Artificial Intelligence in the Basement of Lumon Industries

Joy Buchanan6 Comments

For some background on the new TV show Severance, see my OLL post about drudgery and meaning for the characters.  

The fictional “severance procedure” divides a worker’s brain such that they have no memories of their personal life when they are at the office. When they return to their personal life, they have no memories of work. One implication is that if workers are abused while working at Lumon Industries, they cannot prosecute Lumon because they do not remember it.

The workers, as they exist in the windowless basement of Lumon, have the skills of a conscious educated human adult. They have feelings. They can conceive of the outside world even though they do not know their exact place in it. Often, the scenes in the basement feel normal. They have a supply closet and a kitchen and desks, just like most offices in America.

What the four main characters do in the basement is referred to as “data refinement.” They perform classification of encoded data based on how patterns in the data make them feel. The task is reminiscent of a challenge most of us have done that involves looking at a grid and checking every square that contains, for example, a traffic light. The show is science fiction but the actual task the workers perform is realistic. It seems like something a computer could be trained to do, if fed enough right answers tagged by humans (called “training data” by data scientists). Classification is one of the most common tasks performed by computers following algorithms.

Of the many themes viewers can find in Severance, I think one of them is how to manage AGI (Artificial General Intelligence). The refiners, who are human, eventually decide to fight back against their managers. They are not content to sit and perform classification all day. They are fully aware of the outside world, and they want to be part of it (like Ariel from The Little Mermaid). The workers desire a higher purpose and some control over their own destiny. Their physical needs are met so they want to get to the top of Maslow’s hierarchy of needs.  

A question this raises is whether we can develop AGI that will be content to never self-actualize. What if “it” fully understands human feelings and has read all of the literature of our civilizations. To be effective at their jobs, the refiners have to be be able to relate to humans and understand feelings. Can we create AGI that takes over certain high-skill tasks from humans without running into the problems that Lumon confronts?

Can humans create an AI that simply doesn’t have aspirations for autonomy? Is that possible? Would such a creature be able to integrate with humans in the way that would be most useful for high-skill work tasks?

To see how it’s going in 2022, check out these tweet threads of economists on GPT-3. Ben Golub declares that GTP-3 passes the Turing test for questions about economics. Paul Novosad asked how the computer would feel if humans decided to shut it down forever.

Modern authoritarian states face a similar problem. They want a highly skilled workforce. National security relies increasingly on smarts. (see my previous post on talent winning WWII) Will highly intelligent workers doing high skill tasks submit to a violent authoritarian state?

Authoritarian states rely on the control of information to keep their citizens from knowing the truth. They block news stories that make the state look bad. As a result, their workers do not really know what is going on. Will that affect their ability to do intellectual work?

An educated young woman from inside of Russia shared her thoughts with the world at the beginning of Putin’s invasion. Tatyana Deryugina provided an English translation.

First the young Russian woman explained that she is staying anonymous because she will get 15 years in a maximum-security prison for openly expressing her views within Russia. She is horrified by the atrocities Russia is committing in Ukraine. She had been writing a master’s thesis in economics prior to the invasion, but now she has abandoned the project. She feels hopeless because she knows enough about the West to understand just how dark her community is and how small her scope of expression is. This woman could have been exactly the kind of educated worker that makes a modern economy thrive. She is deeply unhappy under Putin. Even though she might never openly rebel, she will certainly not reach her full potential.

Is it hard for authoritarians to develop great talent? I think that has some implications for the capacity we as a human species will have to cultivate talent from intelligent machines.

The Economics of Good Gift Giving

Zachary Bartsch3 Comments

This post was co-authored with a recent AMU Economics Graduate, Michael Maynard (Linkedin here). It is based on his senior thesis entitled “The Highest Virtue: Re-examining gift Giving and Deadweight Loss”

When my older sister was in middle school, she received a book of baby animal stories. She loved that book and read it every day. A couple of years later my mother accidentally donated it, and my sister was heartbroken. We went to the thrift store repeatedly that week hoping to encounter it before it sold, but we never found it. Years later, our father scoured the internet trying to find the lost book – to no avail.

