The big question about Artificial Intelligence, other than whether it will destroy human civilization, is the effect it will have on employment. Large language models already seem poised to automate large fractions of the work in certain industries, and the technology is only getting more capable.
Having spent the past nearly 50 years working in and with manufacturing automation, I am in the optimist camp.
I've been hearing since the early 50s "you can be replaced with a button" often as part of a screed about how automation was going to put us all out of work. It has not happened yet. Even without AI I do not believe it will ever happen.
Not only have we created new jobs (Like Changeover Wizard?) for all those people displaced by automation, we have generally created better better jobs.
Go read Henry Ford's 1923 autobio "My Life and Work" (Email me at johnhenry@changeover.com if you would like a copy) He describes how he invented what is now called the Toyota Production System..
He was also a fanatic on automation "NEVER put a man to do what a machine can do" He probably created more jobs, paying double or more the going wage, than anyoiine before or since.
Manufacturing output in the US, for all the doom and gloom we hear has increased steadily and dramatically year after year from WWII to today.
Manufacturing employment is down and that is a good thing. Most unskilled and semi-skilled (fully trained in a month or so of on the job training) are really crappy. They are probably not something to be desired, they are something to be eliminated.
So, let's go automation and let's go robotics. And to the extent that AI can improve automation and robotics, lets go AI.
Trying to figure out anything sane about AI (in its current form this AI cycle as LLMs) is massively overcomplicated by the boosterism of a marauding giant blob of money getting anxious that tech is about to turn into a dull industry. There's a hundred years of science fiction to coopt as marketing (even when it's apocalyptic; if your app could take over the world there's no question that it's competent), quasi-messianic cults inside of some of the largest developers high on their own supply, and astounding capital investments that, much like when we did this a few years ago with self-driving cars, aren't getting made back in a timely fashion.
I think comparing it to welding is instructive- because it suggests that the answer for what happens with any realistic automation technology is 'it's a mess and it depends.' AI investors and marketers are keen to present LLM capabilities as completely novel and transformative- and it's true, they do some neat things that were hard to do before- but they're also clearly part of a long chain of search technologies, text generators and linters, and inappropriate personification of simple programs, that stretches back at least two generations.
And my experience both training and using ML/DL-based systems is that they're certainly neat but not hard to back into an inappropriate use case- like any piece of software. When I checked in occasionally for writing help to generate a punchy line, the results were...fine, but I wandered off because because writing is conveying particularities inside my head that the machine isn't privy to (and it isn't that good). I suppose it could write customer service boilerplate- but the competition there is the templates everyone mostly uses already. I've flicked on assorted copilots while coding, and it's pretty nice- but the entire story of programming from machine code on has been about generating more standardized instructions with less typing- this is continuity again. And as a search add-on? It's notably, actively bad.
And then we have welding- a task where most of the volume of the task done in the world happens under repeatable controlled conditions, and so it was mostly but absolutely not completely, handed to machines- over the course of a century, with booms and busts tethered to bigger economic forces, and slowed by the fact that the next weird edge case that breaks the extant version of a technology is always much closer than you'd like to think.
The potential for AI to create massive job destruction was seen by Wassily Leontief (1973 Nobel Laureate in Economics) in his famous “Long-Term Impact of Technology on Employment and Unemployment“, National Academy of Engineering (1983). He describes the great wave of job extinction — of horses. Cutting their wages could not save their jobs, nor will it save ours.
Leontief’s insight was expanded by Erik Brynjolfsson and Andrew McAfee (both of MIT) in “Will Humans Go the Way of Horses?” in Foreign Affairs, July/Aug 2015 — “Labor in the Second Machine Age.”
You touched on it very lightly right at the end, but I want to emphasize that looking at the change in employment of welders doesn't tell you a great deal about AI replacement of welders. The issue is that the current number of employed welders at any point in time is both a function of supply and demand. AI affects the demand for human welders, as it is a substitute, but a simple lack of welders available will also lower the number employed regardless of what AI is doing.
Many economists will say "well, wages should rise until they have all the welders they want" but the problem there is three fold. Firstly, if wages are high in the US for welders, but lower elsewhere, the work will simply move off shore, so that doesn't help us with figuring demand for welders unless we look at global numbers (good luck). Secondly, welding can be replaced to some extent by other processes, such as simply not using metal or riveting, so again we can't tell the effect of AI. Thirdly, and more generally damning, most people do not fully consider their income path when choosing a profession in their late teens. Wages of welders are pretty good compared to low tier college majors, yet many people go into a lot of debt to get a largely worthless degree, for instance. Since people are picking careers based not on actual wages but instead status or potential wages*, wage increases in various low status fields like blue collar work should not be expected to solve any particular shortage in a reasonable amount of time.
