ARTICLE · FROM THE FIELD · Pace of change

A backflip isn’t the interstate

AI is getting smarter fast, and a lot of people expect it to take over everything at the same speed. I don’t think it will. Some work will go to AI very quickly. Other work, especially engineering in the physical world, will move much slower than people think. Not because AI is weak, but because of what that work actually takes. Here’s how I see the difference, from someone who’s engineered the guarding on robot cells.

The same questions apply to a sales pipeline, a purchase order or a month-end close: what stays fixed, what needs judgment, and who decides?

Two kinds of work

Looking things up. Think about legal research: finding precedents, cross-referencing cases, checking what a regulation says. The law library grows every year, but it grows slowly. AI’s ability to search it, read it and connect it is growing much faster than the library is. So this kind of work, research over a body of knowledge that grows slowly, is the kind AI takes over fast. Finding the right standard, checking a spec against its source, summarizing a pile of documents: same category.

Engineering in the real world. Now think about setting up a robotic cell. Every cell is its own problem. How fast is the robot moving? How long does it take to stop? How far is it from the guard, and from the person on the other side? What can be checked in the time available before a decision has to be made? A safety stop has to be decided in milliseconds. Inspecting a part can take a few seconds. Each answer depends on the one before it, and you have to calculate it for this cell, this part, this layout.

That analysis doesn’t get smaller as the model gets smarter. Every new situation is new math.

Why the second kind moves slower

To set up a whole cell on its own, an AI agent has to do all of that engineering, every time, for every decision. And the work stacks up. It isn’t twice as much analysis to handle twice as much; it grows much faster than that, and so does the computing power and data needed to do it well.

AI is improving fast too, so the two partly cancel each other out. My view is that in this kind of work, progress will look closer to a steady line than a rocket, even though both sides are growing fast underneath. Some work will be taken over at a startling pace. Engineering in the physical world will take longer.

There’s outside support for this. Researchers in robotics noticed it decades ago and called it Moravec’s paradox: the things that feel hard to people, like math and looking things up, turned out easy for computers, and the things that feel easy, like moving through a messy physical world, turned out hard. A common way to put it today: robots can be taught a backflip, but reliably opening an unfamiliar door is still very hard. Shankar Sastry, a robotics professor at UC Berkeley, named the other difference in a 2026 keynote: “the physical world is much more unforgiving than a chatbot.” When a chatbot gets something wrong, you ask again. When a robot gets something wrong, it can break expensive equipment or hurt someone.

The stunt and the interstate

Tesla’s new Roadster prototype shows both sides. At its reveal in October 2026, it drove up a purpose-built looping track and used thrusters to fly forward, do a barrel roll and land. That’s an impressive piece of control. I don’t doubt they’ll get computers doing things like that routinely.

But it ran on a course built for it. Doing a planned stunt on a prepared stage is a solved problem you run once. Doing the same thing live on the interstate, with traffic around you and nothing planned in advance, is a far bigger problem. Even Musk drew the line at the event, saying the car is “not necessarily a safety-optimized car.” The same company describes the prime directive for its self-driving system as “don’t crash.” The stunt and the interstate have different top rules.

That’s the gap between a demo and a production line. It’s the real engineering.

What this means for a plant

  • Expect fast gains in the paperwork. Research, standards lookups, document drafting, analysis of the data you already have. These are the places to start.
  • Expect slower gains on the floor. AI will help engineer cells and lines, but it won’t replace the engineering of stopping distances, guarding and checks any time soon. Be skeptical of anyone who says it will run your line next year.
  • Keep the controls engineering. The plant-floor discipline of states, interlocks and commissioning is exactly what makes AI safe to use near equipment. It’s becoming more valuable, not less.

No need to panic about it, and no need to wait for it either. It’s engineering. It’ll get done the way engineering always gets done: one problem at a time, checked as you go.

Sources

These pages sit with the rest of the Advance Now family. The hotbotspot tool library writes the methods up for AI. A.N. Creation Office is where the public AI agents run as open experiments.