Yes, I know, I missed a great clickbait headline
opportunity… But as Wayne Campbell would say, I'm not worthy!
Anyway, I've been studying the issue of potential
revolution in artificial intelligence and robotics extensively for the last few
years. I've read thousands of pages, including most of the recent major
economics
papers,
and general
public
books
(this
one just ordered). And I'm actually doing a good deal of learning about
artificial intelligence, itself, as in studying
textbooks.
Note to economists: You'll actually be very
comfortable with much, or all, of AI. One of the five major types of machine
learning (which basically is AI) is statistical modeling, especially Bayesian
Monte Carlo. And all five, which can be powerfully combined, involve very
advanced mathematics. And the goal of the ML is always maximizing a utility
function! (or equivalently, minimizing a loss function)
So ML will be, in fundamental ways, very familiar
and comfortable to an economist. But there are some important differences. It's
not perfect optimization or bust (or no pub). You recognize that's usually not
realistic. You just try to get the highest utility score you can, even if, as
is usual, you don't know if that's the global optimum. And you don't just assume a
local optimum is the global optimum.
I think there's a lot economists can learn from ML
scientists, but there's also, I think, a lot ML scientists can learn from
economists. For example, with evolutionary algorithms (one of the five ML
methods), algorithms compete to be the highest scoring on the utility function,
and they mutate to evolve, and can even reproduce sexually! where the features
of two algorithms are spliced and combined. In the explanations of this I've
seen, and it's benefits and costs, I have not seen something that occurred to
me quickly from financial economics.
With options, you place much greater value on
assets that have a high variance, all other things equal. As the upside
benefits you far more than the downside costs you. Likewise, consider sexual reproduction. Suppose one species has good members, but not great ones; there's
little variation, but they always reproduce for a good, but not great, average.
Now, suppose another species is poor on average, and has some members which overall are not very good, but they do have some great feature, or features. That second species will evolve to take over the
first with sexual reproduction, because eventually it will produce offspring
which have the best of the greatest members all spliced together in some
offspring. Those offspring, even if initially rare, will then eventually take over because
they will be so super-fit, and become the bulk of the future populations. So,
the lesson is that sexual reproduction of algorithms will be more likely to
have benefits that outweigh the computational costs, the greater the variability (and it might be very beneficial to look for ways to juice up the variability.)
Note, in the natural world at least, there are complications, as genes proliferate in future populations, or not, in a tribe, so the success of a whole tribe is an important factor. And a tribe's success depends on some specialization and variety of members. Also, note that in the natural world, especially, sexual reproduction is extremely costly, yet it has nonetheless evolved to be in every organism beyond the most simple, so the logic for it is very powerful.
This whole thing does give you an idea of the interdisciplinary nature of artificial intelligence, and how the field could benefit from learning from, and collaboration with, other fields -- and vice versa.
Note, in the natural world at least, there are complications, as genes proliferate in future populations, or not, in a tribe, so the success of a whole tribe is an important factor. And a tribe's success depends on some specialization and variety of members. Also, note that in the natural world, especially, sexual reproduction is extremely costly, yet it has nonetheless evolved to be in every organism beyond the most simple, so the logic for it is very powerful.
This whole thing does give you an idea of the interdisciplinary nature of artificial intelligence, and how the field could benefit from learning from, and collaboration with, other fields -- and vice versa.
The AI field is fascinating, and I have gained many
insights already. I will eventually have a long blog post/article on all of
this, but given the size of this issue, it's best not to wait, and get down
some insights and ideas along the way. This is both to help my thinking and
understanding, and hopefully to provide some.
So, without further ado, let's get to today's
topics.
Krugman's
Hot Dogs
The blog of The Bank of England, Britain's central
bank, recently had a sanguine post
on the whole robots/AI's issue. The author, B of E economist John Lewis, uses Nobel
Prize winning economist Paul Krugman's hot dog mini-model as a
justification for his optimism:
Technology can lead to workers being displaced in one particular industry, but this doesn’t hold for the economy as a whole. In Krugman’s celebrated example, imagine there are two goods, sausages and bread rolls, which are then combined one for one to make hot dogs. 120 million workers are divided equally between the two industries: 60 million producing sausages, the other 60 million producing rolls, and both taking two days to produce one unit of output. Now suppose technology doubles productivity in bakeries. Fewer workers are required to make rolls, but this increased productivity will mean that consumers get 33% more hot dogs. Eventually the economy has 40 million workers making rolls, and 80 million making sausages.
