You Cannot Bullshit a Soldering Iron

Blender and 3D animation has always had too steep a learning curve and I have consistently fallen short of making headway in learning these skills.

Five separate runs at it across the better part of a decade. Each one real: a stack of Saturdays, twenty or thirty hours of honest effort, tutorials open, following along.

Every one ended in drift. I’d hit something I couldn’t get past, spend two hours not getting past it, and next Saturday I’d find something else to do.

Eventually I worked out why. Blender is a mastery skill, and so is 3D modeling, animation, digital art, and every discipline stacked underneath them. Learning any of it properly demands the same order of investment I gave networking and software development. Years of consistent practice, most of it unpleasant, before you reach the point where you can play.

I’d already spent those years on two careers. There wasn’t a third set available for a hobby.

Time was the constraint, not ability or interest. I think a lot of working adults have quietly done the same math on something they’d love to do.

Something changed it. Not automatically, and not for the reasons people usually give.

---

Everything I needed was on YouTube and had been for years.

A tutorial answers the question its author decided to answer. The thing blocking you at 9pm on a Saturday is almost never that question. It’s something stupid and specific. Your version has a different menu. The thing you clicked did something else. You can’t search for what went wrong because naming it *is* the skill you don’t have yet.

So the evening goes to hunting instead of building. Three weekends of that and you quit, and you tell yourself you weren’t that interested.

I was that interested. I quit anyway, five times.

This is one of the oldest known problems in education and it has a name.

In 1984, Benjamin Bloom published a paper reporting that students taught one-to-one performed better than 98% of students in a conventional classroom. Two standard deviations. Then he said the part that matters: one-to-one tutoring is “too costly for most societies to bear on a large scale.” He was posing a question. How do we get anywhere near these results under real economics?

Forty years later, the question is still open. Nobody has replicated two sigma. A 2020 meta-analysis of ninety-six randomized tutoring studies found an average effect around 0.37 standard deviations, enough to move an average student to roughly the 64th percentile. Real, repeatable, and a long way short of the myth.

The direction has never been in doubt, though. Personal, responsive, adapted-to-you instruction beats sitting in a room with thirty other people. We’ve known that for four decades and we could never afford to give it to anyone.

Every technology we threw at the problem in the meantime made education more industrial. Learning management systems. Recorded lectures. MOOCs with 90% dropout. We got very good at delivering the same lecture to more people while getting no closer to the thing Bloom actually measured.

Natural language is the first technology that touches the real constraint.

---

Early 2025. I was building 3D environments in Godot for what became Root Cause, and one of the first Blender MCPs had just landed. MCP was new and people were bolting it onto everything, mostly badly. I wanted to see how far it could get.

I ran it the way I run every project. It does the work; I coach. I’m not writing this; I’m directing it.

The loop that emerged.

I’d say: take this tree, put it over there by the dumpster. The MCP would go do that and report back done. I’d look at the scene and it would be wrong. Floating, clipped through a wall, facing backwards. So I’d screenshot it and say here’s what I’m actually seeing. And we’d go again.

Over and over, for hours… for days.

That loop is the whole thing. It wasn’t efficient. What it was, on every single turn, was instruction delivered at the exact moment I needed it. Why that transform behaved that way. What that panel controls. Why what I asked for produced what I got.

I always joke that Ive learned Golang three times on three seperate educational projects but still I would not call myself a strong Golang developer. It never stuck, because afterwards I went back to my normal stack and never actualy applied what I learned.

I got the scene built. I got past a wall that had stopped me five times.

It’s worth being precise about what happened, because it’s easy to overclaim in both directions. AI is still genuinely bad at 3D space. Ask it to place an object at a specific point and it has no reliable grasp of where that point is. Assets, complex scenes, real spatial reasoning: all still a problem, even improved.

I also didn’t come out of it a Blender Pro. I’m entry level at best but even with limited skills I can still build and more importantly slowly solidify my skills as I need them.

