Fine, AI is Going to Take All The Jobs. But When?

Let’s assume Elon Musk, Peter Thiel, and the many other technologists telling us AI will replace all human jobs are right. Okay, great — but when? And which jobs will come first, how fast will they take over, and what can we do to prepare — not as a society, but as individuals who need food, housing, and shelter, and who presently depend on jobs to satisfy those needs?

I’m surprised this question isn’t answered more often, because it’s actually quite easy to figure out. In particular, it’s really easy to figure out what tasks, and therefore what jobs, are going to disappear first.

(If you want the answer go ahead and skip to the bold section header below.)

Keep in mind I am talking about a very specific timeframe: until generative AI is perfect, we need to determine the dispersal and distribution of generative AI’s job displacement: simply put, who’s going to lose their jobs and when, before generative AI replaces all of our jobs?

I don’t want to talk about whether generative AI will replace all our jobs. Elon Musk et al. see this as a utopian future because their value on human creativity and input is low. Maybe it’s other humans’ creativity and input in the case of Musk, but no matter; the point is that, until generative AI is perfect, we need to try to find out who is going to lose their jobs, when, and what will replace them.

Some jobs are already dwindling, which gives us our starting point. ChatGPT’s primary function is to generate (the “G”) a chat (guess where that is in the name) through a process (or “transformation”, the T) of applying its training (“pre-training”) to two texts: the text in its corpus and the text that the user is inputting. Think of GPT as a process of going through (almost!) every possible combination of bits of language that could make sense and finding the one that makes the most sense in that context.

There are, of course, many competing chatbots: Bard, Azure, Character.AI, and others; the technical differences between them aren’t terribly important for my current point, so I’ll just refer to “Generative AI”, but at the end of this article I think you will understand why distinctions between their underlying technologies will become increasingly irrelevant over time, and that there are much more important factors in the unfolding of generative AI that require attention (and can yield tremendous money-making opportunities, good lord). And now you’ve got a damn cliffhanger forcing you to read more of this drivel — rhetoric sucks, doesn’t it?

Anyway. Optimizing for sensible text means chat bots have already begun replacing the first wave of job-dependent victims: programmers. I work at a financial firm, and we process data to produce reports for clients. We have embarked on many new projects that require programming skills we lack, and we use products like ChatGPT to generate the code we ultimately need to use for our own internal analyses. Our overall demand for programmers is down.

Note I didn’t say that our demand for programmers is zero. Rather, if you put together every company’s demand for a programmer on Earth and make that a number, that number has gone down because of Generative AI.

This isn’t the same as saying Generative AI is replacing jobs. At the moment, what it means is that, as Jyoti Bansal wrote in Entrepreneur, “Rather than stealing jobs, AI is making in-demand and over-taxed developers significantly more productive, leading to big gains for innovation.”

That’s great! For now. What it means is that overall demand for programmers was, before AI chatbots, above the overall supply of programmers. This is no secret. One billionaire once called them the most scare resource on the planet.

This is why Bansal is right — AI isn’t replacing programmer jobs. Yet. Because the amount of demand exceeded supply, and the growth in supply from Generative AI has helped those two levels meet.

In the future, supply is going to grow from both more humans graduating from programming programs globally and from growth in the use/capability of Generative AI. It is likely that the latter will grow faster than the former, given the rate of current improvements to AI chatbots over the last year. But even if it stalls, let’s face it — it takes humans not only four years or so to get a university degree, but many more years for mastery. Whether you believe in the 10,000 hour rule or not, it is undeniable that humans take a lot of time to master ideas and skills. Meanwhile, Generative AI can fill in gaps, and over time it’ll fit more and more labor gaps, including ones currently filled by people.

The first and easiest step is for the programmers to use Generative AI to generate much of the difficult and unenjoyable code needed to complete their tasks; the next step is to concentrate assignments to fewer programmers as fewer overall are needed.

This is the first chapter of the broader story of the birth of AI, and we are maybe past the first couple of paragraphs so far. As it scales in skill and speed, either through software shortcuts or hardware improvements, the share of the work programmers do not want to do that Generative AI is used to replace will approach 100%.

Note I say the “difficult and unenjoyable” code; a programmer would never want to automate all of his coding process anymore than a writer would want to automate all of his writing process.

And here, we have to take a step back. Would a writer want to automate any of his writing process?

I earn a living as a writer, so I can answer that definitively: yes. Imagine you are a first-time novelist; you do not simply send your novel to an agent, get accepted, and then get published. There are a lot of documents required to get a book through the book publishing process — even if you are self-publishing! Marketing is necessary (that’s why this brilliant article will get me 5 readers instead of the 5 million my brilliance deserves; not that it matters, nothing here is actionable and I’m not earning anything from Medium, I just need to get these thoughts down on paper ASAP so I can prove people I was right several years from now).