The problem, however, is this: Substitute roll
makers and sausage makers with low-skilled workers and high-skilled workers.
Suppose the economy starts at 4 billion low-skilled workers and 400 million
high-skilled workers, and produces $80 trillion/year in goods and services.
Now suppose that advances in AI and robotics
result in there being, for the purposes of producing pecuniary goods and
services, 12 billion low-skilled workers, but only 500 million high-skilled
workers. With the old production processes, you had a ratio of 1 high-skilled
worker to 10 low-skilled workers. Keeping these processes, you would need just
5 billion low-skilled workers, not the 12 billion you now have, with the influx
from Robotia and the AI Republic. So what happens to the other 7 billion low-skilled
workers, machine and human?
Well, you could go to other production processes
that use less high-skilled workers to low-skilled workers, again, where
low-skilled workers now include the machine kind. But the problem may be, and I
think is, in the real world, that production processes that have a low
proportion of high-skilled workers produce a lot less per worker. And if they
produce a lot less per worker then it does not make sense in the market to do
them, unless they cost a lot less per worker. Thus,
real cost per worker must drop.
Now, the high-skilled workers can be utilized for
the old high-skilled production process, so the market must pay them at least
that old wage. But now to get businesses in the free market to employ a very
heavy low-skilled production process, they will only do it if the wage of the
low-skilled workers drops, and maybe very dramatically, perhaps to poverty
level – or below. And a human subsistence wage need not be a floor, as the
subsistence wage for a machine may be well below that, with the cheap solar
power of the future. And, with how smart and advanced these machines may become,
they may have very low maintenance costs. They may mostly maintain themselves,
and each other.
Essentially, the problem is, what if the roll
makers don't have the skills to make the sausages. Then, you can't just shift
to this new higher production economy with more hot dogs produced, but with a
lower proportion of roll workers and a higher proportion of sausage workers. So
what do you do? You train roll makers to now make sausages? What if sausage
making requires far more education and skill? This may be very costly, and the
benefit may be mostly hard to recoup for the payer of this education and
training. Meanwhile, governments may be unwilling or unable to pay, especially
with the horrifying power of the billionaire funded right, and its government
always bad, always a waste, propaganda machine.
And aside from that, many workers may be too old
to practically learn advanced new skills. And if even a fifth of the population
do not have the cognitive and other abilities necessary to learn high-skilled
sausage making, then do you have a fifth of the population permanently
unemployed if the new robots can do any and all of the roll making at
less than a human subsistence wage?
So, what then? If you can't just shift enough
workers into the high-skilled sausage making, what do you do?
You use a different production process? This may
be a lot less productive. You end up with three rolls per sausage. It generates
a lot less GDP per worker. It won't be done unless the roll workers' wages go down.
But the sausage workers' wages won't go down; if they did, they would be hired away into old style 1-1 facilities, but even moreso, with rolls very cheap now, sausages will command relatively more money (the price of their complement, some would say necessary complement, has gone down). And, if sausages command more money, then so also will the relatively rare sausage makers, all other things equal.
So, the sausage makers' wages go up, but the roll workers' wages will have to go down – and as far as it takes, to employ even the least productive of the roll workers, if they aren't to be unemployed. And this is just the strong trend we've seen over the last two generations for low-skilled workers.
But the sausage workers' wages won't go down; if they did, they would be hired away into old style 1-1 facilities, but even moreso, with rolls very cheap now, sausages will command relatively more money (the price of their complement, some would say necessary complement, has gone down). And, if sausages command more money, then so also will the relatively rare sausage makers, all other things equal.
So, the sausage makers' wages go up, but the roll workers' wages will have to go down – and as far as it takes, to employ even the least productive of the roll workers, if they aren't to be unemployed. And this is just the strong trend we've seen over the last two generations for low-skilled workers.
Now, the B of E blog post author Lewis does
concede, "In the interim, the transition might lead to unemployment,
particularly if skills are very specific to the baking industry. But in the
long run, a change in relative productivity reallocates rather than destroys
employment"
But think about what that could mean here if
robots become able to do almost anything that a low-skilled human could do,
only at a cost below human subsistence. You would have to "in the long
run" give almost every low-skilled human a college education, and not a
Potemkin college education, but the academic and cognitive skills of the
typical graduate of a well-respected major university, like my employer, the
University of Arizona.