* My personal empirics in this realm is my experience teaching college economics. When discussing labor markets I would also bring up a list of college majors by income 5 years after graduation. The vast majority of students were always shocked at the numbers, even if they generally understood that engineering students made a lot more than English majors, and invariably a few students in every class would ask "Why have we never seen this before?" It seems most discussion of career choices before that had been "Go to college and you will get a good job that pays better. If you don't, you will end up a poor trucker or welder."
Another terrific article! You've convinced me. Soon as I get somewhere with a secure connection, I am going for a paid subscription.
Two points I would like to add/emphasize
Robots used to be fairly dumb. You could program it to pick up a part at point A and place it at point B. But the part had to be *exactly* at point A and would be placed *exactly* at point B.
10-15 years ago smart, not AI, just smart, cameras and machine vision systems began being combined with robots. Now the part didn't need to be at point A. It needed to be in the vicinity of point A. It didn't need to be properly oriented. The camera would "see" the part and guide the robot to pick it up properly. This led to literally millions of applications for robots that were not possible before.
Same thing with welding. The weld would always be at the same place and if the part was slightlyh out of position, the weld would be bad.
Another thing you touch on but it can't be emphasized enough. Automated processes require automated quality parts. Quality in this usage meaning no variability. Trying to automate a process using the same parts that had been used manually seldom works. As y9ou point out, the operator can adjust. The automation machine can't
That is why automated processes always (or should always) produce better quality final products than hand crafted products where every product is different. Some better, some worse than the ideal.
I know it’s not the main part of your post, but…When talking industrial robots, whether welding or other tasks, the feedback control and path planning aspects are critical. The ability to generate smooth paths and accurately follow those smooth paths has been the subject of a LOT of work in the control system community. The successes - and limitations - of control systems in industrial processes might make a good topic for this Substack, especially considering that of the principal drivers in control system development over the last 80 years, industrial processes is one of the top two (the other is aircraft control).
Having spent the past nearly 50 years working in and with manufacturing automation, I am in the optimist camp.
I've been hearing since the early 50s "you can be replaced with a button" often as part of a screed about how automation was going to put us all out of work. It has not happened yet. Even without AI I do not believe it will ever happen.
Not only have we created new jobs (Like Changeover Wizard?) for all those people displaced by automation, we have generally created better better jobs.
Go read Henry Ford's 1923 autobio "My Life and Work" (Email me at johnhenry@changeover.com if you would like a copy) He describes how he invented what is now called the Toyota Production System..
He was also a fanatic on automation "NEVER put a man to do what a machine can do" He probably created more jobs, paying double or more the going wage, than anyoiine before or since.
Manufacturing output in the US, for all the doom and gloom we hear has increased steadily and dramatically year after year from WWII to today.
Manufacturing employment is down and that is a good thing. Most unskilled and semi-skilled (fully trained in a month or so of on the job training) are really crappy. They are probably not something to be desired, they are something to be eliminated.
So, let's go automation and let's go robotics. And to the extent that AI can improve automation and robotics, lets go AI.
Trying to figure out anything sane about AI (in its current form this AI cycle as LLMs) is massively overcomplicated by the boosterism of a marauding giant blob of money getting anxious that tech is about to turn into a dull industry. There's a hundred years of science fiction to coopt as marketing (even when it's apocalyptic; if your app could take over the world there's no question that it's competent), quasi-messianic cults inside of some of the largest developers high on their own supply, and astounding capital investments that, much like when we did this a few years ago with self-driving cars, aren't getting made back in a timely fashion.
I think comparing it to welding is instructive- because it suggests that the answer for what happens with any realistic automation technology is 'it's a mess and it depends.' AI investors and marketers are keen to present LLM capabilities as completely novel and transformative- and it's true, they do some neat things that were hard to do before- but they're also clearly part of a long chain of search technologies, text generators and linters, and inappropriate personification of simple programs, that stretches back at least two generations.