A few years ago when I tried to get a book published, I quickly gave up as agents and publishers demanded very long book proposals with, in my opinion, tons of necessary bullshit. If I wanted to write a book proposal today, ChatGPT would write it for me. And I’d look it over. Maybe.

What is emerging is a picture of Generative AI as a tool to replace what I’d like to call “bad work”. This is defined differently by different people even within the same job and responsibility. As a financial analyst, I love writing reports that clarify the data and shape it into a clear narrative with a clear actionable conclusion. One of my colleagues, on the other hand, hates interacting with clients and loves spending hours in Python and R. Different people like different things.

And we both can use ChatGPT to replace the bad work in our jobs, meaning the overall demand for bad work is falling when the output of that bad work is text. Of course, how much text we can produce and how good it is is limited by technology; as GPT technology improves, ChatGPT will increasingly be used to replace all of the “bad work” in text production over time — but also some of the “good work” as well, since obviously some people in the market will be happy to use AI without an expertise filter in the way, in much the same way that some scammers use text generation tools today to produce spam on the internet.

Next is sound.

We know this simply by computational complexity; text involves the combination of 37 alphanumeric characters in English (A-Z, 0–9, and a space) excluding special characters, which computationally are trivial in their complexity relative to the alphanumerics. But putting all of these together is much harder than manipulating sound waves. If we limit sound to 44.1kHz (that’s CD-quality), that means a computer has 44,100 data points to analyze per second; double that for Stereo. But that isn’t even as bad of a problem as dimensions: soundwaves operate in three-dimensional space, text is merely one-dimensional.

After that we’ll get to images, video, and gaming. Indulge me for a bit first.

AI can already use its pre-trained techniques to generate convincing sound at a reasonable rate of computing power, which has made for some incredibly hilarious (and shocking) deepfakes on the Internet. The industry has tried to pivot Generative AI towards music production, which I think may slow the development of Generative AI’s sound applications and the decline in jobs here, so this is likely to be a much slower transition than what we will see with text because of social constraints — the world is rightly terrified of people creating deepfakes of world leaders that start the apocalypse.

As for the bad jobs in sound? Well, imagine you’re Radiohead and you’re working on a new track, but Jonny Greenwood really sucks and keeps playing the wrong notes! Are you going to do Take #357 or just go with the one that’s kind of close and tell your AI to change all the B-flats to B’s?

Now, I’m making fun of Radiohead to trigger all you aging hipsters out there — they’re one of my favorite bands and Jonny Greenwood is an obvious genius. My point is that there are a lot of little things in sound production — moving a beat 30 milliseconds to see if it sounds better there or not — that is the “bad work” of the “good work” of being a rock star. And “rock star” is such good work that it became lazy marketing-speak from HR drones trying to convince people to trade 80 hours a week for a lower middle class income in an expensive urban area. I’m financially independent, assholes, so don’t think you’ll get revenge by never hiring me again.

I digress again. Sorry, free text is free for a reason.

Anyway, this thought exercise demonstrates two things: even in the best “good work” there are still elements of “bad work,” and AI will lower demand for that bad work over time in audio as well. But at a much slower pace than text, with the noticeable impact still to come, and it may take several months or even years because of the cultural constraints on sound-generating AI we have in place.

Similar constraints won’t come for images as the technology improves; they already have. Issues of likeness and obscenity, both for legal as well as moral reasons, has caused ChatGPT et al. to set very confining guardrails here.

Oddly enough, AI might be the first technology that is not applied to pornography before other kinds of content, breaking a trend that began with paint and carried on through the camera, the film camera, the VHS tape, and the internet. This actually does not surprise me; those technologies were all mediums to represent human creative output, not to recreate, as AI does. Thus we want to have more human restraints on very sensitive topics like pornography and less AI input.

Let’s pause for a minute and ponder what we mean by “sensitive”. Pornography exists specifically to stimulate, and it can stimulate both sexual desire and many other emotions: disgust, rage, indignation, and so on. That’s why it’s so close to art and also why it’s pretty easy to tell what pornography is when you see it in the extreme cases (your BangBuses, your FakeTaxis, etc.), but pretty hard to tell in its most banal cases (Victorian ankle porn, anyone?).

Art exists to do all of these things, except sexual desire. Maybe. No one has a proper objective definition of art, and many will disagree on this point in particular, but it’s irrelevant to my point: art and porn both serve to create emotional and physical sensations. They depend on sensitivity. We need a new term for this — a “sensitivity-dependency” that one could quantify and map out. Many ad agencies do precisely this; it’s also what I trained to do for my PhD, and it’s also what I’ve seen some hedge funds try to do when incorporating the linguistic ideas behind ChatGPT into their trading strategies.

If you haven’t heard of this, it’s called “sentiment analysis”, and what we mean by analyzing sentiment is identifying the parameters of sensitivity: who is sensitive, when, by how much, what caused it, how do they respond. After all, sentiment and sensitivity mean the same thing: they come from Latin “sentire”, “to feel”.

