Alex Tabarrok has a problem with AI:
Recently, I have seen two innovations in retail, AI cashiers and human cashiers but working remotely from another country such as the Philippines and making much lower wages than domestic workers (examples are below). I fear that the AI cashiers will outcompete the Philippine cashiers leading to the worst of all worlds, AIs doing low-productivity work.
I'm not picking on Alex over this, but it gives me a handy excuse to take on two aspects of AI at once.
The first is the complaint that AI will mostly just displace low-productivity tasks. But this is only true in the short run. It might be true of AI right now, this very second, but to focus on what AI can do at this moment is badly misguided. There's no question that current AI models have limitations, but those are (a) inevitable and (b) won't last long. Don't obsess over this. Bloggers and university professors are in AI's gunsights too.
The second aspect is that, in the short run, AI will displace low-productivity tasks. Indeed it will. AI is slowly but surely going to replace (practically) everyone. That definitely includes low-productivity workers like cashiers (3 million in the US), truck drivers (2 million), grounds maintenance (1 million), etc. This is still a few years away from happening in volume, but it is going to happen and it will make millions of workers unemployable. We remain woefully unprepared for the seismic shift this will cause.
That's nothing. The use of tractors and other machines caused the replacement of over 90% of farmers (and their horses and mules). As a consequence the unemployment rate is now over 50% (about half the population were farmers in the 19th century). Or is it?
This is lazy. Kevin's contention is that AI will replace *all* labor, not just specific industries. The agricultural revolution took place over many decades, and there were many other forms of employment available. (And it still caused great upheaval.)
The AI revolution Kevin posits will take over (practically) all forms of labor, and will happen quite quickly. University professors won't just be able to go become museum consultants (or engineers, or sex workers, or...etc) because AI will take over *all* those jobs.
You can disagree with his contention, and many do. But that argument has become much harder to make in the last year or so.
Just curious how you envision AI taking all the sex worker jobs.
We already have "inflatables." Imagine cybernetic, fully responsive models. And ChatBot phone sexers. Hey, Harlan Ellison wrote about this stuff back in the 1960s.
Harlan Ellison wrote about it in the 60s, and it's still a fringe thing, sort of like how hands and dildos have existed forever, and yet people still go out looking for other human beings.
Sex workers, probably a bit more time. Porn? Imagine the ability to enter in requests for the clip you want, and wait just a short time for it to be generated to your specifications. We're not that far away from "Deep Fakes" on steroids.
This is lazy. Kevin's contention is that AI will replace *all* labor, not just specific industries.
I don't think it's lazy. That is indeed Kevin's contention, but so far there's not a shred of evidence it's coming to pass. Sure, you say, "But AI is still in its infancy!" But is it? It's becoming a huge sector, and, while it's still got a long way to go, we are seeing it rolled out in myriad fields, and yet this dynamic coincides with one of the strongest employment markets in US history. Wouldn't we at least begin to see some signs of the impending jobs crash Kevin is predicting? When are those signs going to start arriving?
I have no doubt large swaths of the labor market will be eviscerated by AI. But technology as a jobs-killer is nothing new. Talk to blacksmiths or travel agents. But Kevin is suggesting a crash in employment on net—that is, a breakdown of the traditional process whereby employment growth in emerging sectors made up for the losses in the old. I'm now officially a skeptic on that part. I just don't see the evidence.
And one other thing I'll add -- I haven't seen any evidence that any jobs have been lost to AI yet. Maybe some radiologists have gotten a little more efficient, though I think all those experiments are still just that -- experiments reading X-ray scans.
Lazy perhaps, but every time someone says 'this time is different' they have to come up with a better justification for that belief than "Wow, that's cool."
Yes, LLMs are fun. They're also completely untrustworthy, perhaps fundamentally so. Their writing is *boring*. And while people talk about how they're getting exponentially better, what they really mean is that they're getting exponentially more expensive to train. Which means that we're rapidly running out of data with which to train them, so this may be as good as they get.
And then factor in copyright infringement lawsuits, and things get really ugly. These things can't be trained without massive and continual copyright infringement, and sooner or later, that's going to stop.
