Neuromancer

How will AI affect inequality?

PNRJ current events, ethics, futurism, inequality, public policy , , , , , , , , , , , ,

Oct 15 JDN 2460233

Will AI make inequality worse, or better? Could it do a bit of both? Does it depend on how we use it?

This is of course an extremely big question. In some sense it is the big economic question of the 21st century. The difference between the neofeudalist cyberpunk dystopia of Neuromancer and the social democratic utopia of Star Trek just about hinges on whether AI becomes a force for higher or lower inequality.

Krugman seems quite optimistic: Based on forecasts by Goldman Sachs, AI seems poised to automate more high-paying white-collar jobs than low-paying blue-collar ones.

But, well, it should be obvious that Goldman Sachs is not an impartial observer here. They do have reasons to get their forecasts right—their customers are literally invested in those forecasts—but like anyone who immensely profits from the status quo, they also have a broader agenda of telling the world that everything is going great and there’s no need to worry or change anything.

And when I look a bit closer at their graphs, it seems pretty clear that they aren’t actually answering the right question. They estimate an “exposure to AI” coefficient (somehow; their methodology is not clearly explained and lots of it is proprietary), and if it’s between 10% and 49% they call it “complementary” while if it’s 50% or above they call it “replacement”.

But that is not how complements and substitutes work. It isn’t a question of “how much of the work can be done by machine” (whatever that means). It’s a question of whether you will still need the expert human.

It could be that the machine does 90% of the work, but you still need a human being there to tell it what to do, and that would be complementary. (Indeed, this basically is how finance works right now, and I see no reason to think it will change any time soon.) Conversely, it could be that the machine only does 20% of the work, but that was the 20% that required expert skill, and so a once comfortable high-paying job can now be replaced by low-paid temp workers. (This is more or less what’s happening at Amazon warehouses: They are basically managed by AI, but humans still do most of the actual labor, and get paid peanuts for it.)

For their category “computer and mathematical”, they call it “complementary”, and I agree: We are still going to need people who can code. We’re still going to need people who know how to multiply matrices. We’re still going to need people who understand search algorithms. Indeed, if the past is any indicator, we’re going to need more and more of those people, and they’re going to keep getting paid higher and higher salaries. Someone has to make the AI, after all.

Yet I’m not quite so sure about the “mathematical” part in many cases. We may not need people who can solve differential equations, actually: maybe a few to design the algorithms, but honestly even then, a software program with a simple finite-difference algorithm can often solve much more interesting problems than one with a full-fledged differential equation solver, because one of the dirty secrets of differential equations is that for some of the most important ones (like the Navier-Stokes Equations), we simply do not know how to solve them. Once you have enough computing power, you often can stop trying to be clever and just brute-force the damn thing.

Yet for “transportation and material movement”—that is, trucking—Goldman Sachs confidently forecasts mostly “no automation” with a bit of “complementary”. Yet this year—not at some distant point in the future, not in some sci-fi novel, this year in the actual world—the Governor of California already vetoed a bill that would have required automated trucks to have human drivers. The trucks aren’t on the roads yet—but if we already are making laws about them, they’re going to be, soon. (State legislatures are not known for their brilliant foresight or excessive long-term thinking.) And if the law doesn’t require them to have human drivers, they probably won’t; which means that hundreds of thousands of long-haul truckers will suddenly be out of work.

It’s also important to differentiate between different types of jobs that may fall under the same category or industry.

Neurosurgeons are not going anywhere, and improved robotics will only allow them to perform better, safer laparoscopic surgeries. Nor are nurses going anywhere, because some things just need an actual person physically there with the patient. But general practictioners, psychotherapists, and even radiologists are already seeing many of their tasks automated. So is “medicine” being automated or not? That depends what sort of medicine you mean. And yet it clearly means an increase in inequality, because it’s the middle-paying jobs (like GPs) that are going away, while the high-paying jobs (like neurosurgeons) and the low-paying jobs (like nurses) that remain.