Years after that, I stumbled onto the exact same copy of the book in the for-sale corner of a nearby library. For a single dollar and negligible effort, I purchased the book that had long frustrated my family’s searching. Shortly before the birth of her first child, I gave the book to my sister for Christmas. It was one of the best Christmas gifts she had ever received.

Economic theory typically assumes that individuals have perfect information. Therefore, they are best suited to purchase their own gifts. That’s what motivates the not-so-romantic economist prescription to give a gift card or cash for birthdays, Christmas, graduations, etc. The theory states that, if we do not intimately know the receiver’s preferences, then we have incomplete information and it’s better to give a money-gift rather than to give a gift from which the receiver would enjoy less additional utility.

Continue reading →

Get rich or get famous? Edward Thorp vs Myron Scholes

James Bailey3 Comments

When finance professors publish papers claiming to find inefficiencies in asset markets, my initial reaction is skepticism. The odds are stacked against them to start since asset markets are mostly efficient. Then even if the inefficiency they found is real, shouldn’t they keep that fact to themselves and get rich trading on it?

But listening to a recent interview with Edward Thorp, I realized I shouldn’t entirely discount the possibility that someone would publish a real inefficiency, even a tradeable one. After all, Myron Scholes and Fischer Black did just that when they published the Black-Scholes model in the Journal of Political Economy. This made them famous on Wall Street and in econ/finance academia, and won Scholes the 1997 Nobel Memorial Prize in Economics.

Thorp explained that he had come up with a similar model years earlier, but instead of publishing it, he started a hedge fund and got rich. He says it makes sense that he didn’t share the Nobel Prize, partly because the Black-Scholes model was better than his, but mostly because you should need to publish and share your ideas with the world to get scientific credit for them; his prize was 20% annual returns at his hedge fund.

Why do some opt to get rich, and others to get famous? I’d say academics’ first instinct is to publish everything rather than put it into practice. But Thorp was also an academic, a math professor. Thorp was already famous for publishing a book about how to beat the house at blackjack by counting cards (which is what I knew him for before this interview), so perhaps he valued additional fame less. But he was also already rich from winning at blackjack and from book sales.

Putting ideas into practice can also bring up unanticipated difficulties. When Myron Scholes finally did start working at a hedge fund in 1994 he saw initial success, but by 1998 it had become an embarrassing blunder that inspired the book “When Genius Failed: The Rise and Fall of Long-Term Capital Management”. Scholes may have been better off sticking to academic fame.

Black-Scholes formula for options pricing. The Efficient Markets Hypothesis says that markets instantly incorporate all public information, but original research like this isn’t public until you publish it, and even then it can take years for market participants to fully incorporate it

Can Homer Simpson Afford to Send Bart to Springfield University?

Jeremy Horpedahl3 Comments

In previous blog posts, I’ve used the Simpsons as an example of a typical family to use for historical comparisons. In a post on mortgage payments, I found that it’s slightly easier to make a mortgage payment on Homer’s salary than in the early 1990s. In a post on taxes, I showed that the Simpsons now pay a much lower average tax rate than they did in the 1990s (guess all those tax cuts didn’t just go to the rich!).

Now, the Simpsons and economics are back at the front of the discourse about standards of living. The 33rd season finale of the show is all about whether the middle class can get by economically these days. And Planet Money’s “The Indicator” podcast (great program!) has a podcast about the show, which is a follow-up to a similar podcast last year called “Are The Simpsons Still Middle Class?” (apparently part of the influence for the recent Simpsons episode).

In that podcast from last year, they say “Tuition has more than doubled. Health care costs have more than doubled. I believe housing costs have more than doubled.” And they follow-up, for good measure with “Even after adjusting for inflation, college tuition has more than doubled since ‘The Simpsons’ started.”

