In August, UPS announced that driver compensation will increase to an astounding $170,000 over the next few years following hard-fought union negotiations. And, earlier this year, Home Depot said it would spend $1 billion to increase its starting wage, likely in response to Walmart’s decision to also bump up its starting salaries. 

Across the economy, low-wage workers seem to be winning historic gains. 

At the same time, Gallup polling indicates that nearly half of consumers view current economic conditions as poor and are about as pessimistic as they were in mid-2009, when unemployment was more than double what it is today. 

So what’s going on?

A series of economic shocks—including COVID, pandemic-related stimulus spending, the war in Ukraine, and rising interest rates—have buffeted the economy over the last three years, throwing supply and demand out of balance in many industries. 

Nick Bunker, an economist at Indeed Hiring Lab, told me that sectors were differentially affected by these shocks. Some industries engaged in massive layoffs, while others couldn’t get enough labor. Many occupations expanded but some have contracted sharply.

This volatility has created an opening for adventurous workers to boost their earnings by moving into growing industries, Bunker said. But some more risk-averse workers have gotten trapped in jobs with declining real wages.

The practical upshot for workers—especially lower-income workers—is that now is the time to seize the initiative. If you’re in a booming industry, you should ask your boss for a raise—and look for a new job if you don’t get it. If your industry isn’t booming, you should consider switching to one that is.

It’s also a great time to move up the career ladder. Some companies have gotten desperate enough to drop degree requirements for some jobs. Others are offering extra training or career development opportunities to workers looking to switch careers. So now may be a great time to win a promotion or break into a higher-paying occupation—and lock in higher wages for the long term.

Subscribe now

The unexpected compression

Average real wage growth was negative for most of 2021 and 2022, largely because companies can update their prices faster than workers can negotiate wage increases. But over the last year, real wage growth has been positive.

Throughout this period, lower-end workers have fared better than most. Low-wage workers suffered the least wage erosion in 2021 and early 2022, and they’ve enjoyed the strongest real wage growth over the last year.

A comprehensive study of the post-pandemic labor market by economists David Autor, Arindrajit Dube, and Annie McGrew found that low-wage workers have benefited more from the pandemic-era economy than their high-wage counterparts. In fact, low-end pay gains since 2020 have erased a quarter of the increased wage inequality that occurred since 1980.

They credit this “unexpected compression” to the tightness of the low-wage labor market. Low-end workers could afford to be picky thanks to high demand for low-wage labor across many sectors of the economy.

Wage growth has moderated in the past couple of months, but job quits and job openings—measures of labor market tightness—are“still fairly elevated,” Bunker told me. And that’s particularly true at the low end of the labor market.

In response to this tightness, employers have been offering higher wages and better benefits to entry-level workers. The military, for example, has said that high wages in the fast-food industry have made it difficult to meet its recruiting goals.

“This is the most challenging recruiting environment the Department of Defense has probably ever faced,” the National Guard Association said in a June press release.

Workers who shop around have benefited the most. Autor and his co-authors found that half of the total wage gains for low-income workers over the past few years came from job switching rather than negotiating higher wages at an existing job. Proactive workers were disproportionately rewarded for their efforts, while more stationary laborers sometimes fell behind. 

As Rachel, a manager at a fast-casual steakhouse, put it to me, “the power [my workers] have is in leaving, in trying the next thing and shopping around.”

The restaurant industry bounces back

When the pandemic first hit in the spring of 2020, people stopped traveling and started shopping online. The hotel and tourism industries atrophied, while warehouse employment surged.

Warehousing jobs were “a huge growth area in the labor market particularly in the period when COVID was having the biggest impact on people’s decision-making process,” Bunker said. 

At first, wages stagnated for workers in hospitality and shot up for those transporting packages. But as workers flooded out of leisure and hospitality into warehousing, hotels were forced to raise wages as well.

Similarly, the restaurant industry shed more than three hundred thousand workers between 2019 and 2020. It is just now recovering that gap. Meanwhile, spending on food services has not only recovered to its pre-pandemic peak but is actually 6% above it.  

As a result, restaurants can afford to offer higher wages to their remaining workers. That’s good news for those workers, though they may have to work harder than they did before the pandemic.

