spent 3 hours at @saturdayrobotic reading club #28# at @RoboticsCtr listening to @JasonMa2020 from @DynaRobotics on robot foundation models and data scaling. takeaways from the talk: 1. video is a scaling axis on its own. holding action-labeled data fixed while scaling unlabeled human video from 1K to 50K hours keeps improving human-to-robot prediction. action-only training misses this entirely. 2. zero-shot is the wrong goal for enterprise. post-training amortizes gains into a single run instead of pushing cost and latency into every live inference. 3. evaluation is the real bottleneck. automated step labeling, failure detection, and regression tracking across fleets is what actually scales robots from tens to thousands. 4. deployment proof: Dyna2 napkin folding demo under stress (darkness, flashing lights, occluded sensors). at CosmicBrain, this is why we are building the deployment and teleoperation layer for these physical AI models.