AI Is Starting to Threaten White-Collar Jobs
wsj.com
wsj.com
> Manna’s job was to manage the store, and it did this in a most interesting way. Think about a normal fast food restaurant. A group of employees worked at the store, typically 50 people in a normal restaurant, and they rotated in and out on a weekly schedule. The people did everything from making the burgers to taking the orders to cleaning the tables and taking out the trash. All of these employees reported to the store manager and a couple of assistant managers. The managers hired the employees, scheduled them and told them what to do each day. This was a completely normal arrangement. In the early twenty-first century, there were millions of businesses that operated in this way.
> But the fast food industry had a problem, and Burger-G was no different. The problem was the quality of the fast food experience. Some restaurants were run perfectly. They had courteous and thoughtful crew members, clean restrooms, great customer service and high accuracy on the orders. Other restaurants were chaotic and uncomfortable to customers. Since one bad experience could turn a customer off to an entire chain of restaurants, these poorly-managed stores were the Achilles heel of any chain.
> To solve the problem, Burger-G contracted with a software consultant and commissioned a piece of software. The goal of the software was to replace the managers and tell the employees what to do in a more controllable way. Manna version 1.0 was born.
It was not (source: family & friends did that job at several Home Depots recently:)
It takes about 16 - 24 hours of work on a weekly basis to manually update schedule after it has been automatically generated based on policies, requirements, and all known constraints.
Reasons, if you care - I got to hear about them weekly:
1. Department Supervisors are supposed to manage the schedules for their teams, but they did not at most stores, for various reasons from lack of training to lack of will/bother. So the job effectively gets "internally and unofficially outsourced" from a dozen leaders who know their teams, to a single HR / schedule person who now has 150-200 people they know less well (and other work to do as well). This was NOT well understood by corporate/HQ, so entire process is based on flawed assumptions from the get-go.
2. As part of that, nobody's preferences and working hours are correctly in the system. Again, should've been entered upon hire and maintained by DS, instead a single person needs to figure out and manage for 150-200 people.
3. As part of that, people made changes all the time - wedding, funeral, birth, movie, party, picnic, vacation, illness, whatever. Should: Be entered directly in the system by the DS. Actually: people leave voicemail, post it notes, yell in passing, or try to telepathically message, yes, the single person who's job it is strictly speaking not to maintain such information.
4. Software isn't great; training and documentation are worse; processes are not reflective of reality; and nobody does what they're supposed to do, from associates to DSs to ASMs. People TRY their best, mind you. But people are people, so see next point:
5. Basically, the problem there, as I see it: just like economists assume perfectly rational economic actors and physicists assume frictioneless spheres, programmers assume "well defined requirements understood by all and a process that is documented and accurately followed" when they say "this can definitely be easily automated" :->. Don't even get me STARTED on bright eager university graduates who come to companies to do robotic process automation of business processes in a quick agile way with fast ROI.... and the even quicker and painful lesson in real life they soon obtain :O
This is NOT to say I don't feel your pain - I very much do! Not sure if a little bit of background will help or make it worse :-/
Scheduling is NP complete, but the problems are small enough that I think computers can trivially find near optimal solutions, given some scoring function, better than a human can. Why doesn't this happen? Is defining the scoring function too hard? Am I wrong about computer's solving abilities?
Language models have proven themselves reliable with natural speech and highly technical speech (like code), they could help make entering the constraints more approachable. Also, as my example demonstrates, if such a system fails occasionally, it's not worse than the status quo.
I don't have much additional information, but perhaps to summarize:
* Initial iteration of system is (by definition) imperfect
* People who are supposed to maintain constraints and parameters, don't
* Person who typically does end up maintaining the system is not aware of all constraints and parameters, and doesn't have power or channel to change system
* Process for noting and updating constraints is poorly communicated and not enforced
* Through entropy, system, which was not perfect in the first place, gets worse over time, which feedback loops because people keep overriding it and updating it less and less.
My friends'/family experience has been that when a new HR Manager joins a store, they tend to "clean up" everything as much as they can; but then entropy sinks in again - people don't update constraints and requirements, system deteriorates, and provides more reason/excuses for people to not use it properly.
------------
On aside, in my own daily life, I've been working on ERP systems all my life, and as a techie it's a fascinating place to be. I assumed that numbers are numbers, business processes are well structured, defined and documented, and that people who's job it is to use a specific system will be at least somewhat proficient in that system. None of those things are necessarily broadly true :).