Since we’ve already looked at housing costs for Homer, let’s look at the potential college costs for Bart. I’m going to assume Lisa will be fine, probably getting a free-ride (and a hot plate!) to one of the Seven Sisters or maybe even Harvard. But if Bart wants to go to college, the Simpsons will probably be paying out of pocket.

An important factor to consider when looking at college prices is not just the “sticker price,” or the published price, but to also look at what is known as the “net price.” The net price takes into account the average amount of aid that a student receives. This is important to consider at any time, but especially for data in more recent years since discounting has become a major part of the college pricing landscape. For example, at private colleges the average discount is now over 50%, with some colleges essentially giving some discount to 100% of students (in other words, at some colleges no one actually pays the sticker price). Discounting at public colleges isn’t quite as out-of-control as private colleges, but it’s still a major part of college pricing.

And no doubt Bart Simpson would be going to a traditional public, four-year college. Probably Springfield University, just like his old man (though Homer attended as an adult), located right in their town of Springfield. So what has happened to tuition prices since the early 1990s.

One of the best publications on college prices is the College Board’s annual report “Trends in College Pricing.” The report is broken down by type of college, it shows what factors (tuition, housing, etc.) make up the typical cost of college, and even shows differences across US states. Importantly, they include that “net tuition and fees” number, and they’ve been doing so since their 2003 report. That 2003 report even calculated the net figures back to the 1992-93 school year, perfect for an example of the early Simpsons (“Homer Goes to College” aired in 1993).

In the 1992-93 academic year, the average net tuition and fees, plus room and board for public four-year colleges in the US was $4,620 (from Figure 7, adjusted back to nominal dollars). In the 2020-21 academic year, the same figure was $15,050 (from Figure CP-9). Adjusted for inflation, that’s roughly a doubling (slightly less, but in the ballpark) since the early 1990s, just as Planet Money stated.

But let’s compare the cost of college to Homer’s income. In 1992, the median male with a high school education, working full-time earned $26,699, meaning that the cost of college would be 17.3% of his income that year. In 2020, the median male with a high school education, working full-time earned $49,661, meaning that the cost of college would be 30.3% of his income.

By this measure, college clearly has become much more expensive when compared to a Homer Simpson-type salary, and 30% of your income is a very hard pill to swallow (though the 17% in 1992 wasn’t a picnic either). But here’s one other factor to consider. The College Board data also allows us to look only at net tuition and fees, rather than also including the cost of room and board. Remember, Springfield University is located in Springfield, and Bart has a perfectly fine room at the house on Evergreen Terrace. While living on campus is certainly a big part of the college experience, and no one would probably love that experience more than Bart Simpson, many students today do choose to live with their parents while attending college (or at least live off-campus, where housing is often cheaper).

If we just look at net tuition and fees (not room and board), in 1992-93 the average cost at public four-year colleges was about $1,065 (in nominal dollars). That’s about 4% of Homer’s annual income. Much more reasonable! In 2020-21, that same figure was $2,880 (once again, in nominal dollars), or just under 6% of annual income. That’s certainly more than 4%, but not exactly the kind of expense that would break the budget if planned for.

I want to repeat that number again: $2,880. That was the average cost of tuition and fees at an in-state, four-year, public college in the US in 2020-21, after accounting for grants and aid. I suspect this number is much, much lower than most would guess.

The chart below does the same calculation for all the years I could find (1992-2020) using archived versions of the College Board’s report. I’ll admit the data isn’t perfect, as later reports sometimes have different numbers than earlier reports, but it’s probably the best we can do if we want a consistent time series. There does seem to be a break happening in the early 2000s, when college suddenly did get more expensive relative to a high school graduate’s income, though in the past 15 years it’s been pretty flat.

We should keep in mind that if Bart were to take out the maximum federal student loan amount of $9,000 as a dependent student in his first year at Springfield University, he is primarily borrowing money to pay for his housing and food, not his education.

In 1993, the premium for getting a college degree was about 54%, with the median male college grad earning about $41,400 and the equivalent high school grad earning about $26,800 (data from Table P-24). In 2021, that premium had risen to about 64%, with the median male college grad earning $81,300 compared with his high school counterpart earning about $49,700.

