Danny Bernstein

Physical AI is still in its infancy. Nobody really knows what the winning stack is going to look like yet. And that’s what makes this moment so interesting. 1. Models are becoming products. We’re moving from research demos to APIs and real deployments. For the first time, other robotics teams can actually integrate these models into their own products. 2. Tasks are becoming workflows. Doing one manipulation successfully is very different from completing a 30-minute or 2-hour workflow reliably. The industry is slowly moving from “can the robot do this task?” to “can the robot actually do the job?” 3. Deployment is becoming part of training. Teams are putting robots into the real world and discovering failure modes they never saw in the lab. Those deployments create new data, expose edge cases and feed back into the next generation of models. 4. The data question is changing. It’s no longer just about who can collect the most hours. Teams are figuring out which tasks, environments, failures, interactions and types of demonstrations actually improve models. The quality and diversity of data matter much more than the raw number of hours. 5. Evals are becoming their own category. A robot completing something successfully once tells you very little. Can it do it 100 times? 1,000 times? Across different objects, environments and edge cases? Reliability is becoming the real benchmark. 6. A new infrastructure layer is being built. Once you have thousands of hours of data and robots running in the real world, you need systems to manage the whole loop: data → training → simulation → evaluation → deployment → feedback → retraining. A lot of this infrastructure is being built right now. 7. Business models are still being figured out. Model APIs, RaaS, vertically integrated robotics companies, data platforms, deployment companies, integrators… Nobody knows yet where most of the value will ultimately accrue. Everyone is basically learning in production. The next 12–15 months are going to be insane. Not necessarily because we’ll know the final winners, but because we’ll start seeing which approaches actually compound once more robots hit the real world. A few teams I’m watching closely: @DynaRobotics, @Ultraroboticsco, @origin_autonomy, @MeckaAI, @humynlabs, @dreamlabs_ai, @physical_int, @SkildAI, @xdofai, @GeneralistAI, @NVIDIARobotics @Figure_robot, @1x_tech, @FoundryRobotics, @fieldai_

1/ Today we’re unveiling Bombadil, @electricplantco's foundation model that reads the electrical signals of living plants and turns them into a continuous view of their physiology. Trained on 7M+ hours of plant electrical signal across 125+ species. That makes Bombadil the most species-diverse model of directly recorded electrical biosignals in existence. The animal equivalent would be one model that reads the brainwaves of blue whales, horseshoe crabs and fruit flies.