Mike, a waiter in Los Angeles, has kept the same job the past few years. He told me he is now making more money than he did before the pandemic. His restaurant “learned that we were able to operate with less people, with a smaller workforce. So the money has been the same or better,” he said.

Rachel echoed this story, saying her steakhouse has “paid for the people that stuck around and suddenly the jobs that you had four people doing… we have one guy doing. So, yeah, give him $22 an hour to do it.”

Rachel herself is a job-switcher. She previously had a job doing political strategy work, but she started to feel burned out. She took a job in the restaurant industry and has worked her way up into a front-of-house management role at a fast-casual steakhouse.

Workers have also benefited from changing jobs within the restaurant industry. Darryl, a waiter, told me he took the opportunity to move “into more upscale dining,” and, as a result, has “been able to stay ahead of inflation.”

Not everyone has benefited

Karen, a New York security guard, has not seen her salary keep up with inflation. As a result, she's struggling to buy groceries for her family of four. She told me that she has thought about trying to get a new job, but she’s too worried a new job might be worse.

Similarly, John, who has been a delivery driver since the start of the pandemic, told me his commissions have fallen while prices have risen. That’s why he called the current life of a delivery driver “a fight for survival.”

Why have some workers who kept their jobs gained while others suffered? The answer goes back to the newfound worker power identified by Autor and his co-authors. Some workers like John happened to work in an industry—delivery driving—where workers had little bargaining power. As a result, they haven’t been able to get raises to offset the rising cost of living.

I don’t want to second-guess the decisions of individual workers such as Karen and John who may have specific reasons for holding on to their current jobs. But the evidence suggests that a lot of workers in their position—suffering from low and stagnant wages during a generally tight labor market—would benefit from switching jobs.

Are these gains sustainable?

Even workers who have reaped the benefits of a tight labor market can’t help but wonder if heightened salaries are temporary.

On the one hand, many of the gains from the temporarily hot—what economists call cyclical—labor market are transient. Even if low-wage workers are able to get a big raise by switching jobs, they’re still at risk of being laid off if the economy slows.

“For every period with 3 percent unemployment, you've got a period like after the Great Recession, where a lot of people are losing their jobs. And so for an individual who works for 35 or 40 years and experiences downturns and booms, it kind of averages out,” David Deming, a Harvard economist who studies labor markets, said in an interview.

Bunker expressed a similar fear, saying that many low-income workers have found themselves in sectors especially subject to discretionary spending—industries that could contract substantially in a recession.

In a June paper, Demming highlighted one way workers can lock in wage gains. The research shows that educated workers earn more than their non-college-educated counterparts because over time they develop specialized skills and receive promotions. To really narrow wage inequality, Demming told me in an interview, non-college workers should find a path to the sort of “professional jobs” that allow continued advancement over the course of their careers.

“Many people who don't have a college degree end up pigeonholed into dead-end jobs that don't fully utilize their talents and capabilities and don't allow them to grow. And that seems like a real shame and a real waste of potential,” he said.

The tight labor market has meant rising wages for low-wage workers, but may have a secondary effect that could be more important over the long run: allowing less educated workers to get jobs that previously required a college degree.

The Burning Glass Institute, a labor research institution, found that a lot of employers are dropping their degree requirements. This trend began in the tight labor market of the late 2010s, but the pandemic has accelerated it. Similarly, a January survey found that as many as 34% of employers had dropped degree requirements in the last year with the goal of attracting more applicants. 

Moving into jobs that previously required more education isn’t the only way to move up the career ladder. Rachel took advantage of the pandemic labor market to “work [her] way up” by moving to the restaurant sector and then staying in the same job. She received training, became a manager, and now has skills that could translate into permanently higher wages. Similarly, Darrell’s move into high-end restaurants may provide him with a new and distinct skillset.

So while some workers have exploited newfound bargaining power to get raises above the rate of inflation, the long-term winners of this economy may be those who move into jobs or sectors formerly closed off to them.  

Aden Barton is an economics student at Harvard. You can follow him on X.

Subscribe now

When ChatGPT was released last fall, a lot of white-collar workers had a chilling thought: “could this technology do my job?” A May survey by CNBC and SurveyMonkey found that almost a quarter of workers fear losing their jobs to AI in the next few years, including 51 percent of workers in advertising and marketing and 46 percent of those in business support and logistics.

