The Closest Thing I Have to a Perpetual Motion Machine

How an $850 invoice rewired what I think is buildable, and why I now ask students to build a tool instead of write an essay.

An $850 problem

The shift in my thinking didn't come from a conference talk or a think piece. It came from an invoice.

I was paying QuickBooks about $850 a year to do two things for a side business: generate estimates and send invoices. That's it. Not payroll, not inventory, not tax prep. Estimates and invoices. Every year I'd look at the renewal, feel the specific irritation of paying rent on functionality I could describe in a paragraph, and renew it anyway, because the alternative was learning to build software.

Then I built it. A small web app that does exactly the two things I needed and nothing else. It took evenings, not months. It has been running ever since, and the $850 stayed in my pocket.

The money was never the point, though the money is what made it legible. The point was what the experience did to my sense of the possible. I had spent years quietly sorting problems into two bins: things I could make with my hands, and things that required software and were therefore somebody else's to solve. That second bin turned out to be much smaller than I thought. Almost everything in it was reachable.

I should say what I did not conclude, because this is where a lot of AI enthusiasm goes wrong. I didn't decide to hand over my thinking. I still use AI when I write, mostly to dictate and then clean up my own rambling, but I don't let it generate the thoughts. The distinction I landed on is simple and it has held up: it's a poor substitute for the work I'm supposed to do, and an extraordinary instrument for building the things that do that work better.

Toolmaking as a habit of mind

I've been a maker, a builder, a designer, and an educator my whole working life, and if you asked me for my favorite thing to do in a metal shop or a wood shop, it wouldn't be finishing a project. It would be building a jig.

There's a particular satisfaction to building a tool that I've never gotten from building a part. A part is a part. A tool changes what you're capable of making next, and the thing you make next might itself be a tool. It's the closest thing I have to a perpetual motion machine: each fixture, each jig, each shop aid expands the envelope of what's reachable, and the expanded envelope makes the next tool easier to build.

So when I say AI changed my practice, this is the specific sense. It extended toolmaking into a material I had never been able to work in. I can build a fixture out of aluminum. Now I can build one out of software.

Everything turns out to be a simulator

What I did not expect is how consistently my tools have converged on one form. Almost everything I've built is a simulator.

There's a plug in that runs wind simulation on 3D models in Fusion, so I can see how a form behaves in moving air before it exists.

There's one I care about more: a materials simulator for 3D printed parts. Fusion's native simulation treats a printed part as a solid, isotropic object, which is a convenient fiction and sometimes a dangerous one. A printed part is not solid. It's a stack of layers with directional weakness at every interface, and its strength depends on print orientation, infill, and layer height in ways a solid body simulation simply cannot see. My tool models the part as what it actually is. It has changed design decisions I would otherwise have gotten wrong.

There's a plug in for laying out kitchen cabinets, which is less romantic but saves real hours.

And there's the CNC training game I built for my students, which simulates the Tormach mill in our shop so they can snap virtual end mills instead of real ones.

Looking at that list, the pattern is obvious in retrospect. A simulator is what you build when the real test is too expensive, too slow, or too destructive to run as often as learning requires. That describes most of what's hard about both fabrication and teaching fabrication. Physical testing is the bottleneck in both. Every one of these tools is an attempt to make the failure cheap enough to repeat.

The assignment

Which brings me to what I actually ask of students, and why I think it matters more than any policy debate about AI in the classroom.

My students, by and large, do not know what AI is capable of. They know it writes essays. That's the extent of the model most of them have, and it's the reason the whole conversation in education has calcified around cheating. Nobody has shown them the other thing.

So I show them mine. I walk them through the tools I've built, how they work, why I needed them. Then I ask them to build one for a project they're already working on. Not a generic tool, not a demo: something specific to a real need in front of them.

The one I lean on hardest is a polymer weathering simulator. Students model how different polymers behave in extreme conditions, heat, cold, UV exposure, so they can make a defensible material choice instead of grabbing whatever's on the shelf. The learning isn't really about the simulator. It's about developing a feel for where a given polymer shines and where it quietly fails.

These usually live as an HTML file on a student's local machine. No deployment, no hosting, no account. They could turn one into an app or a website easily enough, and some will, but the artifact isn't the deliverable. What they turn in is a screen recording of themselves using the thing they built, narrating what it does and why.

They're not graded on creativity. I want to be clear about that, because I don't want students optimizing for flourish. But I do push them hard on one thing: don't accept the first version. Whatever comes out of the first attempt is a starting point, not a result. Change the colors, change the layout, change what it actually calculates. Make it yours. The students who take that seriously end up with something they understand, and the ones who don't end up with something they merely possess. The difference shows up immediately on the recording.

The part I haven't solved

Here's the honest problem, and I'd rather state it plainly than pretend the pedagogy is finished.

It is genuinely hard for a student to check the work of the simulator they just built. A student who doesn't know polymers cannot easily tell whether the material properties are real or plausible looking fabrications. The tool produces confident output either way. This is the central risk of the whole approach, and anyone who tells you they've solved it in an undergraduate classroom is not being straight with you.

What I teach instead is a disposition, and I've started calling it intuitional engineering.

Does this sound right? Does it look right? What was your first thought when the number came up, before you talked yourself out of it? If anything feels off, even slightly, that's not noise to be suppressed. That's the signal. Go check it. Pull the datasheet. Run a real test. Then correct the tool and run it again.

That last step is the one that makes it engineering rather than skepticism. The simulator is not an oracle to be trusted or rejected. It's an instrument to be calibrated, and calibration means going back and forth between the model and the world until they agree. Students who learn that have learned something that outlives whatever AI tool they're using this semester.

I'll admit the obvious limitation: intuition requires experience, and a first year student doesn't have much. What I'm really doing is starting the clock early, giving them a habit to grow into rather than a rule to follow. I'd rather they leave my class with a slightly overactive suspicion than a well calibrated trust in something that doesn't deserve it.

Where this leaves us

We're in a transitional moment, and education is mostly afraid. I understand the fear. The first visible use of this technology in a classroom was students getting out of work, and institutions responded the way institutions do, with detection tools and prohibitions and a great deal of anxious policy.

But the fear has a cost that nobody is accounting for. Every semester we spend policing AI as a cheating vector is a semester in which students learn that it's for cheating. We're not preventing a mental model. We're teaching one.

The alternative isn't permissiveness. It's a different assignment. Ask a student to build something instead of write something, and the entire dynamic inverts. There's nothing to plagiarize, because the tool doesn't exist yet. The AI does the tedious part, which was never the pedagogical content anyway. And what's left for the student is the part I actually want to grade: what do you need, what would solve it, is this output any good, and what will you change now that you've seen the first version?

I don't know if this scales past my shop or my discipline. I know it works better than what I was doing before, which was watching students learn the least interesting possible thing about the most consequential tool of their careers.