of physical interaction data. Physical Intelligence’s RECAP method demonstrates reinforcement learning post-training that cuts failure rates by half on complex manipulation tasks through value functions and heterogeneous data. The episode closes with the infrastructure shift from physical data collection to compute-scaled simulation and the role of wearable devices as distributed data-generation platforms for physical AI systems.