Likewise, consider “legal services”, which is one of the few industries that Goldman Sachs thinks will be substantially replaced by AI. Are high-stakes trial lawyers like Sam Bernstein getting replaced? Clearly not. Nor would I expect most corporate lawyers to disappear. Human lawyers will still continue to perform at least a little bit better than AI law systems, and the rich will continue to use them, because a few million dollars for a few percentage points better odds of winning is absolutely worth it when billions of dollars are on the line. So which law services are going to get replaced by AI? First, routine legal questions, like how to renew your work visa or set up a living will—it’s already happening. Next, someone will probably decide that public defenders aren’t worth the cost and start automating the legal defenses of poor people who get accused of crimes. (And to be honest, it may not be much worse than how things currently are in the public defender system.) The advantage of such a change is that it will most likely bring court costs down—and that is desperately needed. But it may also tilt the courts even further in favor of the rich. It may also make it even harder to start a career as a lawyer, cutting off the bottom of the ladder.

Or consider “management”, which Goldman Sachs thinks will be “complementary”. Are CEOs going to get replaced by AI? No, because the CEOs are the ones making that decision. Certainly this is true for any closely-held firm: No CEO is going to fire himself. Theoretically, if shareholders and boards of directors pushed hard enough, they might be able to get a CEO of a publicly-traded corporation ousted in favor of an AI, and if the world were really made of neoclassical rational agents, that might actually happen. But in the real world, the rich have tremendous solidarity for each other (and only each other), and very few billionaires are going to take aim at other billionaires when it comes time to decide whose jobs should be replaced. Yet, there are a lot of levels of management below the CEO and board of directors, and many of those are already in the process of being replaced: Instead of relying on the expert judgment of a human manager, it’s increasingly common to develop “performance metrics”, feed them into an algorithm, and use that result to decide who gets raises and who gets fired. It all feels very “objective” and “impartial” and “scientific”—and usually ends up being both dehumanizing and ultimately not even effective at increasing profits. At some point, many corporations are going to realize that their middle managers aren’t actually making any important decisions anymore, and they’ll feed that into the algorithm, and it will tell them to fire the middle managers.

Thus, even though we think of “medicine”, “law”, and “management” as high-paying careers, the effect of AI is largely going to be to increase inequality within those industries. It isn’t the really high-paid doctors, managers, and lawyers who are going to get replaced.

I am therefore much less optimistic than Krugman about this. I do believe there are many ways that technology, including artificial intelligence, could be used to make life better for everyone, and even perhaps one day lead us into a glorious utopian future.

But I don’t see most of the people who have the authority to make important decisions for our society actually working towards such a future. They seem much more interested in maximizing their own profits or advancing narrow-minded ideologies. (Or, as most right-wing political parties do today: Advancing narrow-minded ideologies about maximizing the profits of rich people.) And if we simply continue on the track we’ve been on, our future is looking a lot more like Neuromancer than it is like Star Trek.

What is the processing power of the human brain?

PNRJ cognitive science, futurism , , , , , , , , , , , , , , , , ,

JDN 2457485

Futurists have been predicting that AI will “surpass humans” any day now for something like 50 years. Eventually they’ll be right, but it will be more or less purely by chance, since they’ve been making the same prediction longer than I’ve been alive. (Similarity, whenever someone projects the date at which immortality will be invented, it always seems to coincide with just slightly before the end of the author’s projected life expectancy.) Any technology that is “20 years away” will be so indefinitely.

There are a lot of reasons why this prediction keeps failing so miserably. One is an apparent failure to grasp the limitations of exponential growth. I actually think the most important is that a lot of AI fans don’t seem to understand how human cognition actually works—that it is primarily social cognition, where most of the processing has already been done and given to us as cached results, some of them derived centuries before we were born. We are smart enough to run a civilization with airplanes and the Internet not because any individual human is so much smarter than any other animal, but because all humans together are—and other animals haven’t quite figured out how to unite their cognition in the same way. We’re about 3 times smarter than any other animal as individuals—and several billion times smarter when we put our heads together.