And my experience both training and using ML/DL-based systems is that they're certainly neat but not hard to back into an inappropriate use case- like any piece of software. When I checked in occasionally for writing help to generate a punchy line, the results were...fine, but I wandered off because because writing is conveying particularities inside my head that the machine isn't privy to (and it isn't that good). I suppose it could write customer service boilerplate- but the competition there is the templates everyone mostly uses already. I've flicked on assorted copilots while coding, and it's pretty nice- but the entire story of programming from machine code on has been about generating more standardized instructions with less typing- this is continuity again. And as a search add-on? It's notably, actively bad.
And then we have welding- a task where most of the volume of the task done in the world happens under repeatable controlled conditions, and so it was mostly but absolutely not completely, handed to machines- over the course of a century, with booms and busts tethered to bigger economic forces, and slowed by the fact that the next weird edge case that breaks the extant version of a technology is always much closer than you'd like to think.
The potential for AI to create massive job destruction was seen by Wassily Leontief (1973 Nobel Laureate in Economics) in his famous “Long-Term Impact of Technology on Employment and Unemployment“, National Academy of Engineering (1983). He describes the great wave of job extinction — of horses. Cutting their wages could not save their jobs, nor will it save ours.
https://nap.nationalacademies.org/catalog/19470/the-long-term-impact-of-technology-on-employment-and-unemployment
Leontief’s insight was expanded by Erik Brynjolfsson and Andrew McAfee (both of MIT) in “Will Humans Go the Way of Horses?” in Foreign Affairs, July/Aug 2015 — “Labor in the Second Machine Age.”
https://www.foreignaffairs.com/world/will-humans-go-way-horses
You touched on it very lightly right at the end, but I want to emphasize that looking at the change in employment of welders doesn't tell you a great deal about AI replacement of welders. The issue is that the current number of employed welders at any point in time is both a function of supply and demand. AI affects the demand for human welders, as it is a substitute, but a simple lack of welders available will also lower the number employed regardless of what AI is doing.
Many economists will say "well, wages should rise until they have all the welders they want" but the problem there is three fold. Firstly, if wages are high in the US for welders, but lower elsewhere, the work will simply move off shore, so that doesn't help us with figuring demand for welders unless we look at global numbers (good luck). Secondly, welding can be replaced to some extent by other processes, such as simply not using metal or riveting, so again we can't tell the effect of AI. Thirdly, and more generally damning, most people do not fully consider their income path when choosing a profession in their late teens. Wages of welders are pretty good compared to low tier college majors, yet many people go into a lot of debt to get a largely worthless degree, for instance. Since people are picking careers based not on actual wages but instead status or potential wages*, wage increases in various low status fields like blue collar work should not be expected to solve any particular shortage in a reasonable amount of time.
* My personal empirics in this realm is my experience teaching college economics. When discussing labor markets I would also bring up a list of college majors by income 5 years after graduation. The vast majority of students were always shocked at the numbers, even if they generally understood that engineering students made a lot more than English majors, and invariably a few students in every class would ask "Why have we never seen this before?" It seems most discussion of career choices before that had been "Go to college and you will get a good job that pays better. If you don't, you will end up a poor trucker or welder."
Another terrific article! You've convinced me. Soon as I get somewhere with a secure connection, I am going for a paid subscription.
Two points I would like to add/emphasize
Robots used to be fairly dumb. You could program it to pick up a part at point A and place it at point B. But the part had to be *exactly* at point A and would be placed *exactly* at point B.
10-15 years ago smart, not AI, just smart, cameras and machine vision systems began being combined with robots. Now the part didn't need to be at point A. It needed to be in the vicinity of point A. It didn't need to be properly oriented. The camera would "see" the part and guide the robot to pick it up properly. This led to literally millions of applications for robots that were not possible before.
Same thing with welding. The weld would always be at the same place and if the part was slightlyh out of position, the weld would be bad.
Another thing you touch on but it can't be emphasized enough. Automated processes require automated quality parts. Quality in this usage meaning no variability. Trying to automate a process using the same parts that had been used manually seldom works. As y9ou point out, the operator can adjust. The automation machine can't
That is why automated processes always (or should always) produce better quality final products than hand crafted products where every product is different. Some better, some worse than the ideal.
I know it’s not the main part of your post, but…When talking industrial robots, whether welding or other tasks, the feedback control and path planning aspects are critical. The ability to generate smooth paths and accurately follow those smooth paths has been the subject of a LOT of work in the control system community. The successes - and limitations - of control systems in industrial processes might make a good topic for this Substack, especially considering that of the principal drivers in control system development over the last 80 years, industrial processes is one of the top two (the other is aircraft control).