Robotics
There is a clip I keep rewatching. A robot arm reaches for a crumpled towel, misses the fold slightly, pauses, and adjusts its grip — a small, unremarkable correction. What makes it remarkable is that nobody wrote that correction. No engineer coded the case where the towel bunches to the left. The robot learned it the way a new hire learns: by watching someone else do the job, over and over, until the motion became its own.
I am not a roboticist. I watch this field from the outside, as a designer and engineer who spent years making things move on screens — broadcast graphics, 3D rigs, motion systems. But that outside view is exactly why the current moment feels so strange to me. The rules of the discipline appear to be changing mid-game.
The end of hand-written motion
For most of its history, robotics was choreography written in code. Engineers modeled the physics, computed trajectories, and wrote controllers that assumed the world would behave. Every new task was a new engineering project. A robot that welded car doors knew nothing that transferred to stacking boxes; strictly speaking, it did not know anything at all.
The new approach flips this. Instead of programming the motion, you demonstrate it — a human teleoperates the robot through a task a few hundred times, and a neural network learns a policy that maps camera images to motor commands. Layer on models pretrained on language and internet-scale video, and you get systems that can follow an instruction like "put the cup in the sink" in kitchens they have never seen.
That sentence would have sounded like fantasy a decade ago. Today it is a demo you can find from half a dozen labs. How reliable those demos are outside the demo is, honestly, still an open question — and the labs know it.
The hard part moved from motors to minds
There is an old observation in AI, usually called Moravec's paradox: the things that feel hard to us — chess, calculus — turned out to be easy for machines, while the things a toddler does effortlessly — grasping, walking, noticing — turned out to be brutally hard. Robotics lived inside that paradox for fifty years.
What changed is where the difficulty sits. Motors, sensors, and batteries still matter, and dexterous hands with a real sense of touch remain genuinely unsolved. But the mechanical side has become more of a supply-chain problem than a research problem. The bottleneck now is data and generalization: how do you give a machine enough experience of the physical world to act sensibly in situations nobody scripted? That is a minds problem, not a motors problem, and it looks a lot like the problem language models just crashed through.
The humanoid wave, squinted at
Which brings us to the humanoids. Every few weeks a new biped folds laundry or hands someone an apple, and every few weeks it emerges that parts of the demo were teleoperated or heavily curated. My skepticism is real: reliability numbers are rarely published, and a robot that succeeds four times out of five is useless in a home.
But I do not dismiss the bet itself. The argument for the human shape is not romance, it is data economics. Our world — stairs, door handles, shelf heights, tool grips — is an interface designed for one body plan. Build for that plan and every environment becomes potential training data, and one platform amortizes everything learned on it. It may take far longer than the promotional videos suggest. It also might work.
What a builder recognizes
Here is what this shift keeps teaching me about my own work. Every product I have built solo — a broadcast platform, a QR ordering system, a scheduling tool — began with me writing explicit rules: state machines, edge cases, validations. Specification is expensive, and it always misses the case where the towel bunches to the left.
Demonstration is a different contract. You show the system what good looks like and let it fill in the gaps. Watching robots acquire common sense this way — clumsily, improving with every correction, the way any junior does — makes me suspect the best interfaces of the next decade will be taught rather than configured. Fewer settings panels; more "watch me do it once."
The machines are finally learning the way we always did. The interesting question is what we will choose to show them.