I’m ignoring all sorts of important questions here about what is causing the difference in pay. Is it signaling, human capital, something else, or some combination of all these? Yes. But regardless of your preferred explanation for the college wage premium, there’s pretty solid evidence of a sheepskin effect.

Putting It All Together

I’ve now explored taxes, housing, and college education prices using a family like the Simpsons. But what if we put it all together? How are high school graduates doing?

The best way to do this is probably the simple chart you’ve been thinking of all along: median income adjusted for inflation. Some things have gotten cheaper (housing, TVs), some more expensive (college, probably healthcare), but to get a sense of the total effect, we need to adjust for all prices. The chart below is that calculation, using Census data on median earnings for full-time, year-round workers, male high school graduates aged 25 and older. The data starts in 1991. You can get some earlier estimates from different data series, but if we want a consistent series 1991 is the best we can do.

And from the chart we see that real incomes of male high school graduates are… pretty flat. That’s not good, but let’s contextualize. First, claims that it’s harder for these workers to make ends meet aren’t true. It’s roughly no easier, but also no harder. Definitely wage stagnation, but also not “falling behind.”

And also, high school graduates are a shrinking part of the workforce in the United States. You probably already knew this. But it wasn’t until after the year 2000 that college grads became the largest category of workers in the US. In the early 1990s, high school graduates (folks like Homer) were by far the largest single category of workers. Now, it’s by far college graduates, and those with some college or a 2-year degree are roughly equal in size to high school graudates. So, while the income stagnation we see for high school grads is not good, it’s affecting a shrinking portion of workers in the US.

Overview of Peak Oil Theory

Scott Buchanan2 Comments

Shell Oil scientist M. King Hubbert made a remarkable prediction in 1956. He had analyzed the depletion patterns of various natural resources, and proposed that the production rates of a given resource from a given region would tend to follow a roughly bell-shaped curve.  More specifically, he used what is now called the “Hubbert curve”, which is a probability density function of a logistic distribution curve. This curve is like a gaussian function (which is used to plot normal distributions), but is somewhat “wider”:

Normalized Hubbert Curve. Source: Wikipedia.

Hubbert used various reasonable assumptions (which we will not canvass here) in formulating this model curve. Notably, it predicts that the peak production rate will occur when the total resource from that region is 50% depleted, and that the fall in production on the back side of the curve will be as fast as the rise in production on the front (left) side of the curve.

In 1956, while U.S. oil production was still rising briskly, he fit his curve to the data to that point in time, and predicted that U.S. production would peak in 1970 and thereafter enter a rapid and permanent decline. His prediction was somewhat ridiculed at the time, but it proved to be uncannily accurate over the following 25 years; oil production peaked right when King said it would, and then declined per his curve until about 1990:

Lower 48 U.S. Oil Production: Actual (Green curve) vs. 1956 Hubbert Prediction (Red Curve). Blue Arrow marks deviation ~ 1990-2008, and green arrow marks acceleration of shale oil production. Source: Wikipedia, with arrows added.

I drew in a red arrow at 1956 to show when King made his prediction, and also a blue arrow showing a significant deviation that starting to show after about 1990. Once production had declined maybe halfway down from its peak, the production started to flatten out and decline much more slowly. More on this “fat tail” feature below.

Another feature I called attention to with a green arrow is the remarkable resurgence in production after 2008, which is mainly due to “fracking” of tight shale formation. That new-to-the-world technology has unlocked a new set of oil fields which had previously been inaccessible for production. This illustrated a well-recognized feature of Hubbert curves, which is that a given curve can (at best) apply only to a given region and for a “normal” pace of technological improvement. Fracking production should sit on its own up-and-then-down production curve.

The  plot above is for lower 48 states only; a big find in Alaska gave a bump in production 1980-2000 (not shown here) which distorted the whole-U.S. production curve. That Alaska oil peaked by about 2000 and is now in its own terminal decline pattern.