On aside-aside, similar thing can be observed in MS Project btw. MS Project has an incredibly powerful scheduling engine. It's amazing! But! 99% of MS Project users don't know how to use MS Project. They try to use it as Excel. They type in some tasks, the dates aren't what they want them to be, then they override automatic scheduling by putting manual dates rather than putting constraints into the system... and the negative feedback loop commences. I've rarely seen a MS Project Plan file which properly uses the scheduling engine :-/
Stereotypical discussion with customers after a "bad" scheduling result: "Hey, why did your system not schedule Dave to do any of these jobs? All others are at 100% already, but he has nothing to do." "The job all need a Fubar cert, Dave doesn't have one." "Of course he has a Fubar cert." "Not according to the system." "Oh, we probably forgot it, because all schedulers know it anyway."
As others have mentioned, you don't always have a complete and accurate set of constraints. So you generate a schedule based on what you have. You hand it out to your employees. And they point out conflicts that they forgot to tell you about or that come up on short notice. So you have to rework the schedule to take them into account manually.
You could regenerate the schedule with the additional constraints. But the new schedule is likely quite different from what you originally handed out. People will have assumed the original schedule is final and they will have made plans. You cannot simply replace the schedule.
You might also already be halfway into the schedule period and the new schedule might only be fair if you had applied it from the beginning of the period.
You might be able to account for all of that in your constraints but it's going to get really hairy.
One of the most challenging problems I've every been faced with is machine-scheduling a set of manufacturing facilities. I understand that Home-Depot is a simple case, but I would caution you against trivializing the problem. The amount of compute (and accurate data collection) necessary to beat a (semi-competent) human is astounding.
On side note, I'm mentally preparing myself for a wave of AI powered plant scheduling hucksterism where those selling the systems bill AI as a tool that can magically fix the persistent problem of poor data collection.
What do you think the hardest part of solving this with computers is? Surely if we can get all the constraints (some of which would be soft-constraints) into the computer it can come up with an better solution than a person can. Right?
Is it a problem because the real world is messy and not all the data can go into the computer? Or is it that we just don't know how to define a "score function" that accounts for everything?
Real time, accurate data collection is a non-trivial problem in a high-mix environment and if you've got a handle on the first two, compute infrastructure to crunch all of it and then push it back to the facility being scheduled is expensive.
Finally, systems that are this optimized can be slow to react to sources of error.
The alternative is to pay a couple of guys to spend most of their time walking or golf-carting around a few hundred thousand square feet of facility. Their job is to intimately know the various processes and employees involved, run a very basic scheduling protocol and modify it on the fly as the situation demands. In most instances, it is very hard to prove that the technologically complex version is a true improvement over the good ol' boy approach.
As such, the computational solutions usually end up implemented to boost the companies valuation to potential buyers.
edit: job shop scheduling can work wonders, but it generally fails hard in 'high-mix' environments.
https://en.wikipedia.org/wiki/Lateral_thinking
https://en.wikipedia.org/wiki/Outline_of_thought
https://en.wikipedia.org/wiki/Critical_thinking
https://en.wikipedia.org/wiki/Brain_training
https://en.wikipedia.org/wiki/Higher-order_thinking
https://en.wikipedia.org/wiki/Integrative_thinking
https://en.wikipedia.org/wiki/21st_century_skills
https://en.wikipedia.org/wiki/Method_of_loci
https://en.wikipedia.org/wiki/Mind_map
https://en.wikipedia.org/wiki/Emotional_reasoning
and so on...Why is it better to be smart? Why would you give more freedom and more money to someone who has a larger brain capacity? It's not something most people can do something about.
And even if people can do something about it, it's a race to the bottom. Initially we have given many more people proper education, raising the average bar, and now we have to compete with computers as well.
I'd prefer a society where we would value kindness, bodily flexibility, strong hands, or random other traits as much as having fast processing speed in your brain.
I'm saying it is something a person can do something about, just like an athelete can train for the Olympics. Also, it takes creative intelligence to solve our problems, which benefit even those that don't have creative intelligence skills. From each according to their abilities, to each according to their needs.
I have nothing against working with my hands and have enjoyed it myself in the past. My suggestion is for those young people who want white collar jobs and want to compete with AI, which isn't a great thinker - yet.