What changed is that the thing standing between me and the work got small enough to step over. The years of practice I’d correctly identified as necessary got compressed into the moments where they actually mattered, delivered exactly when their absence was blocking me and with enough resistance to force me through the correct steps.

The finding was about how the two of us work together.

---

Once I saw the pattern I started noticing it everywhere I was stuck.

Soldering. I’m building an LED wall driven by an ESP32, adapted from a project I found on YouTube. My parts didn’t match theirs so the wiring had to change. This was the project where I forced myself to get good at soldering, which is a skill you acquire through your hands, by feel. Nobody picks it up from a video alone.

But the specific way I was failing at it is diagnosable. Instead of scrubbing through fifteen minutes of someone else’s tutorial hoping they’d happen to make my mistake, “Im touching the solder to the tip, why the hell wont it melt?” I could describe what I was seeing and get told what I was doing wrong. On my mistake, in the moment I was frustrated enough to quit.

FreeCAD. The cyberdeck I’m building needs a housing, and the LED wall convinced me I was done downloading other people’s STLs to 3D print. So I stepped off the Blender path entirely and went to FreeCAD: parametric, dimensioned, engineering-discipline modeling instead of artistic modeling. It suits how my head works far better than Blender ever did. I’d never have found that out without being able to move fast enough to compare them.

Cake pop molds. My sister-in-law has made cake pops for years and complains about the same two things. No variety in the available molds, and tools slow enough that her time goes into stamping shapes instead of making anything interesting. Could I print better ones? That turned into a real problem, since 3D printing plastics aren’t food-safe, which pushed the whole thing toward printing masters and casting silicone. Several versions in, still iterating.

Painting. The one I’d have predicted least. I’ve painted abstract for years, partly by preference and partly because I can’t draw and my color theory is bad. Lately I’ve moved toward impressionist and representational work, to make strides in my skills. I now have feedback on mistakes and technique to push a piece to completion instead of a fresh white coat of gesso for the next inspiration.

---

In 2024, researchers at Wharton and Penn ran a field experiment with about a thousand high school students learning math. Two AI conditions, same underlying model.

The first was a standard ChatGPT interface. Students using it improved 48% on practice problems. Then the researchers took it away for the exam, and those students scored 17% worse than students who never had AI at all.

The second condition was the same model with prompts built to protect learning. It made students articulate their reasoning before revealing anything. Practice performance went up 127%, and on the unassisted exam the effect was statistically zero. No penalty.

Same model, same students. Structure was the variable.

The authors call it the crutch effect, and every engineer already understands it, because we have our own version. You don’t paste code from Stack Overflow that you don’t understand. Everyone nods at that. Everyone has also done it out of desperation and paid for it later.

Identical failure. Working output, no comprehension, and a debt that comes due the first time something breaks and you’re the one holding it.

My loop is the second condition. Not by temperament, but because I’d already learned that lesson expensively somewhere else.

---

Robert and Elizabeth Bjork spent decades on what they call desirable difficulties: the finding that conditions which slow learning down often improve long-term retention and transfer. Spacing, interleaving, retrieval practice. Each one feels worse in the moment and each one wins on delayed tests. Their central distinction is between performance, what you can do today, and learning, the durable change. Most study methods maximize the first while quietly destroying the second.

My claim is that AI shortened the unpleasant stretch. Their research says the unpleasant stretch is where the learning lives.

Two things save the argument and both come from the research itself.

The first is Bjork’s own qualifier. A difficulty is desirable only if it engages effortful processes the learner is capable of executing. Push past that and you’re flailing rather than retrieving. Insufficient prior knowledge produces overwhelm instead of challenge.

Five abandoned Blender attempts and 3 Golang classes are stories about the second kind of difficulty, the kind that produces zero learning because you never get far enough in to struggle productively. AI moved me from overwhelmed to challenged, which is exactly the window where Bjork says learning happens.