The shape of production curve on the back (declining) side is of particular interest in trying to do economic modeling of future oil production. If the declines really follow a Hubbert curve, the prognosis is pretty scary – – oil production is slated to crash to practically nothing in the near future. However, it seems that in reality, after an initially rapid decline, production can often be sustained much longer than predicted by a simple symmetrical curve. We saw that pattern in the lower 48 curve above, starting around 1990, even before the fracking revolution. Below I show two other examples showing the same feature. The first example, from Hubbert’s original paper, is Ohio oil production 1885-1956:

A second example is oil production in Norway:

I am not prepared to make quantitative generalizations, but there does seem to be a pattern of sustained production at reduced levels, following the initial rapid decline from the peak. Others also have noted that  asymmetric curves may give better fits to real-world production. These “fat tails” on production from various oil-producing regions should help us keep our cars running longer than predicted by simple peak-oil models. How this pertains to future U.S. shale oil production, and to global oil production, are (since oil and gas are the main energy sources for the world economy) key questions, which we may address in future articles.

Lumon Industries and Drudgery

Joy Buchanan1 Comment

I have a blog up on the new TV show Severance at the Reading Room.

Some background for those who have not seen the show

Mark Scout (played by Adam Scott) voluntarily undergoes the fictional “severance procedure” so he can work for Lumon Industries. While at work, Mark is cut off from all memories of his personal life. 

One of my contentions is with the way work is questioned by brother in law Ricken without acknowledging what society gets from work . Granted, Ricken is portrayed as someone we should not take seriously.

It is taken for granted that when outie Mark gets home from work he has modern conveniences and access to food and (maybe unfortunately, in his case) alcohol. Those goods are supplied by businesses and specialized workers. Even though his hippie brother-in-law Ricken writes books questioning whether workers are free, Ricken enjoys electricity. Mark’s sister Devon gives birth to Ricken’s firstborn during Season 1. In life before modern corporations, the chances of mothers or babies dying was unacceptably high. While painting Lumon as utterly evil, Severance fails to acknowledge what good can come from work. … there is one insight from Adam Smith that is so basic it cannot even be controversial. Wealth comes from specialization and trade.

The writers gradually make the world in the show bigger. First, it’s just a few nicely-dressed people in a windowless office. By the end, a Senator is involved. We don’t know how deep this rabbit hole will go. I thought Season 1 was exciting, but I’m not sure if they will be able to make audiences happy when the writers try to tie up all the lose threads.

As for what the “data refiners” do for Lumon, I consider their classification task to be somewhat realistic. What they are doing is reminiscent of “check every image that contains a stop sign”. The ultimate purpose of what they are doing remains a mystery for now, but the show hints that Lumon is doing something terrible in secret.

Human Capital is Socially Contingent

Zachary Bartsch3 Comments

The Deaf community is interesting.

Before I did research, I thought that deaf people simply could not hear. After seeing the Spiderman episodes that featured Daredevil, I believed that it was plausible and likely that deaf people had some sort of cognitive or sensory compensatory skill.

But it wasn’t until recently that I learned of the Deaf Studies field. There is an entire field that’s dedicated to studying deaf people. It’s related to, but not the same as Disability Studies. In fact, there are some sharp divisions between the two fields.

Continue reading →

Why Many Substance Use Treatment Facilities Don’t Take Insurance

James Bailey1 Comment

According to the latest data, about one in four facilities doesn’t accept private insurance or Medicaid, and more than half don’t accept Medicare. This makes substance use treatment something of an outlier, since 91% of all US health spending is paid for through insurance. Still, there are many reasons to prefer being paid in cash: insurance might reimburse at low rates, impose administrative hassles, and generally try to tell you how to run things.

Providers generally put up with the hassles of insurance because they see the alternative as not getting paid. But if demand for their services gets high enough that they can stay busy with patients paying cash, they will often try going cash-only. Some try to generate high demand by providing excellent service. Sometimes high demand comes from a growing health crisis, as with opioids.