There are plenty of snake oil companies that have attempted to sell a silver bullet towards being able to increase somebody's fluid intelligence everything from luminosity to hucksters which claim to be able to teach you how to have an eidetic memory, but it's mostly a con.
Interesting that you decided to pick athletes because I would argue that genetics play a far larger part in determining one's aptitude for being able to participate at the Olympic level than anything else.
Do you think that the 10,000 other diligent athletes that can't even get close to Usain Bolt don't train just as hard as he does? What about the thousands of equally studious physicists who will never get close to a Richard Feynman or an Enrico Fermi?
I will grant that through diligence and dedication one can increase one's ability to pattern match to solve certain problems.
Why is it better to be smart? Okay....
Why is it better to be beautiful? Why is it better to be physically fit?
I don't think anyone in the universe would disagree that we wish society was based more heavily on a culture of empathy, i'm just not sure I understand what your point is.
preferences aside, you "get" things because you either make them yourself, or you barter the value you are able to generate. money is just the fungible medium to mediate such barter, it makes the value you generate exchangeable in a generic way instead of limiting you to barter only with people who has things you want / need. if you are smart, you can generate more valuable things and provide more valuable services for other people. in return you get more things. it is perfectly natural. it is not random. your bodily flexibility or strong hands don't produce things, by themselves, that you can exchange for things you need / want. you can convert those skills to generate value (for yourself to survive, or for others to exchange with things for your survival) but by themselves they don't mean much.
I am not saying it is a good deal for most people, but it is a perfectly natural order in how things work.
I think it would be much more prudent to learn how to install AC, or run electrical conduit.
Based on past experience, it seems more likely that what will happen is that skill sets will change, people will become more productive resulting in more stuff getting done that necessitates more companies employing more people, etc. Which sucks of course if your skills are the ones that are automated away or become obsolete, but that's something been going on for half a century or more.
The hypothetical arrival of AGI, of course, especially if it's doable on commodity-level hardware, would be a different thing and the societal upheaval could be catastrophic. But I feel like the current crop of AI is just a continuation of the previous thing.
People often bring up horses and cars, but think about the timescale of how long it took to make that shift. Horses still played a role in WW1, decades after the automobile was a thing.
The other factor is the non-existent cost barrier. Think about the transition from typewriter to word processor -- it took a relatively long time for computing to become affordable enough to be accessible meaning that the transition from typewriter to word processor manifest over a large span of time allowing for labor to adjust.
AI is practically being given away for free and prices are constantly dropping.
that's if hype will be supported by quality eventually. I think it is still not clear if AI can automate even some trivial clerks jobs without failures.
The authoring tool allows retailers and brands to generate hundreds of articles a day which would have taken a team of human writers days if not weeks to put together. Some of the work straight up couldn't be done previously because the number of people required would be too exorbitant for the returns.
Some of the content augments existing content on the retailer and brand websites. Some of the content is feeding into SEO.
And its good enough that we are ranking #1 in Google just a few weeks in for some customers in competitive search terms.
I think search engines won't have a spam problem for long... but AI training will start to have a spam problem as a side effect of developing into a marketplace. The influence market moves from people to machines.
This is all of the work we do before even getting to the LLM.
The other question is: what sort of economy would even remain? One where the same AGI runs everything and engages in adversarial games against itself?
But what exactly is AGI? This is the first problem. Where to draw the border between harmless and dangerous AI? Careful legislation needs to give clear limits. Too little and real AGI doesn't get banned. Too much and useful AI like LLMs get banned. Unclear law is unenforceable and people start ignoring it.
The second problem is detecting. Police can just look at media and they know. But they can't look at abstract data and immediately say, hm, looks suspicious. They need to look at software behavior, and this is a lot more work. This is like catching someone stealing red-handed. But what behavior exactly is illegal, even if they could watch AGI doing something?
And finally, AGI is probably clever and might hide by playing dumb.
Anyone who tells you otherwise has a shallow understanding of the tools and hasn't built anything more than demo projects.
In terms of work, with zero-friction reallocation, of course automation is a good thing. There are tons of other things which need to be done besides packing boxes in a warehouse or driving people around. But we don’t have anything like perfect reallocation, and we don’t have business models for many of the things which need to be done.
To put it in perspective when I was learning about neural networks in ~2009 it was still actively debated in the field whether neural networks were even the right paradigm for general intelligence. Back then one even knew how to make deep neural networks work well. They sucked so bad and had sucked for so many decades that most in the field just assumed that they probably weren't going anywhere without some significant advancements.
