Clay Tech

"clay-works make things real"

translated from clazytech.com

Moore's Law Shows Why AI Cannot Read Human Emotions

original Society & Ideas

The other day my wife read an article claiming “AI that reads people’s emotions and gives recommendations is going to develop further, and so on” and called it “downright frightening.” She used to be a nurse. Medicine is not omnipotent. Some patients can be cured and some cannot. Some people are certain to recover but can’t stop worrying right now, and some people’s condition deteriorates before their test results even come back. In situations like that, a nurse has to seriously think through “what does this person want,” “what are they thinking right now,” “what kind of support can I offer,” and then turn that into action. It’s an extremely difficult profession. What does he or she like, and what do they need right now.

I once talked with a young man, a first-year doctoral student at Tohoku University. He researches cancer treatment using ultrasound, and he told me, “I don’t think curing disease is the most important thing a doctor should do.” Cancer treatment doesn’t end once the tumor is removed. There’s the care that comes afterward, which involves not just the patient but the family too, and if it’s already too late, there’s palliative care to consider. It runs very deep. Analyzing needs and making recommendations. At first glance, that looks like territory where AI would excel.

There’s a debate about whether AI could move into this kind of domain. It will probably heat up even more from here. But I believe AI cannot enter this domain. Here’s my reasoning. This isn’t an emotional argument like “I’m against AI!” or “We don’t need strong AI to replace humans!” It’s entirely a logical matter. (For what it’s worth, I use AI constantly in my own work, in other contexts.)

Meanwhile, there’s something that seems distant but is actually closely related: “Moore’s Law,” which is in the title of this post. If you don’t know it, look it up on Wikipedia. In short, it’s the prediction that CPU processing performance doubles every 18 months, an accelerating progression.

I believe it was around 2005 when a very famous person polled opinions on Twitter, asking, “Do you affirm or deny Moore’s Law?” I denied it. The reason is simple: silicon has physical limits, so process miniaturization cannot continue indefinitely. Indeed, the industry moved toward multicore processing. In other words, it gave up on linearly increasing processing speed within a single core, and instead pursued raising overall processing speed by making parallel processing more efficient. Put simply, Moore’s Law was indeed wrong (which makes sense — I doubt even Moore himself expected his statement to be quoted for this long).

As an aside, I once heard from someone in Israel who had been on the team that first developed multicore processors at Intel. Apparently, at the time, many people still believed in Moore’s Law, and headquarters in Santa Clara had rejected the multicore idea itself. The fact that the team went ahead and built it secretly in Israel anyway struck me as the mark of people who were both strong-willed and wonderfully untamed (meant as a compliment).

The idea that technology will keep advancing infinitely is an illusion.

Behind the scenes, engineers solve technical limitations through shifts in thinking, or they swap out entirely different solutions underneath something that looks the same on the surface, crossing the chasm that way. Piled up behind that are the struggles, challenges, and disappointments of countless engineers.

The motivation to go that far only arises where economic rationality exists. In other words, it has to be something that can become a business or help someone, something that can be monetized from there. This is the basis of my own view that “emotion-reading recommendation AI will not develop.”

First, there are several technical hurdles here. One is the generalization of vital sensing. For example, if you think about ways to read emotions in real time, you might come up with things like:

That’s roughly the list. Needless to say, none of this is data that’s just lying around on the internet. It has to be captured in the field. That means someone has to develop a device. Most recommendation AI up to now has been based on behavioral data from the internet (an enormous amount of it), and compared to that, vital data captured in the field is, frankly, extremely thin as a database. It would be a huge mistake to think you can do here what was possible in a world that lived entirely online.

Now, suppose you somehow manage to develop the device. You then have to get users to wear it, or install it somewhere in the field. In other words, you have to persuade someone with something like, “Please wear this (or let us place this here) because it offers such-and-such benefit.” On the internet, when collecting information, there are prompts asking for permission to use cookies, and terms of service, both of which most users blow past in a second, and yet they’re packed with rules. In other words, our behavioral data has already been harvested everywhere, over and over. (This has been true for over 20 years.)

Now, suppose you somehow manage to persuade someone. Then data collection begins. Without data, AI accuracy won’t improve. And having worked with biological data many times myself, I have a personal sense of this: Biological data varies enormously between individuals, and it’s genuinely brutal to deal with. That’s really all there is to it. The time it takes to get stable results, the number of subjects, the ingenuity required in the algorithms — in other words, the investment required. Unless you’re prepared for that, it’s very hard to build an AI that handles biological data. This isn’t a matter of the AI technology itself, nor of supercomputer processing speed. It’s an entirely physical, mechanical, solemn saga demanding enormous labor and trial and error.

But that’s fine, isn’t it. Humans can push through for the sake of a goal. The desire to grow is one of the levers of human self-actualization. Many people devote their fleeting lives to clearing the challenge in front of them and standing atop that summit. Many more will keep appearing in the future.

So let’s reconfirm what that summit actually is. “AI that reads emotions and makes recommendations.” Wait — AI alone has no value by itself, so we need to think of it tied to some kind of service. So then: “A streaming service that plays music matched to your mood that day.” “A robot that comforts you when you’re feeling down.” ”

Hmm. Do people really want that so badly? At the very least, I can’t imagine there’s anything here substantial enough to be a game changer.

In other words, emotion AI is just one more instance of the “illusion that technology will somehow keep advancing.” There will surely be people who research it as a hobby. But the amount invested there is probably less than a hundredth of what a “large corporation that got serious about it” would invest. I think the key that brought AI this much attention was the major progress made in “image recognition technology.” Image recognition is extremely useful. It’s now used practically everywhere in real-world work. At this point there’s no need to explain further. That happened because, starting with Google, enormous numbers of “X that got serious about it” emerged — not just companies but universities, research institutions, startups that sensed a business opportunity, and individuals with strong personal interest. Emotion recognition hasn’t shown anywhere near that level of opportunity. At best it’s “might be interesting to have.” And that situation hasn’t changed for a very long time. From the lack of economic rationality and growth potential, the prediction follows mercilessly: that technical domain probably won’t develop at all.

Originally published in Japanese at https://clazytech.com/2020/01/419/. Translated with LLM assistance and reviewed before publication.