This may be a very, very long run, with massive, or
catastrophic, poverty, unemployment, and homelessness in the interim. And the
public will have to dramatically change their attitude about the size and value
of government, because the private sector won't come close to funding this,
given the severe free market problems with education. There's good reason government
funds the vast majority of education in every first world country.
The key intuition here is if you just add a ton of
low-skilled workers, including low-skilled workers from the countries of Robotia and the AI Republic, without high-skilled workers, or with proportionally way
less high-skilled workers, then you're going to lower the marginal product and average product of low-skilled workers greatly, and thus their market wage.
You could say, like Clouseau, problem sol-ved,
just have low-skilled workers become high-skilled workers, like the roll makers
in Krugman's model just becoming sausage makers. But that is one incredible
endeavor to suddenly massively increase the world's, or any country's,
education level. I would love to do so, and would certainly vote to
invest in it, but as I said in a recent
post:
How are we supposed to get the vast majority of men, and women, up to this level of skill and education?
To do so would take a regime shift in our politics, and in public understanding of economics. By and large, one of our two major parties not only does not believe in global warming, or evolution for that matter, they don't believe in externalities, asymmetric information, natural monopoly, contracting limitations and costs, and basically anything that says the pure free market is imperfect (except in cases where it benefits the rich). But providing a massive increase in the education, skills, and general capabilities for most of the population is something that free market companies could only extract a small fraction of the benefits from in profits. And therefore they alone would grossly underprovide this.
The externalities, contracting and enforcement problems and costs, adverse selection and other asymmetric information, and so on, are profound and enormous. This is why general education has historically been predominantly publicly funded. To say that now, so that most of the population won't go the way of horses, we have to enormously increase our investment in Heckman-style early human development, education, public nutrition, healthcare, and more, from prenatal until at least well into a person's 20's, is to say that we should have an unprecedented increase in governments' size and roles.
Right now, this is impossible, as the Republican party is dogmatically against any government, except for a small number of areas; mainly military, police, courts, prisons, and perhaps minimal public infrastructure and education.
A recent
OECD paper put it dryly, but seriously, "If the tasks that complement
machines become increasingly complex and demanding, the employment prospects
for workers lacking certain skills may deteriorate." (page 23)
Are robots and AI like shipping containerization?
Lewis gives the example of the revolution in shipping
containerization, which plummeted the workers needed per ton to transport goods, yet workers were redeployed, and the amount of shipping increased dramatically to counter:
Take the humble shipping container. Transporting goods in pre-packed locked containers, which can be lifted straight onto a lorry or train, yielded enormous savings relative to having cargo transported in crates which needed loading and packing individually at each port. Their inventor estimated that the combined savings on labour costs, time at the dockside and insurance for breakage and theft reduced the price of a tonne of cargo 39-fold. Bernhofen et al calculate this led to an eight fold increase in bilateral trade between countries with container ports. Whilst employment fell, productivity of labour increased nearly 20 fold. For the shipping industry this wasn’t a massively disruptive technology- though trade patterns changed, the industry became moreconcentrated and ironically less profitable.
But by reducing the cost of trading, containerisation opened up the possibility of new supply chains and trading arrangements that were previously too expensive to undertake. And, inso doing, the resultant trade flows led to a substantial spatial reallocation of economic activity.
A point I'd like to make here is that with the kinds of robots, machines,
and AI's we may see in the future, it won't be just a clever idea that
eliminates many specific jobs in a specific business, so demand increases in
other businesses and workers move there, or to different jobs in the same
business. It will instead be whole classes of work eliminated. A whole class
of work, or a whole class of skill or ability, say dexterous movement and
sorting with good visual recognition, will be eliminated from pretty much every business, every industry, in the
entire economy – permanently – as happened to horses.
It's whole
skills that will be eliminated from employment by these new robots and
AI's, across every business, not some specific jobs that will be eliminated in
a specific business, so I'll take my skillset elsewhere. Your skillset may no
longer be needed anywhere. It may be replaced everywhere, or replaced in 90% of
businesses.
And again, there are problems with, well, then
production will just increase 10-fold to sop up all those low-skilled workers.
There are serious bottlenecks in high-skilled workers and raw materials,
and there are inelasticities of demand, at least for certain types of products.
Is
increasing GDP share of capital the only distributional concern, or even the biggest?