But for one white-collar profession—translators—this stopped being a theoretical question years ago. ChatGPT is based on a neural network architecture called the transformer that was first used by Google for machine translation in 2017. So competition from AI has been an everyday reality for human translators for more than five years.

Subscribe now

Marc Eybert-Guillon started his career as a translator in 2017. In 2020, he founded From the Void, a firm that helps video game makers localize their games for foreign markets.

“It's that meme of the guy with the noose around his neck,” Eybert-Guillon told me. The condemned man looks over at the guy standing next to him on the gallows and asks “first time?”

“We’ve been ‘in danger’ of being taken over by AI for 10 years now and it still hasn’t happened,” Eybert-Guillon said. “But we keep getting told that it’s going to happen.”

There are two big reasons AI hasn’t put many human translators out of work. First, human translators still do a better job in specialized fields like law and medicine. Translation errors in these fields can be very expensive, so clients are willing to pay extra for a human-quality translation.

Second, there has been rapid growth in hybrid translation services where a computer produces a first draft and a human translator checks it for errors. These hybrid services tend to be about 40 percent cheaper than a conventional human translation, and customers have taken advantage of that discount to translate more documents. Translators get paid less per word, but they’re able to translate more words per hour.

But while AI software has not put human translators out of work the way pessimists might have predicted, this isn’t an entirely positive story for translators either.

“I think rates for translators have stayed largely the same for 10 or 12 years,” said Mark Hemming, a translator in the United Kingdom. “I think it is harder to get work now. I think it's harder to get well paid work as well.”

The main lesson, though, is that workers in other industries don’t need to worry about AI taking over their jobs overnight. Automation is a gradual process that plays out over decades, not years or months. And there are plenty of ways human workers can adapt.

Some translators do work AI can’t

Barbara Leon is a freelance translator in Spain. She focuses on translating legal documents like contracts, trusts, immigration records, and divorce agreements between English and Spanish. There’s a robust market for human translation of documents like this because errors can be expensive. Moreover, some of the documents Leon produces are submitted to courts or government agencies that require sign-off by an expert human translator.

To translate a legal document accurately, Leon doesn’t just need to be an expert on the English and Spanish languages. She also needs a deep understanding of legal systems on both sides of the Atlantic. 

“If you put ‘trust’ in ChatGPT it's going to translate it to confianza,” Leon said. “But that's not what it means.” In reality, Leon says, there are 20 or 30 different ways to translate the legal concept of a trust to Spanish. Figuring out which meaning to use in any given sentence requires a sophisticated understanding of American law, Spanish law, and the context of the specific document she is translating.

It’s a similar story for Marc Eybert-Guillon’s work localizing video games. His firm provides translation for a wide variety of text, from character dialog to the labels on in-game items like weapons or magic potions. Often he needs to translate a single word or short phrase—too little context for an unambiguous translation.

To gain the necessary insight, he may play a pre-release version of the game or ask the game’s developers to explain it. This kind of exploration is beyond the capabilities of software like Google Translate or ChatGPT.

There’s also an important element of creativity to video game design, Eybert-Guillon argues. Machine translation software might provide a literal translation of a word like “dagger” or “pistol” when a more whimsical or obscure term might be a better fit. It’s hard to provide a good translation without understanding the gaming culture in a particular target market.

Mark Hemming told me about a friend who specializes in doing translations related to outdoor sports like climbing, mountain biking, and skiing. A company selling forks for mountain bikes might hire her to translate its product descriptions and installation instructions. Translators who aren’t into outdoor sports may not know the right vocabulary—and ChatGPT probably wouldn’t either.

In short, when you look closely at the market for translators, what you find is actually a bunch of distinct sub-markets. The skills required to translate text in a video game are different from those needed to translate a legal document or the installation instructions for mountain bike parts. Each of these translation tasks requires context beyond the four corners of the document being translated—context that a computer program likely doesn’t have. So clients that care about quality are willing to pay extra for an experienced human to do the job.

Subscribe now

Lower cost drives higher demand

While some customers are willing to pay a premium for high quality, there are plenty of other customers who are perfectly willing to accept lower quality work if it saves them money.