Then in the early 2010s the field exploded after a few minor algorithm tweaks allowed deep neural networks to really begin working. But even then they kinda sucked at a lot of tasks outside of niche applications like image recognition. But then a few years later we made a few more algorithmic improvements and suddenly we had pretty decent chatbots. And then a few years after than OpenAI created GTP-3.
What I'm saying is that basically went from most experts in the field doubting whether neural nets would even work to GPT-3 in about 10-12 years. That's absolutely insane progress.
With this rate of improvement the "foreseeable future" is probably less than a decade away. Assuming progress continues at this pace I strongly suspect we'll have LLMs able to beat the best humans in software development in 5 years.
Humans are on the cusp of learning that just like everything else so far, there is nothing special or unique about the meat version of intelligence. In fact it's likely hilariously weak compared to purpose constructed general intelligences. Similar to a bird racing a jet, a cheetah against a V12, or horse pulling against a truck.
was it consensus among experts? How did you measure it?
Your takeaway seems to be software will always do things ahead of when we expect, whereas my takeaway is we're incredibly bad at guessing what sorts of tests and benchmarks mean software will be able to totally replicate, best, and replace human reasoning and decision-making. Beating games, predicting protein folding, forming reasonable-sounding paragraphs, and scoring high on the LSAT have all turned out to not be enough.
Some work requires direct human level control. Other work just requires a human in the loop. We've already exploited most of the things that can be automated without human level reasoning, and LLMs have made inroads on a whole bunch of things that do require reasoning.
Augmentation - any significant force multiplier applied to individuals - will result in individuals performing more work or producing more output while the overall workforce is reduced in size. Corporations exist to maximize profit, not to provide meaning and worth to human existence.
AI is taking us to a place where human labor is detrimental to value. Hourly wages won't be sensible in the face of software that can do any cognitive work a human can, for orders of magnitude less cost and overhead, with no feelings or rights to account for. Current economic models won't make sense at all without scarcity of labor, intellectual and otherwise.
No profession is safe, and things will only accelerate from here. We should probably start thinking about what a social contract means in a post-scarcity world. What does money mean? How should resources be fairly allocated if working for money is actually counterproductive even if companies were willing to hire you? What does the market look like if it's irresponsible to allow fallible human control over a task which can be performed to a higher level of quality and consistency for orders of magnitude less cost?
It won't happen all at once. It might take a few years, or it might just pop up and happen this summer, or it could take decades, but AI replacing human jobs is a zero sum game. Once machines can more efficiently and effectively perform any cognitive task that any human can achieve, it's no longer strictly economical to pay humans for those tasks. We'll either have to do pretend-work so we can justify paying humans for work that's less cost effective than paying a machine, or we'll need some sort of Universal Basic Income. I'd love to see alternative ideas, but there don't seem to be any.
We're in a very interesting period of human history.
We will not create ENOUGH new jobs.
They think this is like going from loss of blacksmith jobs when fewer horseshoes were needed to jobs gained in huge Ford motor factories.
If we don't protect our resources and enable basic income, we will find ourselves banging on the gates of robo guarded walls begging for food.
The best they can do is say that middle management is feeling the squeeze. Google for "middle management squeeze" for a cornucopia of articles from every year for the last two decades.
I am personally skeptical that no-code solutions will replace software developers at a significant scale.
Creating a boilerplate project with no integrations or existing logic is one thing. It's a very different thing to use generative AI to maintain existing systems, integrations, and business logic without a human that knows these systems.
The landscape will change for software developers. Probably dramatically. But anyone predicting the end of software developers has as much credibility as someone who says they can talk to dead people.
Such as the example provided in the article:
>At Chemours, a DuPont spinoff, the company has trained close to 1,000 office and lab workers in AI applications over the past three years. As a result, finance professionals who used to prepare certain reports with a lot of copying and pasting between systems and spreadsheets now do it much faster because of their training with no-code analytic tools
Seems like a great use case for llms, large scale text generation that acts as scripting glue. AI will improve efficiency for dumb tasks meaning smaller teams can take the same workload. This will either lead to layoffs indirectly or more head count for r&d and other investment areas.
Imagine we went back and time and described a social media job where people work endless hours, so that other people can sh*tpost online... They'd probably look at you and wonder how people are clothed and fed in our time with so many diddling away (even stressed out!) at such jobs.
