AI Will Not Replace Most Jobs, Despite What People Claim

Since it’s a common theme, I’m sometimes asked for my personal opinion, and this is no exception.

【Affirmative counterargument ①】 Most jobs will be automated. That’s only natural, since there are many things AI does more accurately than humans. But the same thing has happened at various points throughout history — the Industrial Revolution, the period of high economic growth, the IT revolution. So there’s no need to fear it needlessly; we should simply understand that a gradual, natural shift in “the range of work humans should do” will occur.

【Affirmative counterargument ②】 AI overwhelmingly surpasses humans. Soon there will be nothing left for humans to do. Even the work of developing higher-precision AI will come to be handled by other AI, and humans won’t even need to develop AI anymore. Here, humans will be called upon to find a new sprout of growth. What AI is good at is rational action. So humans should turn to the irrational actions unique to humanity — selfless service, non-profit activity, or art. Freed from rationality, humans need to move toward a new stage.

【Critical counterargument ①】 In reality, what current AI can actually do is severely limited, and it’s a serious mistake to treat it as though it were omnipotent. This is not the first AI development boom in history — far from it. Of the “strong AI” and “weak AI” discussed in the past, only weak AI, and only in specific fields, has produced remarkable results, and people are simply making a fuss because it seems likely to generate large economic benefits. The day computers essentially acquire intelligence is still a long way off — in fact, there’s no sign that such an age is coming at all.

【Critical counterargument ②】 AI’s progress is remarkable, but it carries many problems. Because AI is, by nature, an extension of statistical processing, it’s strong with probability. In other words, for the task of identifying “this is probably roughly right,” the balance of speed, accuracy, and cost is excellent. Conversely, AI’s weak points are individual optimization and exception handling. But the real world is overflowing with exactly that kind of thing. I can already hear the counterargument coming: “AI’s strength is finding patterns in massive datasets where patterns aren’t obvious at first glance.” But it’s obvious to anyone that running a giant supercomputer for days just to handle a handful of extremely rare cases in the real world lacks any economic rationality whatsoever. So no person or company will ever do such a thing.

All of these counterarguments seem, to some degree, to hit the mark.

As partly explained above, I think many engineers are well aware of AI’s technical constraints.

There’s a famous prediction about processor speed called “Moore’s Law.” It began in the 1960s when Gordon Moore, then an engineer at Fairchild and later the founder of Intel, stated in a paper that processor performance would double every ◯ years (the number ◯ was revised over time and settled around 18 months). But by the 1990s, many engineers had recognized its limits, and after 2000 came the era of multi-core processors that continues to this day. When I visited Israel once, I heard an episode about the engineers at Intel’s Israel branch who devised the multi-core processor. Apparently headquarters in Santa Clara wouldn’t listen to their proposal at all and instead pushed a different development theme on them. So they worked on that assigned theme while, in nine months, also building a multi-core processor prototype on the side, and then brought it with them to Santa Clara.

It’s a fairly exciting story, but the lesson here has two sides: most people don’t know about technical limits and don’t even try to notice them, while a handful of sharp engineers noticed and broke through the difficulty by exercising innovation.

I believe many technical challenges get resolved through this kind of innovation process. For example, recent “AI-generated” work is genuinely impressive. It used to always have a telltale artificial feel to it, but in recent years that has improved enormously. People often credit a technology called GAN (Generative Adversarial Networks) for much of this, but back in 2012, during that shocking moment when Google used deep learning to get a Stanford supercomputer to teach itself to recognize “cats,” surely no one had even imagined something like GAN.

If we take an optimistic stance on technical constraints, does that mean an age of AI triumphant is coming? I don’t see it that way. It relates to industrial structure.

There’s no field harder to topple a giant in than AI development. There’s nothing wrong with developing a new AI. But is there really anyone today motivated to develop image recognition AI from scratch and beat the giants? The reason is data. Data volume is critical to improving learning accuracy. The GAN example mentioned above is a typical case of this. AI technology suppliers who have already launched image-related services own an overwhelming volume of image data, and that volume keeps growing constantly. Winning as a latecomer is not a simple matter.