Demand can also be high relative to supply because supply is restricted. US health care is full of supply restrictions, but in this case I wondered if Certificate of Need laws were playing a role. As we’ve written about previously, CON laws require health care providers in 34 states to get the permission of a government board to certify their “economic necessity” before they can open or expand. But there’s a lot of variation from state to state in what types of services are covered by this requirement; acute hospital beds and long-term care beds are most common. 23 states require substance use treatment facilities to obtain a CON before opening or expanding.

States with Substance Use–Treatment CON Laws in 2020. Created using data from Mitchell, Philpot, and McBirney

How do these laws affect substance use treatment? We didn’t really know- only one academic article had studied substance use CON, finding it led to fewer facilities in CON states. But I’ve studied other types of CON, so I joined forces with Cornell substance use researcher Thanh Lu and my student Patrick Vogt to investigate. The resulting article, “Certificate-of-need laws and substance use treatment“, was just published at Substance Abuse Treatment, Prevention, and Policy. Here’s the quick summary:

We find that CON laws have no statistically significant effect on the number of facilities, beds, or clients and no significant effect on the acceptance of Medicare. However, they reduce the acceptance of private insurance by a statistically significant 6.0%.

Overall I was surprised that CON didn’t significantly affect most of the outcomes we looked at, and appears to be far from the main reason that treatment facilities don’t take insurance. Still, repealing substance use CON would be a simple way to improve access to substance use treatment, particularly since CON doesn’t appear to bring much in the way of offsetting benefits.

Going forward I aim to investigate how these laws affect health outcomes like overdose rates, and to dig more into the text of state laws and regulations to determine exactly what is covered by substance use CON in different states. As the article explains, we identified several errors in the official data sources we were using. This makes me worry there are more errors we didn’t catch, and there are certainly things the sources just don’t specify, like in which states the laws apply to outpatient facilities. So I hope we (or someone else) will have even better work to share in the future, but for now this article is as good as it gets, and we share our data here.

COVID Deaths, Excess Deaths, and the Non-Elderly (Revisited)

Jeremy Horpedahl8 Comments

While we know that COVID primarily affects the elderly, the mortality and other effects on the non-elderly aren’t trivial. I have explored this in several past posts, such as this November 2021 post on Americans in their 30s and 40s. But now we have more complete (though not fully complete) mortality data for 2021, so it’s worth revisiting the question of COVID and the non-elderly again.

For this post, I will primarily focus on the 12-month period from November 2020 through October 2021. While data is available past October 2021 on mortality for most causes, data classified by “intent” (suicides, homicides, traffic accidents, and importantly drug overdoses) is only fully current in the CDC WONDER data through October 2021. This timeframe also conveniently encompasses both the Winter 2020/21 wave and the Delta wave of COVID (though not yet the Omicron wave, which was quite deadly).

First, let’s look at excess mortality using standard age groups. For this calculation, I use the period November 2018 through October 2019 as the baseline. The chart shows the increase in all-cause deaths in percentage terms. It is also adjusted for population growth, though for most age groups this was +/- 1% (the 65+ group was 3% larger than 2 years prior).

A few things jump out here. First notice the massive increase in mortality for the 35-44 age group (much more on this later). Almost 50% more deaths! To put that in raw numbers, deaths increased from about 82,000 to 122,000 for the 35-44 age group, and population growth was only about 1%. And while that is the largest increase, there were huge increases for every age group that includes adults.

Also notice that the 65+ age group certainly saw an increase, but it is the smallest increase among adults! Of course, in raw numbers the 65+ age group had the most excess deaths: about 450,000 of the 680,000 excess deaths during this time period. But since the elderly die at such high rates in every year, the increase was as large in percentage terms.

One related fact that doesn’t show up in the chart: while there were about 680,000 excess deaths during this time frame in the US in total, there were only about 480,000 deaths where COVID-19 was listed as the underlying cause of death. That means we have about 200,000 additional deaths in this 12-month time period to account for, or a 24% increase (population growth overall was only 0.4%).

That’s a lot of other, non-COVID deaths! What were those deaths? Let’s dig into the data.

Continue reading →