Hopefully all those people who will be out of work don't think focusing on activates like "eating" or "getting healthcare" are important because they won't be able to afford them.
People tend to take whatever jobs they can get that will pay them enough to keep working and will fit their schedules, but with very little consideration paid to what they think is important or meaningful. I don't see that changing, especially as more and more people will be increasingly forced to compete for fewer and fewer available jobs.
> Imagine we went back and time and described a social media job
It might sound strange to someone from the past, but companies hire people for "social media" jobs because doing that makes the company more money than not doing it. Businesses wanting more money should be relatable at least.
Tech companies axe 34,000 jobs since start of year in pivot to AI
Example: "Despite reducing its workforce by 5% or roughly 10,000 workers, Microsoft's overall employee headcount remains well above what it was on March 11, 2020, which is right around when the World Health Organization (WHO) declared COVID-19 a pandemic." [1]
[1] https://steve-taplin.medium.com/big-tech-employee-numbers-be...
How many articles nowadays use pictures generated by AI instead of stock images created by an artist?
That probably saved me 20 minutes of looking through the manuals.
That very carefully chosen prompt, btw, was: "what is wrong with this code"? I have since refined it to be "what is wrong with this code, assume the tables exist".
No, it is not going to mean that anyone can code, but if developers are 10 to 20 percent more efficient that means that the same problems need at least that fewer developers, more when you take into account communication overhead between devs.
Companies loudly announce they're laying off workers to cut costs. A short time later they're advertising those same jobs, but people are hired, or even hired back, at a lower wage.
What used to be a well-paid job 15 or 20 years ago, no longer has job security or the same benefits. It may have even be contracted as a 'gig' job, with no job security or benefits.
This phenomenon can be seen in how lower priced restaurants can cause people to go out more and then create more demand for restaurants. (I think this has been studied by someone somewhere)
That said this doesn't always happen, part of this has to do with the balance between lowering prices and increasing demand, etc., and part of this goes to whether or not automation can completely cover the important aspects of the job (for the typical use case) it's replacing. For example, how travel comparison and search sites cover the typical use case for a travel agent roughly (I'm guessing).
However in the coming age of restricted energy availability, capitalism will become obsolete and disappear. It's already happening, we're contemplating a new form of techno-feudalism growing upon us.
I'll add this aside : some of these people with "Nobel prizes" in economics go as far as saying things like "+4°C will only marginally affect the economy as only outdoors activities like agriculture, which represents a couple percent of GDP, will be impacted". Because as you know, agriculture as part of GDP perfectly represents its actual importance in our lives: economists don't eat I suppose :) A joke I like is "the brain is only 2% of the weight of an economist, therefore its ablation will only have a marginal impact on this economist's ability to function".
Overall, Americans have been increasingly working more for less pay for decades, all while productivity (and profits) have risen. Even at the SWE engineer, we should never expect innovations to ease our labor burdens unless we fight for that to be the case. By default, we'll do one of two things:
1. Be expected to do a lot more work, since AI makes it easier
2. Get laid off, because 1 person is now doing what a team did with AI (even if they're doing it worse)
this is the kind of society you want to live in?
I am a 79 year old who has successfully changed his career several times. I now have a business in cybersecurity but in my past I have been several things including a Vocational Expert.
I call bs. At minimum, an extreme exaggeration.
hahahaha
I see it as a move towards the middle where the need for high paid knowledgable people in certain subject areas will decrease, the need for people who can perform remedial tasks and rote memorization will decrease, but the pool in the middle with enough knowledge to query an LLM in a fashion to get the answers will be a necessity.
I have worked with so very few principal engineers and architects that "get it" because they are not hands on.
Find some good staff / principal engineers that care about growing and mentoring people and use the tools available to augment their efforts with less body counts.
And on the flip side, junior engineers can be much more performant in their roles and grow into that mid-level engineer much faster by allowing them more opportunity to learn conceptually with less rote coding.
The doers are going to thrive and the talkers and fakers are gonna struggle.
Also good managers inspire. Inspire growth, productivity, learning. Good luck with chatgpt
I agree there appears to be a great potential where this becomes a force multiplier reducing the need for the more experienced engineers to sit on top. To other sibling comments, yes we need better methods for managing tech debt with these emerging tools but that doesnt make OP wrong.