The farm workers were easily absorbed into the huge labor pool required to manufacture ‘tractors and other machines.’ And they made more money per hour worked because they were in high productivity jobs. The farm workers who remained were also more productive than before they had tractors and other machines to use. This is how the Industrial Revolution economy worked. But we are not in that economy anymore. We are passing through the tail end of the Information Age economy that started out by transferring labor from less to more productive employment, and will end with workers displaced by technology having few options for more productive employment because the new tech requires so little labor. The AI economy will finish that job and replace almost all labor because the machines will design and produce the machines. Nobody has explained how this will succeed if nobody has money to buy anything, though. A glib answer like ‘We can have and do whatever we want because everything will be almost free,’ - but that particular economic model won’t convince a single investor to finance the tech development needed to make it happen.
I have never had a cashier job that didn't also include cleaning the floor and stocking shelves and answering questions from people who can't use instrutions.
So pretty sure AI isn't replacing cashiers. Also, an automated ordering system isn't much in the way of AI.
I dunno. If a grocery store just swaps out AIs for cashiers, I agree that won't work. But what if it's just a front-end (on a website?) to a food warehouse? Robotic warehouse automation is very much a thing, and will only accelerate. Robotic AI-driven warehouse loads a robotic AI-driven truck that delivers to your home. Problem with your order? Call the ChatGPT v6 powered customer support line.
That may sound dystopian. But to me it doesn't sound impossible.
That's easy. You replace the cashiers anyway and hire a cleaning company who charges double the hourly rate of the cashier for the 10% of time they spent on cleaning.
Then you fire the cleaning company once VacBot 9000 rolls off the assembly line.
I've seen a lot of automation recently in ordering food at restaurants and paying bills there. AFAICT, it hasn't impacted restaurant employment at all.
Basically, I agree with @Crissa -- automating the last 10% of a job will be 99% of the effort.
Of course, if we are all out of work, who can afford the services now provided by tech?
I agree that automation poses a serious problem, and the first adapters might get a short term benefit, but Henry Ford's willingness to pay his workers enough to by a car increased his market. Removing the incomes of your customers is not a winning play, medium to long term.
This has long been my concern. When almost nobody is able to work for pay, who buys all the stuff the machines make? Oh sure, export it to other countries. Fine. But you've still got a US population that's close to 100% unemployed, and thus having zero income. How are they/we supposed to live?
Let's presuppose continuing Republican blockage of government action, so "basic income" proposals will remain as dead as Ronnie Reagan.
But you've still got a US population that's close to 100% unemployed, and thus having zero income.
No, you don't "still got" that. That's an utterly speculative claim made by dystopians about the far future that so far has no backing whatsoever in terms of evidence.
It’s the obvious result if you don’t account for the fact that the huge investment required to get to that point will never be made by capitalists who will see the flaw. It is one thing to invest in labor saving tech that will eliminate labor in one area and require MORE of it to build all these new labor saving devices. That was the industrial Revolution. But it is quite another thing to go too much past where we are now with tech replacing jobs but creating little need for new, better jobs. If AI can have one person replace 20, ultimately there will be 19 fewer jobs. But this will end as the customer base for everything shrinks too much.
We're already running into this. People are just pretending it isn't so. Right now, eliminating a low cost taxi driver means hiring 1.5 higher cost self driving taxi dispatchers. The theory is that self driving cars will keep getting better and better and that they'll need less and less human supervision.
Will this happen? People have been talking about self driving cars for at least ten years now. As Bullwinkle used to say before trying to pull a rabbit out of a hat, "This time for sure, Rocky."
The existing economic system cannot survive. NONE of the existing economic systems anywhere in the world can survive. Andrew Yang has some suggestions. My view is that Yang is being unimaginative but at least he's giving the matter some thought.
100% agreement. We will have to re-imagine economics. It's like a science fiction question, like just what IS the economy in Star Trek where they appear to have basically every material thing that they need.
I wonder if something like a Soviet-style state socialism (albeit one hopefully far more benevolent) might be best suited to this new techno- economic system. Or maybe we will all just start merging with AI and robotics and all this will become moot.
On a personal level, I am glad that I am near retirement, with a Federal Pension and Social Security and investment income to live on. But I have two kids just entering the workforce. Maybe I should not be in a hurry to retire so I can leave more money to them; I suspect the rentier class might be some of the few people who will be relatively economically safe in this new world.
Don't forget admin jobs.
In AR/AP, you need a human or two to blame when things go wrong.
Offices with high-status visitors will need receptionists, and execs will still want an office servant, at least until their peers go robot.
