It's so fascinating that physical labor seems to be the main concern when it comes to robots doing it better/cheaper than humans. If anything, we know that automation is coming after office desk jobs first. Those jobs are much easier to automate. Language models can read a manuscript and spit out a summary/judgment with much better my friend that has read so many books and evaluate book projects. The "robot" has read more books, can memorize more of the manuscript as it reads it and does it much much much faster.
Maybe publishing houses are not comfortable replacing her with an AI but it will eventually happen.
We've been automating away office jobs for a lot longer than we've been putting ML in robots to automate factory work, though.
For example, the way business mail used to work was that the bureaucrat in question would record their message onto a tape, and then send that tape off to a special department full of typists to actually turn that voice recording into a letter. That whole concept is not only gone, but it's such a foreign idea that it sounds like something you'd write for a dieselpunk novel. The moment we started putting computers on people's desks, we expected everyone to know how to type. Same thing goes for a lot of other office tasks, which are now comfortably managed by software suites we literally call "Office".
That being said, the new wave of machine-learning powered automation scares me. Not because I'm worried that my job will be taken by software, but because said software will barely work. For factory jobs, the risks are obvious; that's why we put these robots in cages[0]. However, these office jobs are still making critical decisions that will increasingly be handled by automation. We already know how much having to deal with Google sucks; and they are pretty much addicted to automating away all their support staff. In your manuscript example, it could be that the ML model just starts burying specific genres of book or books with specific types of characters in them, for stupid reasons.
[0] Or if you're Amazon, you put the workers in cages, because Dread Pirate Bezos hates them.
> Not because I'm worried that my job will be taken by software, but because said software will barely work.
If said automation works like most corporate initiatives I've been a part of, it'll require 5 employees to implement, update, and maintain for every 1 that it saves, meanwhile costing millions of dollars per year to some vendor for a support license. Some workers might be let go but they were on the chopping block anyway. A few years later the whole thing is scrapped and the cycle starts over again.
One of the most interesting automations I saw was a system intended to let the company handle more calls with fewer people by pushing people into not talking to a human.
Except the reason they had so many calls in the first place was largely because every other business process was fucked, which kinda meant you needed a human.
I call this automating feeding coal into a dumpster fire.
"Broken process vs broken execution?" is one of the first questions everyone should keep in mind during automation discovery. But it's (usually) a nuanced call.
This is why automation to cut costs in a lax labor market is much less interesting than automation in a tight and tightening labor market.
The fear of wages and workers, not just idle "gee more profits would be nice", is what makes automation actually go, vs floundering out as some middle managers doing a b2b steak dinner grift
<<Not because I'm worried that my job will be taken by software, but because said software will barely work. For factory jobs, the risks are obvious; that's why we put these robots in cages[0]. However, these office jobs are still making critical decisions that will increasingly be handled by automation.
Just to build on this a little. Even if they do work, general population will have little to no understanding on how they work. They will be little black boxes that govern our daily lives with little to no way to correct it if things go awry. As much as I am amazed by what ML can do already, we need some basic customer facing documentation on how it is supposed to work.
Not just general population, but actual employees!
We've already seen this with legacy code. Whereas a manual process featured someone you could have a conversation with, how many 10+ year old apps are there that have some quirk that nobody remembers or understands?
Automation is going to make that worse, because it goes from "that process 3 people know about" to "that process no one has thought about in 5 years."
This is one of the dystopian incoming realities. Automation will be extremely pervasive in daily life and especially in urban environments. But because it will be dumb and there will be no profit in making it smarter or to improve respectful interaction with humans, we will be the ones corralled to allow the automation have free reign. Whatever noble intentions the small minority working on this might have, it will mainly be another gut punch to human dignity in the name of capitalism.
I know this is a bit of a low effort argument, but every ML "enhanced" product I use ay my job is a variation on "you entered data xyz - here are some more examples where people entered .xyz. and this is the result they got:..." or "you are entering data at time ab:cd - here is what other people searched for at similar times:...".
“For example, the way business mail used to work was that the bureaucrat in question would record their message onto a tape, and then send that tape off to a special department full of typists to actually turn that voice recording into a letter. That whole concept is not only gone, but it's such a foreign idea that it sounds like something you'd write for a dieselpunk novel. The moment we started putting computers on people's desks, we expected everyone to know how to type.”
It’s true that executives now do their own typing, but that is not automation. It’s actually a rare case of modern work becoming less specialized, with a whole category of highly specialized workers (typists) ceasing to exist.
If the executive uses voice-to-text technology, then that would be a case of automation.
If you try using a vintage typewriter you will reconsider. It was a ton more work than now.
(1) Typewriter keyboards were actually physically strenuous... find an old manual typewriter and type a couple pages on it and see how whether you feel like typing a dozen more. This was rectified beginning in the 60s I think, but manual models were still around for a while due to cost.
