Did Numerical Control Make Master Machinists Obsolete? Asianometry on Forty Years of Skills Moving House

When I carve a wheel, if I go slow, the spoke slides in but won’t hold; if I go fast, it bites but won’t go in. Not too slow, not too fast — I get it in my hand and it answers in my heart. I can’t put it into words, yet there is a knack in it. I can’t teach it to my son, and my son can’t learn it from me.

—— Zhuangzi, “The Way of Heaven” (Warring States period; my translation)

Wheelwright Bian spent his life making wheels. Cut too slowly and the joint is loose; cut too fast and it won’t fit. The measure lives in his hands, he can’t say it out loud, and he couldn’t even teach it to his son. That passage kept coming back to me while I listened to Asianometry’s September 10, 2026 episode, “Did Numerical Control De-skill Machinists?” Once the machines arrived in the shop, where did that unspoken measure go?

What this episode is about

When numerical control (NC) machine tools started to appear in the 1960s, many people worried they would “de-skill” machinists: the feel an old hand built over a four-year apprenticeship would be replaced by a roll of punched tape. The episode starts with nineteenth-century machinist culture, moves through NC and computer numerical control (CNC), and ends by holding the story up against today’s AI coding.

Key points

1. Half of a machinist’s craft lived in the ears and hands. Machinists read 2D blueprints and turned them into 3D parts in their heads, so they needed geometry, trigonometry, and algebra. The host quotes the 1887 edition of The Complete Practical Machinist: feed and speed on the lathe are the most intricate and deceptive part of the turner’s art, the part requiring the most judgment, perception, and watchfulness. A tool can sound fine while underperforming; a wrong sound can mean a malformed part or a lost finger. Old hands used gauges, and they also used feel to sense changes as small as half a thousandth of an inch. An apprenticeship ran four years and 8,000 hours. At the end, the company handed the new journeyman a set of tools or a cash bonus, and, the host adds, most importantly, he was now qualified for marriage.

2. Subdivision started in the nineteenth century. In 1883 a young machinist named John Morrison testified to the US Congress that the trade had begun splitting up in the mid-1870s, and then the pieces split again. In sewing-machine work, he said, “a man is not considered a machinist at all. Hence it is merely laborers’ work.” Riveting a boiler once took nine skilled craftsmen; with a hydraulic riveter it took one skilled worker and eight laborers. NC is one more step on that timeline. What set it apart was binding metalworking to data.

3. Early NC disappointed owners who hoped to cut skilled labor. NC was invented in the 1950s and first served the Air Force and aerospace, specced far beyond what small job shops needed. Its functions were hard-wired; upgrading meant opening the controller and rebuilding it. Surveys by Lawrence Williams and Brian Williams in the mid-1960s found shops still needed operators with the same skills. A failure meant expensive downtime, so owners put their best people on the machine.

4. The payoff came from utilization. As NC improved, one NC drilling machine replaced three conventional ones, and one NC mill replaced two or three. Productive machine time rose four to five times, making batches of dozens to thousands of parts economical. A 1982 paper on Swedish shops recorded one lathe firm hiring 22 NC operators instead of 44 lathe operators, and another replacing 63 qualified machinists with 21 NC operators. NC operators trained for six to twelve months; a lathe man needed four to five years. In Sweden, skilled workers willing to take night shifts were scarce, and NC kept the machines running at night — the benefit owners stressed most in interviews.

5. The further people stood from the machine, the surer they were that skill had dropped. A 1980s Canadian survey found nearly 80% of manual machinists said NC machining required “less skill.” Those who had operated NC machines themselves split half and half. Managers and NC programmers leaned toward “less skill,” though less lopsidedly than the manual machinists — perhaps, the host suggests, because they were one step removed from the work.

6. Wages didn’t match the perception. A 1989 study using 1981 Bureau of Labor Statistics data compared 80,000 Class A machinists in the US machine-tool industry at $9.72 an hour with NC operators at $9.51 — a 21-cent gap. Class B and C machinists, who did more repetitive work, earned $8.54 and $6.41, and some NC operators came from those ranks, which meant a raise. In the 1970s the microprocessor brought CNC; vendors added keyboards and screens to the machines, operators could edit programs on the floor, and some design power flowed back to the shop.

7. The new skill was reconstructing the logic of a failure. One operator told researcher William Cavestro that you mustn’t go straight to the breakdown: you retrace how the incident developed, identify the signs of failure, and don’t just push any old button. A Hurco operator said that when the drill works too hard, “it’s just not gonna sound right,” and the fix is in the program — run it slower or make the spot bigger. At a manual drill press, the same fix was pressing a little less hard. A mid-1990s survey of Taiwanese CNC operators found 80% had both machining and programming experience. Jeffrey Keefe’s 1991 study of three BLS surveys spanning 30 years found aggregate skill fell about 1%. Underneath, setup operators, semi-skilled operators, and unskilled workers saw their jobs change or disappear, and tasks migrated to NC operators.

