skara

Thats awesome! A developer used Al-powered 4D Gaussian Splatting to convert flat video footage into dynamic 3D spatial scenes. The system reconstructs different camera angles and depth information from ordinary footage, making it possible to navigate scenes in three dimensions. We're moving from recording video to digitally recreating reality itself.

some tentative thoughts on humanoid robotics 1) i'm not sure there is such a thing as asi for humanoid robotics, at least in a commercially valuable sense 2) once you have a robot that can take text instructions and turn them into decent management of a particular robot embodiment, you have captured most of the value 3) at that point, you can deploy it onto a production line, you can use it for unpacking and stocking shelves in a convenience store, perhaps it can be used in the home, etc... 4) at scale, like in a warehouse or on a factory floor, they can be coordinated using reasoning models, which receive telemetry and issue commands to individual robots and troubleshoot their behavior; 5) you can imagine further advances in robot operating models that create incremental value in particular industries, but the core of the value has already been captured 6) you can also imagine more general models that can operate multiple kinds of physical embodiments well, not just one; but i tend to think that this is mostly a cost improvement for operators 7) we should assume foundation models for robotics will follow a similar trend to llms, where the open source models trail the frontier by 9-24 months 8) this is true in part because there are well resourced players, like nvidia, that would train these models and have an incentive to open source them, to avoid concentration of their customer base 9) so, the companies that get the software advantage first have 9-24 months of lead time, but their models will saturate much more quickly than language model intelligence saturates 10) at that point, most of the value goes to whoever can produce the robots at scale and get them out at good gross margins, not to whoever produced the software 11) so, the winners look like the people that are good at building and running the factories and tooling them, plus the people that are good at training language models for discovery and operations that support this 12) my gut is that companies like figure and physical intelligence are on the wrong side of robotics on the long term; they are too invested in the software 13) tesla is maybe on the right side; chinese hardware companies are certainly on the right side; as these companies specialize in building at scale 14) there will also be many niche uses of robotics that both require further capability unlocks to be fully valuable and are vlm-shaped, like very small air gapped drones for war 15) but, i suspect this is not the majority of the economic value for humanoid robots and the majority of the value saturates on intelligence

every university should have one of these in their robotics department

boston dynamics uses mjlab btw :)

Atlas hauling a 50 lb mini-fridge - Practiced the maneuver for millions of hours in a virtual environment. - Focused the training policy on full-body engagement rather than just hand-grasping, allowing the robot to leverage its entire frame for the lift.

I built something that finally gives agents the option to see and measure the physical world, reason within it, and accelerate your business with better and faster decisions. Pulling out insights from 3D data and cross-referencing across data silos, completely automating all the steps. From point cloud to the sent pick-up price, date, and picking the best buyer. If you have other companies our agent should integrate with, tell me!