Sentiment analysis is a big deal in AI (AWS markets the hell out of it, for example), because it is about perfecting these constraints so that generative AI will produce the right stuff and not the wrong stuff in an artistic or pornographic context. It does this by creating a limit function on how sensitive its output is: not too offensive, and also not too salacious. Not too scary nor provocative.

A clear problem has emerged. What provokes me does not provoke you and vice versa, which is why you will have a large proliferation of AI services attuned to different sensitivity levels as it is used for higher and higher levels of sensation, and people will have to find what works for them.

This is all necessary work before Generative AI can really begin producing “good work” as it moves from images to video and the final frontier: gaming. Sure, Generative AI can produce “bad work” in video and gaming today: its biggest video accomplishment in terms of commercial success (here defined as viewership and engagement rather than monetization of viewership and engagement) was Nothing, Forever. This is very much “bad work”. Its popularity was not in its quality as a sitcom, its popularity rested on the absurdity that resulted from Generative AI’s imperfections in recreating Seinfeld.

To make the jump, it has to get better at making more lifelike video, and the jump in complexity from images to video is massive, making this a bottleneck that is likely to last for years, which will only set up for the even more computational difficulty of gaming — and that leads us into talk of holodecks and weird scifi stuff too far ahead for me to talk sensibly about.

My expertise, whatever I may have, can’t apply beyond the point when we start to replace movies with AI-produced videos. The time it takes for this process is likely to be extended significantly by the need to create different AIs that fit different sensitivities, and I do expect tremendous chaos during this sorting process. It is going to intersect questions of morality and absurdly large economic interests. When I can use AI to generate hardcore videos of Mickey rawdogging Donald, I’m going to offend people because of its tastelessness and immorality and Disney for using their intellectual property (I know the copyrights are expiring, please don’t nitpick irrelevant details just to feel clever — and if you don’t find the mental image I’ve put in your head tasteless and immoral, well I’ll just let you think about that for a bit).

Okay You Scrolled Past the Bullshit, Read This

In short, AI is going to replace a lot of jobs, and it’s going to take a very long time due to five primary factors:

1. As we replace “content”, for want of a better term, computational complexity grows. Text is one-dimensional, sound three-dimensional, and images multidimensional, and videos collections of many multidimensional images. This spreads the process out, probably for many years, on its own.

2. As we replace more and more computationally complex content, demand for “bad work” in the fields that have that content output will decline. First there’s decline in text — that is a lot of people, programmers, lawyers, writers.

3. Demand for “bad work” is not the same as demand for labor, so the impact of this decline in bad work demand, particularly in labor markets where demand had already outweighed supply (e.g. programming), will result in the decline in actual jobs in this sector to decline more slowly than the work, taken in aggregate, would at first suggest. This is reinforced by early-stage sustained demand for highly specialized human filters of AI output.

4. There will be a significant threshold that makes the trend of replacing AI with bad work slow down significantly when AI has mastered text, sound, and image and begins to make good video content with reasonably little computing power. When we have the tech to be able to tell an app “Make me 7 seasons of a Star Trek TV show that don’t suck”, society will make this solipsistic media utopia further out of reach as we try to figure out the legal, remunerative, and ethical issues of replacing video.

5. After a protracted period, we’ll finally start to build Holodecks and Starfleet’s enlisted will finally fulfill their destiny of following Commander Riker with a bucket and a mop.

There You Go — But Wait, How Long For Real?

How long will all of this take — well, we aren’t going to master 100% of text then 100% sound then 100% images sequentially: simultaneous improvements mean these will grow together, and human adoption of the technology will vary significantly as well. But it will take many years, and only after that will we finally be able to seriously talk about what we do with AI-generated good video.

That conversation is going to take decades. I think I will see the technology to make those 7 seasons of a Star Trek show in my lifetime, but I don’t think I will see those actual shows.

For workers, however, there won’t be one set rule for when they’ll get a pink slip because they’ve been replaced by Generative AI, and for many that won’t happen for many, many years. And they will have a place at the table to say when they lose their jobs and how because they’ll be the primary appliers of AI tools to bad work. That doesn’t mean they will get their way, but they will be able to advocate for themselves.

Which of course means we need a society where everyone is listened to in good faith, treated with compassion and humanity, and whose real concerns are addressed in a way that is fair for everyone.

And, as everyone knows, we don’t have that society — and after thousands of years of saying we want that kind of society, we still haven’t made it. Which, as someone who has read about stated and revealed preferences, tells me we don’t actually want that society and never will build it. Which means AI’s unfolding is going to make the Twitter culture war of the late 2010s look like a quaint game of table tennis between toddlers.

Unfortunately, AI is a beautiful and powerful tool that humanity is not really mature enough to use in a way that won’t be very, very painful for a very long time.