Further thoughts

”Will AI make my line of work worthless?”

Near the end, the host mentions a February 2026 case study by Ryan Lopopolo at OpenAI: a brand-new product built by a team of AI agents, with humans monitoring and steering. Programming in the US today, like machining back then, is one of the dominant high-paying jobs. My first reaction was to worry for friends who write code for a living.

The episode gave me three layers to work through.

First, watch where the tasks move. In those two Swedish shops, headcount fell by half to two-thirds. What got removed was “tending one machine.” The people who stayed earned within 21 cents of Class A machinists. The machine absorbed the repetitive part; the part where you reconstruct the logic of a failure moved into the new role. The host says programmers will still read and judge code to diagnose problems — the difference is that to fix it, you tell the agent instead of opening the editor.

Second, look under the average. Read on its own, Keefe’s 1% suggests nothing happened; only by breaking out job categories do you see setup operators and laborers vanish as groups. Industry data fools me the same way: flat total revenue in a sector can hide one group of companies shrinking while another takes their orders. Now when I see a sector average, I ask one more question: who is being averaged with whom?

Third, notice where the person making the call is standing. In the Canadian survey, the surest voices were the manual machinists who had never run NC. I have the same habit with new technology — the tools I haven’t used are the ones I judge fastest. So before I form a view on a tool, I try to find someone who has used it hands-on for at least three months.

”If I learn the new thing early, how long does the premium last?”

In a 1992 oral history, Bill Bowman said employers increasingly expected machinists to program their own machines. He took the time to learn CNC, and in 1989 he earned about $30,000 a year — roughly $81,000 today. As more people learned the same skills, his premium compressed. Several years and layoffs later he was making $20,000, about $9 an hour, while doing more programming than before.

It reminded me of something I run into watching markets: a view is valuable while few people hold it, and once it spreads, prices absorb it. Bowman’s story puts the same dynamic on a paycheck. In the years when CNC programming was scarce, he earned a premium; when more people could do it, the premium shrank, and the layoff cycle pushed it lower still.

The reading I take from this: when sizing up the premium on a skill or a theme, ask how many people can do it now and how many will a year from now. While few can, the premium holds. Once courses and tools cut the ramp to six to twelve months, the premium is on a countdown. Bowman’s raise came and went within a few years.

”The sales numbers for a new technology look great — is it widespread yet?”

In 1980, 25–30% of machine tools sold in developed economies were NC-enabled, yet the installed base was far lower. Companies strained their finances to buy these expensive machines because rivals who adopted them pulled ahead on cost and lead time. Twenty years passed between invention in the 1950s and the acceleration of the 1970s, and early buyers dealt with hard-wired functions and had to assign their best staff to babysit the machines.

I got two readings from this. One is to separate “share of new sales” from “share of everything in use.” The first is flow, the second is stock, and headlines about adoption usually report the first. The other is to ask who captures the benefit. NC’s gains — utilization, night shifts, small batches — went to the shops that bought the machines, and those same shops stretched their balance sheets to keep up. A technology spreading doesn’t mean every adopter makes money on it. When I look at AI numbers now, I ask two things first: is this this year’s additions or the total already in use? And are the savings adopters capture being eaten by what they spend to keep pace with rivals?

References

• Asianometry, “Did Numerical Control De-skill Machinists?”, September 10, 2026
• The Complete Practical Machinist, 1887 edition (the passage on lathe feed and speed quoted in the episode)
• John Morrison, 1883 testimony to the US Congress (subdivision of the machinist trade)
• Jeffrey Keefe, 1991 study of three Bureau of Labor Statistics skill surveys spanning 30 years
• Bill Bowman, 1992 oral history
• Ryan Lopopolo (OpenAI), February 2026 case study on building a product with AI agents

One thing to take with you

When machines come in, skills move house. They move to the judgment that hears when something is off and can say why.

Here’s something I’ve tried: pick a task you do every week — making a pot of soup, running a meeting, reconciling an account — and break it into three to five steps. Next to each step, write one word: “machine” if a tool or someone following instructions could do it, “ear” if you’re the only one who would notice when it goes wrong. Then take one “ear” step, find a beginner, explain it in three sentences, and see whether they can do it by following what you said. The step you can’t put into words is Wheelwright Bian’s knack.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.