And there are various roles where "the human touch" is a significant part of the service, and the consumers will be able to still pay for it after.
There are some roles, especially in small businesses, where it will not be economic to replace the weird mix of physical tasks, specialist knowledge and vendor relationships that make things work, at least at first.
Most other backoffice roles are going to be automatable soon.
I do agree with Kevin that we are not ready for this. Hell, our economics experts - constantly talking about unintended consequences of good intentions - cannot even see the implications of the intended consequences here.
I am always surprised at how few people see this. It's probably going to be a rough ride for the 2 or 3 decades. In the near-to-medium term, I think groundskeepers are a lot safer than lower-tier lawyers and coders.
I work for one of those AI powerhouses, and I haven't seen any evidence that LLMs are contributing anything to software development tasks yet.
AI still can't think critically or creatively to solve problems, which even for some menial tasks/jobs is still necessary. That's what the quote was talking about for "worst of both worlds" - and I'm sure it will happen for some things.
But human hands are still going to be needed in roughly the same numbers that they're needed in right now. I don't think it's going to be that rapid of a shift. For all that "AI" (really, LLMs and algorithms) can do for us right now, their applicability is very limited.
Barring a seismic advancement (see what I did there?) towards AI that's more true to the label, what you're predicting isn't currently in the cards because there's no convincing sign that such advancement is coming. On the contrary, it sure looks like there's a big ol' hard stop coming in "AI" capabilities.
All the journalists claiming that AI will just improve until - *whamo* - it becomes intelligent would be amusing if it weren't so pervasive.
- Even scientists at OpenAI are saying that their LLMs are not suitable for all tasks.
- There is no cure for GenAI "hallucinations." They are a product of the statistical representations at the heart of the algorithm. Few companies can afford to have an unsupervised "agent" making promises that cannot be kept.
- AI companies have run out of data to train their models. If your super-intelligence requires more data than exists on the internet to be fully trained while human beings get by with their parents, mud pies and Thomas the Tank Engine, then I would say you are hurdling toward a dead end.
- As you said, the current models are cute but nothing like the "breakthrough" that will be needed to reach AGI and no one has any idea what that breakthrough is going to look like.
Hello, AI autumn.
Well said. Maybe Kevin and the journalists optimism is well-placed, but when the current leaders in the field are seeing a dead-end with current methods I listen up. I mean the state of the art now is using the models to generate synthetic data to train on. This is an ouroboros and can't get you very far.
Then there is the whole meatspace problem. It is one thing to have an LLM that can answer questions, it's a completely different one to have an actual robot that can traverse the real world. We don't even have reliable automated vehicles and that is a relatively easier problem.
@Scot_F -- I think you're absolutely right. People aren't paying attention to how much data is required for training these suckers, and they're running out. Each new generation of these models requires exponentially more data, and they're probably close to running out already.
On top of that, they're going to get their asses sued to, as Ralph Cramden might put it, the moon. I don't see how they'll get the right to steal all the IP in the world; there's way too many powerful people who want to protect their IP. The owners of that IP will insist on laws protecting their interests.
PBS did a NOVA show on AI, AI developers have been able to code creative thinking in their algorithms. They showed an example of the Chinese board game "GO". The AI made an extremely odd move early in the game that all the human observers thought was a mistake. The AI ended up winning. In retrospect it turned out that move allowed the AI to win.
I see generative AI running into the same challenge as self-driving cars: The technology can handle 95% of a given job pretty easily, but getting that last 5% is incredibly hard. And because those "hard" tasks are interwoven with the regular ones in unpredictable ways, you can't simply hand the 5% to a human and leave the rest for the machine.
On top of that, don't ignore the amount of work required to take a prototype to production. And then consider how long it takes new technology to be widely adopted.
Don't get me wrong, I do think we will get to AGI eventually. But it's not going to happen nearly as fast as the hype would have it. Most people's jobs are safe for the next couple of decades at least.
As is often the case in these discussions we need some definitions here. I will just say to that end:
Robots AI
Automation AI
A smart text generator will only get you so far. It's not going to clean a toilet or, for that matter, install or repair an AI enabled kiosk of some kind. I am trying to think how an AI would replace me at my job or all of the people who work for me and I frankly can't imagine it. There is just so much creativity, personal communication, and detailed knowledge of a specific business involved. To even get an AI to help out a little bit would completely amaze me.