The second is sequence. An MIT study last year got a lot of press for finding weaker neural engagement in people who wrote essays with an LLM. What I find interesting is the reverse condition: participants who worked unaided first and brought the model in afterward showed higher recall and re-engagement.

Order matters!

---

That MIT study measured essay writing, and the choice of domain explains a lot.

Essay writing gives you no ground truth. You read the paragraph, it sounds fine, and you have no independent way to know whether it’s any good. Nothing catches you.

Now look at everything in this post.

The tree is floating or it isn’t. The joint holds or it’s cold and it fails. The print fits the housing or it doesn’t. The mold releases or it tears. The build compiles and runs, or it doesn’t.

You cannot bullshit a soldering iron.

That’s why my experience diverges from the alarming studies, and it has nothing to do with me being special. The crutch effect requires a domain where you can’t tell you’re wrong. Makers, builders, and engineers work in domains that tell them immediately, unforgivingly, and for free.

Which produces actual advice: Stop trying to one-shot with AI, instead learn with AI where reality grades your work. If your domain doesn’t grade you, manufacture the grader. A test, a print, a dry fit, a person who’ll tell you it’s bad. The maker world has always known this, and it’s what jigs and test fits and prototypes are for. We scaffold physically because the consequences are physical.

---

My own advantages are worth naming.

I have unusually high tolerance for being bad at things in public. I’ve spent twenty years learning how to learn. A decade of CCIE-shaped grinding taught me what sustained difficulty feels like from the inside. I’m comfortable in the unknown to a degree that isn’t normal and isn’t a virtue, it’s just how I’m wired.

The Wharton study is the way out. Both groups got the same model. One had structure supplied by the prompt design and I supplied mine out of habit.

The scaffolding doesn’t have to live in your personality. It can live in the tool, the process, a checklist, or a person you show your work to. Make it articulate its reasoning. Make yourself say what you expected before you look. Keep something in the loop that will tell you the ground truth but also provide constructive criticism and encouragement.

Design your agents and workflows to automate this instead of the actual work, and you don’t need my temperament.

---

Engineering education has known for at least twenty years that it doesn’t produce practice-ready engineers. That’s the mainstream position in the field, not my opinion. The Carnegie Foundation funded a multi-year study on it. The Royal Academy of Engineering published reports from both industry and academic sides flagging the same gap. Australia’s engineering deans called for rebalancing theory against practice and building in real-world problems. Interview working engineers and you get the same answer: school didn’t supply the practical skills, and they picked them up afterward, on the job.

School never taught engineering judgment. Apprenticeship did. The junior years, the grunt work, the tickets nobody wanted, the bad version you wrote and then had to fix. That’s where judgment came from.

That layer is being automated away right now. Stanford’s Digital Economy Lab found a 16% relative employment decline for 22-to-25-year-olds in the most AI-exposed occupations, and the finding held against every counter-explanation they tested. Most of the recent recovery in software job postings has gone to senior roles.

So we’re removing the part of the pipeline that worked while the broken part stays broken.

What we lose specifically is verification capacity, the ability to look at plausible output and know it’s wrong. You get that from having produced the wrong thing yourself and felt the consequence. If nobody does the entry-level work, nobody develops the judgment to check the work.

Just like cell phones and the internet, society will take a while to adjust to life with AI. Both of those adjustments happened to us rather than being chosen. We handed over our memory for phone numbers and our sense of direction and only noticed the trade long after it closed. Neither cost much. This one is different, because the thing on the table is judgment, and judgment is what you were going to use to evaluate the trade.

Hopefully this time around we notice while it’s still a choice.

It isn’t a policy choice, or not only. It gets made one project at a time by people deciding how they’re going to pick the thing up, and that decision is identical at work and in the garage. Most people are currently making it by default.

The default is to reach for the answer. It works, it feels like progress, and it costs nothing you can see today.

---

My last post covered five and a half years of this from the professional side: the projects, the sharp edges, what it cost. https://abstractryan.substack.com/p/the-long-way-around?r=j7qo3