A third reason is that even if you have sufficient computing power, that is surprisingly unimportant; what you really need are good heuristics to make use of your computing power efficiently. Any nontrivial problem is too complex to brute-force by any conceivable computer, so simply increasing computing power without improving your heuristics will get you nowhere. Conversely, if you have really good heuristics like the human brain does, you don’t even need all that much computing power. A chess grandmaster was once asked how many moves ahead he can see on the board, and he replied: “I only see one move ahead. The right one.” In cognitive science terms, people asked him how much computing power he was using, expecting him to say something far beyond normal human capacity, and he replied that he was using hardly any—it was all baked into the heuristics he had learned from years of training and practice.

Making an AI capable of human thought—a true artificial person—will require a level of computing power we can already reach (as long as we use huge supercomputers), but that is like having the right material. To really create the being we will need to embed the proper heuristics. We are trying to make David, and we have finally mined enough marble—now all we need is Michelangelo.

But another reason why so many futurists have failed in their projections is that they have wildly underestimated the computing power of the human brain. Reading 1980s cyberpunk is hilarious in hindsight; Neuromancer actually quite accurately projected the number of megabytes that would flow through the Internet at any given moment, but somehow thought that a few hundred megaflops would be enough to copy human consciousness. The processing power of the human brain is actually on the order of a few petaflops. So, you know, Gibson was only off by a factor of a few million.

We can now match petaflops—the world’s fastest supercomputer is actually about 30 petaflops. Of course, it cost half a month of China’s GDP to build, and requires 24 megawatts to run and cool, which is about the output of a mid-sized solar power station. The human brain consumes only about 400 kcal per day, which is about 20 watts—roughly the consumption of a typical CFL lightbulb. Even if you count the rest of the human body as necessary to run the human brain (which I guess is sort of true), we’re still clocking in at about 100 watts—so even though supercomputers can now process at the same speed, our brains are almost a million times as energy-efficient.

How do I know it’s a few petaflops?

Earlier this year a study was published showing that a conservative lower bound for the total capacity of human memory is about 4 bits per synapse, where previously some scientists thought that each synapse might carry only 1 bit (I’ve always suspected it was more like 10 myself).

So then we need to figure out how many synapses we have… which turns out to be really difficult actually. They are in a constant state of flux, growing, shrinking, and moving all the time; and when we die they fade away almost immediately (reason #3 I’m skeptical of cryonics). We know that we have about 100 billion neurons, and each one can have anywhere between 100 and 15,000 synapses with other neurons. The average seems to be something like 5,000 (but highly skewed in a power-law distribution), so that’s about 500 trillion synapses. If each one is carrying 4 bits to be as conservative as possible, that’s a total storage capacity of about 2 quadrillion bits, which is about 0.2 petabytes.

Of course, that’s assuming that our brains store information the same way as a computer—every bit flipped independently, each bit stored forever. Not even close. Human memory is constantly compressing and decompressing data, using a compression scheme that’s lossy enough that we not only forget things, we can systematically misremember and even be implanted with false memories. That may seem like a bad thing, and in a sense it is; but if the compression scheme is that lossy, it must be because it’s also that efficient—that our brains are compressing away the vast majority of the data to make room for more. Our best lossy compression algorithms for video are about 100:1; but the human brain is clearly much better than that. Our core data format for long-term memory appears to be narrative; more or less we store everything not as audio or video (that’s short-term memory, and quite literally so), but as stories.

How much compression can you get by storing things as narrative? Think about The Lord of the Rings. The extended edition of the films runs to 6 discs of movie (9 discs of other stuff), where a Blu-Ray disc can store about 50 GB. So that’s 300 GB. Compressed into narrative form, we have the books (which, if you’ve read them, are clearly not optimally compressed—no, we do not need five paragraphs about the trees, and I’m gonna say it, Tom Bombadil is totally superfluous and Peter Jackson was right to remove him), which run about 500,000 words altogether. If the average word is 10 letters (normally it’s less than that, but this is Tolkien we’re talking about), each word will take up about 10 bytes (because in ASCII or Unicode a letter is a byte). So altogether the total content of the entire trilogy, compressed into narrative, can be stored in about 5 million bytes, that is, 5 MB. So the compression from HD video to narrative takes us all the way from 300 GB to 5 MB, which is a factor of 60,000. Sixty thousand. I believe that this is the proper order of magnitude for the compression capability of the human brain.