For some reason, Bender Rodriguez comes to mind...
AI is currently at a stage of development where it won’t replace entire categories of jobs. But it WILL make people in those jobs much more productive. If it makes the job of a low level lawyer/paralegal, many varieties of analyst, coder, editor, copywriting in general, strategic planner, medical diagnostician, administrator, etc. even 2x more productive, that will reduce employment in those areas by half. More likely, within 5 years it will make them 5x more productive.AI will do what people in these jobs once did, and a few humans will remain to coordinate input and check output and prepare final versions of whatever they are paid to produce. And a tiny few more will figure out new applications to expand the usefulness of AI. This is exactly what happened after Google and other tools gave us instant and free access to a world of information. When combined with the advent of desktop publishing, digital imaging tech and variable data printing, the result was a complete transformation of an industry I was in at the time: advertising. I used to write expensive marketing analyses and marketing l/business plans for clients before the internet was useful for anything. We charged tens of thousands of dollars for these things because accessing up-to-date data and almost every kind of intel I needed was hard and took a long time. But a couple years later I could do it in half the time. A couple years after that we could charge a fraction of what we charged before because the clients could access Google just like us. And then the entire art/production end of the biz was obliterated. One person could do in a day what used to take graphic artists, paste up artists, photographers, typesetters and print shops a month to do before. Then the clients hired people like that and the agency had to figure out what to offer that the clients couldn’t do. This advanced the capabilities of everyone in the industry, which is how progress works. But the net result was that far fewer people are needed now in that field than were needed before. Some of those people were young enough to learn and grow with the tech. But in the end it is all musical chairs, and the chairs always win.
Barf. Law firms already are heavily automated. When we started a company in 2008 I compare the legal incorporation docs with the one we started in 1999, and the only differences among maybe 200 pages were a couple of paragraphs that were obviously added to disambiguate something they had to argue in court.
You don't need an AI to handle maybe 95% of legal docs, you just need a template.
How does energy use factor into the predictions?
At which point AI is expected to require just sensible amounts of energy instead of today's gargantuan demands?
Apparently AI will just magically solve the energy crisis for us, enough to surpass their own energy use! At least that's what all the AI boosters have been telling me. XD
Who will buy all the goods and services that the robots will now be producing in the next few years? If everybody is thrown out of a job then demand will collapse - so what would be the point of investing further in robots? Would the robot revolution be completed while the economy is collapsing?
The robots-are-coming people are far from having a grasp of everything that is entailed in automation. They refuse to learn from the past.
Ha ha. The robots won't be free. Robot companies are going to want to make a hefty profit. Your business will likely find 10 people to work for $15 per hour rather than pay a $500,000 per year subscription for a single robot to do their jobs.
We are in for several decades of dumbing down what AI means and AI always being a few years away from putting everyone out of work.
We can already sew this as we have recategorized automated cash registers, cars and lawnmowers as AI.
Literally all automation, hardware and software will be sold as AI in order to capture that extra juicy premium.
I'll get in early on $1,000 worth of AI tulips!
I think that literally the only two things Kevin Drum is more concerned about than the normal person is "AI will replace all the jobs" and cancer, and he seems pretty blithe about cancer.
Anyway my bigger point is that I think Mr. Drum knows more about technology than he does about economics. An economy will always find an equilibrium point, economies only collapse when there is also a failed state situation going on - that seems more pertinent to discussions of the future than whether or not a robot is going to be cutting your hair or jerking you off.
It's a combo effect of robotics + AI that will displace the low-wage jobs, whereas the white collar high-wage jobs will not require robotics.
In about 50 years or so we'll have robotic-organic replacement bodies and AI that is indistinguishable from humans. When the two are paired together, the end of humanity will follow, thus saving the planet from climate change.
AI won’t have energy requirements?
You mean they'll be a cashier at a dollar store?
Right now, stores are pulling out self-checkouts because they don't save that much and too many people can just walk out. How would an AI cashier be better?
It would invoke the security system to fire a tranquilizer dart into the person trying to leave the store ? It would invoke the robot to hog tie the offender ? It would communicate a "request for arrest warrant" to the police API so that police sends an android officer to arrest the offender ?
