My latest at Forbes:
You can’t fine-tune your way to Artificial General Intelligence (AGI) for robots, declares 1X CEO Bernt Børnich, revealing the ambitious launch of their new World Model Lab. Led by Sam Sinha, a founding researcher from Luma AI, this lab isn’t just an upgrade to their existing foundation model; it’s a full-throttle acceleration towards fully autonomous humanoid robots that can work in teams and truly think.
Forget patching up existing models. Børnich argues that every significant AI breakthrough stems from feeding models the right kind of data, in the correct order — text, then images, then video. The critical error many make? Treating robot data as a mere afterthought, a superficial layer atop general pre-training. Sinha sharply agrees, calling this approach “fundamentally broken.” He insists: “You need to see your most important tokens from step zero.”
With 1X now scaling up production of its Neo humanoid robots, a torrent of rich, live data is flowing in: visual streams, precise proprioceptive data from joints, pressure and grip feedback from Neo’s advanced hands, and even web-scale human video. This isn’t just data; it’s the raw material for a new kind of intelligence. The big bet? Robotics is simply another AI scaling challenge, and Neo’s human-like design — including 22 actuated degrees of freedom in its hands — is key to leveraging vast human video for robot training. My latest at Forbes.