I once advised an AI-based medical startup to increase the processing accuracy and speed of their equipment. They were building a “system” that could automatically classify cells. Being a system meant they were developing the whole chain from IN to OUT. IN is the stage of acquiring and organizing information — the camera that captures the images, the focusing technology, the mechanical technology for precisely positioning the cells. It’s an area full of troublesome challenges. Then, based on the input data, AI automatically judges various pieces of information about the cells, and OUT is completed through final classification. In other words, my advice meant: if you intend to win on AI alone, you’ll end up crushed by the giants in the end, so you have better odds specializing in the part you’re good at.

A society without game changers is boring. And yet society is steadily heading toward a state of gridlock, including through the entrenchment of capital. The more people demand of AI, the more AI itself needs to acquire scarcity. What this means is: no one is going to pay hundreds of millions of yen at this point for an AI that recognizes the species of an animal from a photo. Not just AI — technology in general becomes commoditized and eventually obsolete. That holds true whether it’s a hardware business, a software business, or a content business.

But self-generated scarcity, in most cases, sooner or later hits a wall.

Take television, for example. How interested are you in an 8K TV? Do you even remember that 3D TVs existed? Television is already an area where securing scarcity has become difficult. This isn’t especially negative — it’s not the industry collapsing, nor is it anyone’s negligence — it just seems like a perfectly natural course of events.

Take the mail client you use on your computer, for example. Did you know there was a time when people paid money to buy mail clients? Can you say what’s good or bad about any particular mail client? I’d say the very concept of a “mail client” has already faded, and this all happened within the last 20 to 30 years. Did you know that Famicom game cartridges (not including the console) used to sell for more than ¥10,000? Now you can find toys at Don Quijote with 128 titles packed in for around $5.

Scarcity is extremely important in a capitalist society. This is not the same as monopoly. Creating scarcity through monopoly is not permitted in modern society (that’s what antitrust law is for). Scarcity is only born when something makes users think, “This has special value! I want it!” — and someone creates that ahead of everyone else. There’s a very difficult balance here: emphasize “being ahead” too much and the pursuit of value falls short; emphasize value or performance too much and you lose to others, especially to the giants. But when there’s tremendous open possibility on the “value” side — as in the current field of AI development, where expectation and belief in growth are strongly pervasive across society as a whole — there’s no doubt that opportunities are scattered everywhere. But this, too, will eventually run its course.

AI development requires substantial cost, and so does actual operation. The processing power demanded of edge devices keeps rising, which inevitably becomes a factor squeezing costs. Naturally, many innovations will resolve these issues. But I believe an imbalance is bound to emerge somewhere.

At that point, I think we’ll start hearing people say, “Isn’t this better than AI?”

There’s a passage in Yanagi Muneyoshi’s “The Way of Crafts” that goes something like this: “A culture that ignores machines cannot stand, and I do not deny that machine-made products possess a peculiar beauty of their own. At the same time, a culture in which humans become slaves to machines cannot stand either. Society, therefore, should not demand machines without limit.” Unlimited demand is exactly what excessive mechanization is. The concern that the sage Yanagi Muneyoshi pointed out 100 years ago is precisely what stands before us right now.

Humanity’s capacity for misunderstanding and lack of imagination can sometimes be extraordinary. A typical example is the “human washing machine” from the Osaka World Expo. It was supposedly a system where you get into a machine, press a switch, and your bath is done — but who on earth wants such a product? What’s even more astonishing is that this washing machine wasn’t just an exhibit; a major manufacturer actually pursued its development for a while. If Muneyoshi were alive, I imagine he would declare that nearly every machine solemnly presented under the theme of “predicting the future” was extremely ugly and unnecessary.

I’m a father of four, and recently my six-year-old eldest daughter has become able to wash her own hair, which fills me with a quiet sense of her growth. That said, her technique still isn’t great, and she asks me, “Papa, how was it? Did I do a good job?” When I point out spots she missed on the back of her neck or around her ears, we go through an exchange like, “Hmm… how about now?” “Yeah, you did a great job.” I’m happy being in the midst of this wasteful, inefficient process. I truly, from the bottom of my heart, am glad the human washing machine never spread to ordinary households.

The content of this post is an excerpt (original text) from the following book. If you’re interested, please consider picking up a copy.

The Shape of a Happy IoT Startup

The Shape of a Happy IoT Startup

Originally published in Japanese at https://clazytech.com/2022/08/1078/. Translated with LLM assistance and reviewed before publication.