Oh dear lord. 'AI' cooks, cashiers, gardners? Tell, me, Alex, where do they keep their hands? You're an idiot, Alex.
ai will be able to sell any product or notion to us. Can democracy survive that? We’ll weep at the beauty of ai music and lyrics and poetry and art. It’ll crack us up with its jokes —we’ll be falling in the aisles. Standup comedians talk about “killing” the audience — maybe ai comics won’t be exaggerating.
ai. Not Albert.
AI generated music video,
https://www.youtube.com/watch?v=D9Mxl341_Rw
"(b) won't last long"
What is your evidence for that claim? Oh, that's right! You don't have any. You're (again) extrapolating a trend under the assumption that it will continue just as it has in the past. That is almost never true.
Look at self driving cars. You did the same thing there. But the trend you extrapolated was based on low hanging fruit. When the technology hit difficult cases, the development pace slowed a LOT. If I'm remembering correctly, full self driving is about 5 years behind your prediction.
Extrapolation is fraught with peril, and you should be a lot more humble and circumspect in your claims based on it. After all:
https://xkcd.com/1007
+1 for a perfect xkcd
And I don't see AI replacing university professors any time soon. You can't do original research by taking existing research and throwing it in a blender.
Kevin has more faith in these language models than I do. You can optimize your tree-climbing skills all you want. You’ll still never reach the moon.
This is just religion for Kevin Drum; he never responds to any of the cautions or counter-arguments in the comments. I’ve actually worked on automation systems for much of my career, and I’ve stated reasons why the future won’t look as KD imagines. So I’m done with the issue; climate change is a greater threat to the future well-being of the human race, let’s tackle that one.
I've grown more skeptical of the Kevin's jobs crash thesis. It's possible he's right, who knows? But I don't really begrudge his not frequently responding to his commenters. If anything I think that's a better approach than heavily moderated sites where you never know when the hammer is going to come down on you. It can be a bit intimidating. Kevin, for better or worse, doesn't seem particularly bothered as to whether his readers agree with him or not.
Oh, it’s his blog, he can do as he likes. I would prefer he elaborate arguments, or provide evidence, but on this particular topic he just wants to be the Oracle. Bah.
We remain woefully unprepared for the seismic shift this will cause.
1) Maybe one reason we're unprepared is that nothing of the sort is going to come to pass.
2) But if it does come to pass, the actual remedy—giving money to people—actually isn't all that complicated from a technical standpoint. We saw hints of this in the depths of the pandemic, and the US seemed surprisingly adept* at implementing it.
*Or maybe not so surprising. I've always maintained America is perfectly skilled at "socialism" based on those rare occasions when we've been allowed to practice it.
I've no doubt that AI will make it technically possible to perform all sorts of tasks with machines. I remain sceptical that it will ever make economic sense to replace a minimum wage grounds maintenance worker, for example, with a very expensive sophisticated machine. I don't mean sophisticated only in its software, but in the complexity of manufacturing a mobile machine able to do all the things a human body can do, from tip-pruning delicate shrubs to digging ditches to rescuing the neighborhood cat stuck up a tree.
AI is what we don't know how to do yet. A long time ago, playing chess was something intelligent people did, so teaching a machine to play chess was AI. Min-Max optimization algorithms solved that problem, and suddenly playing chess wasn't AI any more. Then Expert Systems were going to replace all the experts. But that didn't live up to the hype. Neural Networks were hot in the sixties and again in the 1980s, and now for a third time in the 2020s. We will soon learn what the current machine learning large language models cannot do.
One application of technology is looking at existing processes for doing things, and figuring out new ways of doing them that are better, either better results or lower cost. If your livelihood involved doing things the old way, it was tough for you. Your investment in learning skills and purchasing tools to do things the old way no longer pays off, and you have to find a new line of work, learn new skills, re-start as a beginner in new technologies, and that's hard.
This isn't new. It happened when steam engines came out. It happened when electricity became universal. It happened when internal combustion engines happened. It happened a couple of times with computers: first with mainframes, and again with PCs, and now with smart phones (which are really portable networked computers). It will keep on happening.
Kevin has no idea what he is talking about. I'd be thrilled if AI could do grounds keeping better and at a lower price than the humans, mostly Hispanic, that are doing it now. There is approximately 0 chance that that is going to happen in my lifetime. For example, as near as I can tell, robotic tree pruners (not harvesters) are just barely an idea right now. Heck, let me know when robots can whack weeds without destroying desired plants for 50 bucks a weeks and I will sign up now.