This is the issue you always hear, and the one
Lewis discusses (and downplays). But you rarely hear a similar issue that's
perhaps more important, and I think is probably far more important in the short
and medium run: It’s not just that robots and other AI's may increase the share
of the pie going to capital owners. It's that they could cause massive
increases in inequality in how the labor share is distributed among laborers.
These technologies can send the superstar, or winner-take-all, phenomena to the moon. Moreover, an increasing share of
workers, due to revolutionary advance in these technologies, may find that the
wage the market will pay for their education, skills and other characteristics
drops below minimum wage. Or well below, for those who think cutting the
minimum wage is the answer.
What happens if these robots and AI's make it so
that 10% of the population is basically unemployable at a wage above minimum?
then 20%, 30%,… Suppose at a grocery store cameras watch everything shoppers
put in their carts, and know what it is; the cart is mechanized and follows
you, the cameras also recognize your face and have your payment info on file,
so no need for any workers at checkout other than a security guard. Dexterous
robots can do the vast majority of stocking and unloading,… We are not far from
this now, and going the rest of the way looks not that hard in the next 30
years, from what I've learned of this technology.
Then, about 90% of the grocery store jobs are
gone. Where are those 90% going to be redeployed? They're low-skilled laborers,
where they can be redeployed are similar places where robots can also do those
kinds of manual jobs; fast food, manufacturing, waiting tables,… And, as we've
seen, no, people are not willing to pay much more, in money and time, for the
human touch, to be able to chat with the checker, or teller, or waiter. They'd
rather have more badly needed time and money to spend with their own families
and friends.
You could say, ok, the economy would expand to
have just 10 times the production then, but you hit the bottleneck of lack of
high-skilled laborers to do this that I talked about earlier. In addition, you
hit a lot of inelasticities of demand; people aren't going to just buy 10 times as many groceries, so we need 10 times as many grocery stores. Those who do still have jobs and money
can only eat so much (the wealthy with money, I would think, would tend to buy high prestige items that require a high proportion of highly skilled labor to produce). And costly raw materials can also become bottlenecks.
So, this is not so easy to dismiss with, in the
past,… We actually did have to increase educational levels and skills to
prevent mass unemployment with past technological advance. But, (1) Government
rose to the challenge. We didn't have such a dominant and billionaire powered,
government always bad, pure free market always good, right wing. And, (2) The
societal level of education and skill necessary with the kinds of robots and
AI's we're probably going to see will be very high indeed. It will take an
incredible increase in human capital investment, from prenatal to college and
beyond.
A simple
test of the veracity, or completeness, of normal-employment robot/AI arguments
Finally, I'd like to note a key problem with
Lewis's post. His argument applies to any
technology. There's no discussion of the specific technology here, AI and
robotics, at least as far as what it's capable of. What the technology is, and how capable it
is, or may become, is irrelevant to his arguments. These arguments are supposed to work for any technology, so no need to consider
how good this technology might realistically get, what it might be capable of
doing. That doesn't matter; his arguments will always apply. It wouldn't matter how good engine and motor technology got, we would always find alternative uses for horses, product prices would drop, production would expand,... Except these arguments didn't work for horses.
And with humans, and specifically low-skilled humans, you have the same technology irrelevance in Lewis's arguments. Even if robots and AI become capable, over
the next generation, of doing any
pecuniary work at all that a low-skilled human can do, at less than a human subsistence
wage, this will still not be a problem for low-skilled humans "over the long
run", where presumably "the long run" will not be long enough to
ruin, or end, their lives.
At least explicitly and clearly, Lewis does not discuss how the long run could be generations, and it may be extremely hard to reach this long run. Nobel Prize winning economist Jan Tinbergen talked about "a race between education and technology" to prevent massive unemployment and/or income inequality. It is possible, depending on the technological advance, and
the resources put into education, that technology can run far ahead for a very
long time, even that education could never catch up for a potential
super-technology like AI. I discuss this in a guest
post at the blog of Haverford College economist and Roosevelt Institute fellow Carola Binder.
At least explicitly and clearly, Lewis does not discuss how the long run could be generations, and it may be extremely hard to reach this long run.
So, I'd like
to offer a rule. Any time someone offers an argument that robots and AI cannot
cause a massive unemployment/poverty problem, if those arguments do not
consider how good the technology may, with significant probability, get, then those arguments have a fatal
flaw, or are at least significantly incomplete.