(2) Even once electric typewriters were invented, dealing with minor typos, let alone more major textual surgery, remained a huge hassle until the personal computer / full word processor came about allowing on screen editing... Just imagine typing most of a page only to realize you forgot a sentence near the middle, or even made just a small typo. which is around when typists stopped being a thing, because then, and only then, typing had become so automated that a reasonably skilled bureaucrat wasn't really saving much time (and losing latency) by using the typing pool, especially for brief, urgent memoranda.
It is a case of automation working so well, you don't even notice it at all and thus think nothing was automated, until you think about it in more detail.
The automation with a word processor is not the input (the typing), it's the page layout and reflow, and being able to edit and get WYSIWYG before committing to the printed page.
My best typing experience, typewriters manual or electric, or computer keyboards, was the IBM Selectric. Second best was on an AT&T 3270 (!?!) with a mechanical clicking keyboard, similar to original IBM PC, but even better. I mean using a 3270 isn't great fun but the typing was great.
To extend your excellent summary of office automation, you can think of most government functions as a manually operated AI. There are piles of rules and regulations to administer, and that is ripe for automation. However, can you imagine the horror of, say, a machine efficient IRS? The only thing that makes a lot of the regulatory regime survivable is the inefficiency of the bureaucracy. A hyper efficient bureaucracy would be suffocating.
I think you're right about automation coming for office jobs, but it's not that odd that physical labor would be the focus, a couple hundred years ago, like 95% of people worked in agriculture, and automation is the reason only 1% of people do today. It's harder to suss out how many factory jobs have been automated away compared to how many are just being done elsewhere, but a lot of factory labor has been automated as well.
It's almost like automation is coming for office jobs because it already claimed the low-hanging fruit elsewhere.
Human computers were automated by computers and created the programming jobs. But we have many more of those than ever and it seems to increase even more.
There are types of jobs that don’t necessarily get reduced even though productivity is increased.
For example the number of military soldiers is mostly a political decision. Managers, lawyers and other bureaucrats are increasing in numbers and power despite having better educated workers and more sophisticated tools.
The economic and political games we play are not aligning with reducing work time to pre industrial levels. Technological progress cannot be thought of as independent of societal factors.
Agree with your broad point, but minor nitpick about soldiers: we've substantially cut armed forces personnel, because they're hellaciously expensive with benefits.
While there are no human computers who actually add and multiply, there is a lot of programmers. But this is not because they replaced human computers 1:1, it's because electronic computers have replaced huge swaths of everything around us. I go to my kitchen, of all places, and I see my microwave oven, my coffee machine, my bread-making machine, even likely my gas stove all run by small embedded computers. To say nothing of the advanced computer attached to a bunch of radio interfaces most everyone carries around these days.
I think your last assumption is correct. Agriculture, textiles, automobiles, and food production (not restaurants) have all seen significant automation, while office jobs have actually increased.
Very good point, the physical labour we still see has survived 100+ years of attempted automation so is very resilient to it (relatively speaking). Whereas there is still a lot of office works that you can look at and quite easily think "there must be a way for a machine to handle that".
Even before you need language models, though, there's an insane amount of "digital manual labor" that involved people shuttling files around and validating / cross referencing data in ways that would be done far more correctly and efficiently by software. In my opinion, low code tooling threatens many more jobs than AI does in the short term.
This seems like one of those problems where the last 5% is going to take 99% of the effort and time.
A lot of people would be out of jobs if data was truly interchangeable, and there were robust ways of formatting data that would work for everything people wanted to do, and software was bug-free and exported/imported perfectly every time, but... even something "standard" like date/time data is all sorts of hard to have people and their systems actually do correctly in a general way.
There's definitely a lot of spaces where this is true but a huge fraction of spaces are not even that complicated.
I volunteer my time for various pandemic assistance and basic income work and the number of finance companies (banks, payment processors, etc) that rely on a human being regularly downloading a CSV from one system and uploading it unchanged to another is _absolutely_asinine_.
That does sound near-trivial to automate, but then I've got to wonder how frequently "regular" is? Saving someone 15 minutes a day is gonna be hard to justify spending programmer time on - or paying for a 3rd party solution? Are there people moving those spreadsheets for like 20+hrs a week?
When I temped at Bank of America, there were 20s of people who effectively just shuttled CSVs between points A and B, and a whole on premise army of programmers who made our work redundant. So...
I have a feeling a lot of people commenting never worked outside IT. I worked in marketing twice and in sales once just to see if I would like other career than programming. The amount of stuff I automated was amazing.
The problem is, it's hard to automate stuff for people in a scalable way. There is recent YC startup Axios, but when I tried to automate something with their software, the learning curve was too big.
I think the best way to automate the office work would be "automation houses" where people sell software that automates stuff. I think we have this for big companies, but smaller ones are left doing boring jobs
There would be no need for 75% of automation if all products had (1) a programmatic interface that covered 100% of functionality & (2) a data standard.
Unfortunately, that's something that almost no customer is mature enough to ask for, so down it goes on the priority list, if it makes it at all.
