Robotics Didn't Need Better Hardware. It Needed Better Brains.

For years, robotics felt like a slow, grindy discipline where progress was measured in millimeter precision and years of bespoke training for a single robotic arm to fold a towel. The conventional wisdom was simple: machines fail in the physical world because reality is messy, unpredictable, and starved of training data. To fix it, we thought we needed millions of hours of physical trial-and-error in specialized labs.

That assumption is falling apart fast. Frontier multimodal models—built primarily to understand text, code, and video—are quietly leapfrogging dedicated robotics software. It turns out that when a model develops a deep, generalized understanding of how the world works, it transfers directly to physical space. Common sense physics, spatial reasoning, and intuitive cause-and-effect don't require ten thousand hours of a robot bumping into a table if the underlying brain already understands what a table is and why glasses break when dropped.

We spent a decade treating robotics as a mechanical problem when it was always an intelligence problem. Modern hardware—servos, cameras, compact batteries—has actually been good enough for a while. What was missing was an agent inside the shell capable of looking at an unfamiliar kitchen, figuring out where the sponge lives, and adapting when something slips out of its grip.

This changes the development curve completely. Instead of building physical robots from the ground up, we are watching general frontier intelligence get poured into existing mechanical frames. The hard part of robotics wasn't building the hand; it was giving the hand a reason to move.