The manufacturing sector have been heavily into automation without AI for quite some time.
Take a tour of a modern auto assembly line and if you're like me, you'll be shocked by 2 things --- how few people are involved and the lack of lights (robots don't need them).
At the Hyundai assembly plant in Mobile Alabama, only about 24 hours of human labor goes into building each car.
At an average rate of about $30 per hour, less than $1000 of human labor goes into each new car.
This doesn't leave a lot of room for AI to have a major impact.
The column mentions manufacturing automation and claims that even though automation has been wildly successful, that automation gives less than 30% increase in total productivity (edit to add: in those particular tasks where it is most effective). That's part of his intuition for why LLMs are unlikely to give much more than that on any particular task.
But even a in aggregate 30% increase in TFP can have massive implications on a job market, as is already being seen.
The issue is a lot of people (especially policymaking adjacent) have an incentive to either use a "skynet is coming" story or a "there is nothing happening" story.
The reality is it's somewhere in the middle, and plenty of white collar jobs are heavily ripe for significant reductions in headcount.
The article also doesn't claim there will be no impact, and indeed acknowledges that for particular roles it could be very consequential. The headline is just that when you average over everything the economy does, a 0.5% to 1% productivity gain is a more plausible outcome than 10-30%. That's the "somewhere in the middle" conclusion that this particular economist comes to (as of last year).
The reality is somewhere in the middle for now. Over long enough timescales, that reality will trend close to full automation (or all the way), unless society rejects AI and policy intervenes to stop it, or unless there's some hard final barrier that no amount of time, money, compute, ingenuity or labor could surmount. The latter seems improbable (are we at AI's vigil - or its nativity?) so Policymakers anticipating the former are only being responsible.
Excuse me if I'm mistaken, but you seem to assume that the assembly plant is already maximally optimized and that only the human labor could be improved by AI. That would discount the potential effect of AI on re-engineering plant with unorthodox insights or through comprehensive fast simulations that haven't been feasible before. Then there's any new engineering techniques it may arrive at, new materials, the scope it will bring to robotics etc etc
So basically, you're suggesting AI will replace or have a significant impact on the industrial engineers whose job is to promote efficiency?
That's expecting a lot from something that still struggles to count letters in words or take orders at a fast food drive thru.
> something that still struggles to count letters in words or take orders at a fast food drive thru.
You're expecting quite a bit more if you think that here in 2025 we're at the end state of AI development.
AI has an unpopular but really clever cousin namely intelligent automation (IA). They're already helping the humanity since we know how to automate [1],[2],[3],[4].
[1] Logic, Optimization, and Constraint Programming: A Fruitful Collaboration - John Hooker - CMU (2023) [video]:
https://www.youtube.com/live/TknN8fCQvRk
[2] "We Really Don't Know How to Compute!" - Gerald Sussman - MIT (2011) [video]:
https://youtube.com/watch?v=HB5TrK7A4pI
[3] Google OR-Tools:
https://developers.google.com/optimization
[4] MiniZinc:
https://www.minizinc.org/
No one should be surprised. Manufacturing is producing the same thing over and over again. Little to no AI should be needed.
Ok, so we agree that manufacturing is pretty much out.
So how about service jobs? How about one of the lowest level service jobs imaginable --- taking orders at a fast food drive thru?
IBM and McDonalds spent 3 years trying to get AI to take orders at drive-thru windows.
Here are the results:
https://apnews.com/article/mcdonalds-ai-drive-thru-ibm-bebc8...
I can actually see this bein' one task that current levels of language models would excel at, honestly... Given the limited list of items on a typical fast-food menu, and the accuracy of even some of the lowliest modern language models and speech recognition, I see no reason why fast-food order-taking needs to be handled by humans at all anymore, especially if you confirm the final order with the human ordering before proceeding; I could honestly see that bein' much more accurate than a human doing that job. (I can't count how many times over the years I've had a human order-taker completely screw the order up despite them repeating the order back exactly as given. A well-designed LLM-based system likely shouldn't have that problem. What it repeats back should end up bein' exactly the order that the system pushes through to completion.)
would excel at, honestly.
You would think so --- but well financed tests in the real world suggest otherwise.
Typical AI fanatic behavior - presented with the evidence that it doesn't work and goes "hmm, this should work perfectly!"
If that doesn't sum up AI hype and apologia then I don't know what does.
Yeah, no. I'm not an "AI fanatic" by a long-shot, but whatever... I use A.I. sometimes, and other times I don't. When I do use it, I use it for what it's good at. When I don't, it's because it's simply not capable of the task at hand. Simple as that. :shrug:
It's not clear what the problem is. Is it that the mic quality is not good enough for an AI? Is it that the AI is not smart enough? Is it that people generally don't like AI taking orders? Is the latency not good enough?
Or is it that people prefer to preorder on the phone instead and pick up?
After years of effort, I'd say that most of the simple non-AI issues were examined.
Lots of videos on TikTok illustrate the problem.
https://www.businessinsider.com/tiktokers-show-failures-with...
How much is it because they were still too early?
Would they do significantly better with a model like Claude 4 than I’m guessing something worse than GPT3.5?
Once you incentivise people to use AI by unbundling the costs of using humans vs using AI, you will see a lot of people fall in line. Although legally I do not think it will be an easy implementation. I am sure a lot of people already order via web or app nowadays.
There's no AI needed for that either. Just order on your phone, drive there, scan, machine dispenses your order and goodbye.
Much like how when you go to one of these places >>right now<< you just walk up to a kiosk, input your order, pay, then collect your order at the desk.
Couple more years and we'll rediscover that vending machines exist.
Ok? Taco Bell is taking ai orders at drive throughs right now.
Not "full self driving", human supervision is still required.
"a Taco Bell employee is still always listening on the other end of the ordering system with the ability to intervene"
Yup, automation has been happening since (before) the 1970s, finance since the '80s, media since the '90s, digitalization since '00s and all of it more ever since. AI currently has the least impact of any major development and will for some time. Hacker News isn't the real world.
> Yup, automation has been happening since (before) the 1970s ...
Yeah, significantly before the '70s, unless you're specifically talkin' about robotic automation. Folks been automating human labor with automated machinery of various kinds for quite a long time before that.
Ned Ludd did nothing wrong
None of that automation thought for itself, or could undertake its own automation. The potential in all those former waves was limited by the skills of human beings, but the limit on AI eventually is only compute. These are not the same thing at all.
A lot of automation isn't what people think of as automation. Is the average supermarket automated? A bit, but not a lot. But 90% of what they are actually selling has been automated in some way. Groceries have been cooked, picked, preserved, packed, shipped or otherwise processed with the help of automation. So regardless how smart an AI becomes it doesn't provide much more value as a cooking robot. We have automated cooking for decades by processing food. But we are still limited by physics, society and need.
AI will probably make music free. But it is already almost free with cheap instruments, recording equipment and distribution. And even before music wasn't that expensive. You can argue that we lose value in not performing it ourselves. That is some impact, but not one that strictly replaces the other. You can choose to have society where you teach music and it will still provide value over AI.
I do realize that the idea is often not that we will have cooking robots, but that AI will change chemistry or biology to where food is something else. Still hard to say if or when that happens, and what impact it would actually have.