At least web apps are slightly better, as some sort of data packet is actually transmitted (even if user-action initiated), so easier to draw an interface line.
This is how competition is snuffed out - the companies that realize this have already completed the switch or have started it and they'll be the ones to outlast all their competitors that have fallen behind with slower processes.
This has been true for as long as people have written software. We eliminate ops with clouds, assembly with compilers, code with libraries, software with SaaS and platforms.
What makes low-code/no-code different from everything else?
I consider it yet another tool in the abstraction stack that allows folks without software backgrounds to use their expertise to fill their operational needs.
Disclaimer: my intention with low/no-code tooling is likely broader and includes everything from WYSWIG form builders to Airtable automations, etc. The key to me is the usage of (1) syntax-free (2) domain-specific tooling to (3) configure software to meet (4) custom business needs (what a mouthful).
I don't buy it. Low code can make ICs much more productive, but you don't fire the other three... They have a ton of institutional and domain knowledge. You just find things for them to do that wouldn't have been worthwhile without the leverage that came from low code.
While i'm not in favor of firing folks (especially if they do have institutional, domain knowledge)...i firmly believe that orgs will endeavor to force or influence or sneakily trick folks who have domain knowledge into codifying such knowledge which gets input into digital processes which will - also think - get codified by low code tools. As a technologist i like the automation part, but as sympathetic human, i hate that humans will be impacted. By impacted of course i mean that crappy orgs won;t try to find more relevant roles for these impacted humans, they'll resort to crappy things likke layoffs. (Separately, i do believe very much in universal basic income...and in fact having something like UBI in place keeps both my tech brain and human brain happy in a sort of win-win scenario where we get more aujtomation, but humans are taken care of.)
What you'll find in a lot of workplaces you can't automate judgement of something like a book, because of the management's ever shifting definition of what they want.
Twilight is a big hit, and suddenly they want a bunch of teen paranormal romance, and it's even fine if it's poorly written by normal standards if it has the right themes. Your AI from before is largely useless.
I disagree on the order of things. Both are equally on the chopping block.
Physical labor has reasons to be focused first compared to office jobs because as a society we've scaled that up far more (more jobs that involve physical labor than office jobs/those that don't), all that physical labor and scale comes at a large cost, and physical labor is at times easier to automate compared to office work. You can decompose the steps of delivering food to a table or lifting a box and dropping it down somewhere else. Naturally, we've converged to optimize for simplicity when it comes to physical labor because people in physical labor don't like wasted effort or operations that change all the time, whereas office jobs can have many conditional branches unpruned.
There are a few publishing companies in NYC that pitch manuscript ideas and ghost writers are hired to write it. Currently a profitable company, only works around cookbooks and logical book ideas that AI could pick up (what's trending on google, etc), but either way it's coming.
Majority of Excel jobs I have seen in regulated industries are because people use the tools that they are familiar with for everything. Do you really need a spreadsheet to generate reports? Project Management Tracking? Timesheets? Task Lists? Calendars and Schedules?
The majority of the jobs people do are updating crap in spreadsheets that don’t need to be there. If it weren’t for SOPs and audits, this stuff could be wiped out easily.
The only protection for these office workers is the status quo and inertia to change. Excel provides some task automation but still requires recurring human input and maintenance. The new wave of AI does NOT.
Sure in 2022...but when spreadsheet software first arose....many people lost their jobs and many many more people were hired in jobs that never existed before.
I dunno, the job of customer service phone rep seems like it should be easy to automate, and they've been attempting to do so for decades, but how many people get absolutely frustrated with such systems?
I interned in sales. One of my jobs was logging into CRM, counting how many people have tasks to do today / tomorrow / overdue.
Count how many deals moved since last day. Count deals value in the basket for specific stages.
It would take me 15 minutes a day, before I spent 15 minutes automating it. I know this is extreme example, but one that happened in my job. That was a startup with some of the best VCs in Europe.
Another international unicorn, had me create a marketing campaign and then copy the data from the marketing software into an excel. Just copy it like date, name of campaign, number of target users.
I guess I spent like 30 minutes a day doing this. Just copying stuff from one web app to a spreadsheet.
Automating customer service is hard, but work is so manual people in the programming field I think have no idea. Also made me unable to work in marketing or sales as I would just want to spend time automating stuff.
Summary maybe, but a judgement I doubt. Your robot would have to be able to address how the book fits into the current zeitgeist which is not something written down. It's ephemeral and intrinsic to the climate of current affairs, pop culture and most challenging, how people actually feel about the combination of all those facts.
Humans can fail to grasp all that too, but I think you vastly underestimate the complexity of culture and it's tendency to change rapidly.
Agree. Robotics get the bad rap because it can physically be pointed at vs. an AI in the cloud. Robots mostly take redundant physical labor whereas software is taking the cush jobs away.
This story has stuck with me since I read it many years back. And then seeing it come to life in so many ways where machines are "optimizing" processes and workflows for humans is scary.
Maybe publishing houses are not comfortable replacing her with an AI but it will eventually happen.