Even more interesting is the fact that the human brain is almost certainly in some sense holographic storage; damage to a small part of your brain does not produce highly selective memory loss as if you had some bad sectors of your hard drive, but rather an overall degradation of your total memory processing as if you in some sense stored everything everywhere—that is, holographically. How exactly this is accomplished by the brain is still very much an open question; it’s probably not literally a hologram in the quantum sense, but it definitely seems to function like a hologram. (Although… if the human brain is a quantum computer that would explain an awful lot—it especially helps with the binding problem. The problem is explaining how a biological system at 37 C can possibly maintain the necessary quantum coherences.) The data storage capacity of holograms is substantially larger than what can be achieved by conventional means—and furthermore has similar properties to human memory in that you can more or less always add more, but then what you had before gradually gets degraded. Since neural nets are much closer to the actual mechanics of the brain as we know them, understanding human memory will probably involve finding ways to simulate holographic storage with neural nets.

With these facts in mind, the amount of information we can usefully take in and store is probably not 0.2 petabytes—it’s probably more like 10 exabytes. The human brain can probably hold just about as much as the NSA’s National Cybersecurity Initiative Data Center in Utah, which is itself more or less designed to contain the Internet. (The NSA is at once awesome and terrifying.)

But okay, maybe that’s not fair if we’re comparing human brains to computers; even if you can compress all your data by a factor of 100,000, that isn’t the same thing as having 100,000 times as much storage.

So let’s use that smaller figure, 0.2 petabytes. That’s how much we can store; how much can we process?

The next thing to understand is that our processing architecture is fundamentally difference from that of computers.

Computers generally have far more storage than they have processing power, because they are bottlenecked through a CPU that can only process 1 thing at once (okay, like 8 things at once with a hyperthreaded quad-core; as you’ll see in a moment this is a trivial difference). So it’s typical for a new computer these days to have processing power in gigaflops (It’s usually reported in gigahertz, but that’s kind of silly; hertz just tells you clock cycles, while what you really wanted to know is calculations—and that you get from flops. They’re generally pretty comparable numbers though.), while they have storage in terabytes—meaning that it would take about 1000 seconds (about 17 minutes) for the computer to process everything in its entire storage once. In fact it would take a good deal longer than that, because there are further bottlenecks in terms of memory access, especially from hard-disk drives (RAM and solid-state drives are faster, but would still slow it down to a couple of hours).

The human brain, by contrast, integrates processing and memory into the same system. There is no clear distinction between “memory synapses” and “processing synapses”, and no single CPU bottleneck that everything has to go through. There is however something like a “clock cycle” as it turns out; synaptic firings are synchronized across several different “rhythms”, the fastest of which is about 30 Hz. No, not 30 GHz, not 30 MHz, not even 30 kHz; 30 hertz. Compared to the blazing speed of billions of cycles per second that goes on in our computers, the 30 cycles per second our brains are capable of may seem bafflingly slow. (Even more bafflingly slow is the speed of nerve conduction, which is not limited by the speed of light as you might expect, but is actually less than the speed of sound. When you trigger the knee-jerk reflex doctors often test, it takes about a tenth of a second for the reflex to happen—not because your body is waiting for anything, but because it simply takes that long for the signal to travel to your spinal cord and back.)

The reason we can function at all is because of our much more efficient architecture; instead of passing everything through a single bottleneck, we do all of our processing in parallel. All of those 100 billion neurons with 500 trillion synapses storing 2 quadrillion bits work simultaneously. So whereas a computer does 8 things at a time, 3 billion times per second, a human brain does 2 quadrillion things at a time, 30 times per second. Provided that the tasks can be fully parallelized (vision, yes; arithmetic, no), a human brain can therefore process 60 quadrillion bits per second—which turns out to be just over 6 petaflops, somewhere around 6,000,000,000,000,000 calculations per second.

So, like I said, a few petaflops.