“A lot of use cases don't demand perfect fluency or accuracy,” said Arle Lommel, an analyst at the market research firm CSA, in a recent interview. “Obviously probably people would prefer something done by the best translator in the world, but usually the alternative to that is no translation, and that doesn't benefit anyone.”

So while there continues to be robust demand for fully human translation, there’s also growing demand for a hybrid approach where a computer generates the first draft and a person checks it for accuracy.

“The process of editing the output of machine translation was initially perceived to be viable only for very specific types of content,” said Alon Lavie, a professor at Carnegie Mellon and an executive at the machine translation company Unbabel. “And now I think the tide has kind of turned and most types of content, most use cases, I think it's clear that just editing the output of automated translation is more cost-effective, faster, and more efficient to do.”

Lommel estimates that this kind of service typically costs about 60 percent as much as a traditional human translation. And many clients take advantage of this discount to have more text translated.

While translators don’t like that discount, Lommel argues that the technology will enable them to translate more words per hour. In the long run, this might even lead to higher incomes.

Many of the translators I spoke to remain skeptical of this approach. They told me that machine translations are often so bad that it takes more work to fix them than it would have taken to produce a translation from scratch. Some found the experience so frustrating that they’ve stopped accepting this kind of work.

Yet Lommel says his firm’s data shows that machine-augmented translation is growing faster than conventional human translation. Because customers who buy cheaper semi-automated translation services don’t care as much about quality, these services can probably get away with using less skilled translators.

I see a parallel here to the way smartphones have changed the taxi industry over the last 15 years. Before the advent of smartphones, driving a taxi required extensive knowledge of local streets. Indeed, New York once required cabbies to pass a geography test before they could get a license.

Smartphones changed that. The Uber and Lyft apps supplied turn-by-turn directions that allowed unskilled drivers like me to drive without any special training or experience. That was not a positive development for existing taxi drivers because it devalued their hard-won navigational skills. But it was good for consumers, who got more ubiquitous and affordable taxi services. And it was also good for many new drivers—especially those who used Uber or Lyft as a temporary or part-time source of income.

If this analogy is right, then I would expect translation software to lower barriers to becoming a translator, which may reduce both the quality of translations and the average wages of translators. I can understand why this might drive experienced translators crazy. But it’s also easy to see why many customers would see it as a good trade. As Lommel says, a low-quality translation is often better than no translation at all.

Data from the Bureau of Labor Statistics is consistent with this interpretation. The number of translators and interpreters rose 19 percent between 2010 and 2021: from 58,400 to 69,400. But their median income, adjusted for inflation, fell by 8 percent. In 2010, a typical translator made $60,357 (in 2023 dollars). By 2021 it was $55,483.

CSA estimates that global revenue from translation, interpreting, and other language services grew by 23 percent (in inflation-adjusted dollars) between 2012 and 2022. Yet when CSA surveyed translators on the state of the industry, many of them were quite negative about their career prospects.

“It's a declining business,” wrote one translator in response to a CSA survey. “I still make decent money because I am a specialist (finance, tech, agriculture). I am setting up a new, completely unrelated business and hope to stop translating as this business replaces my current income.”

Automation is good for consumers

The advent of machine translation has not been catastrophic for human translators, but machine translation software does seem to be putting downward pressure on wages. And it’s easy to imagine that pressure intensifying in the coming years as AI technology improves further.

At the same time, machine translation has been unambiguously positive for consumers of translated content. Multinational companies, government agencies, and other organizations collectively produce billions of documents every year and only a tiny fraction of these are ever translated into even one other language. There are dozens of languages around the world with millions of native speakers.

This means that there is a ton of value that could be created by translating more documents into more languages. For example, companies tend to see a big increase in sales in a country when they translate their marketing materials into that country's native language. So cheaper translation services would enable a lot of companies to expand the market for their products around the world.

So when thinking about the impact of automation on workers, it’s important to remember that every worker is also a consumer. We all benefit when automation in other industries lowers the cost of the goods and services we buy. And while having your own industry automated isn’t fun, it’s usually not catastrophic.

Pessimists will counter that this time is different—that AI will soon become so sophisticated that it will put all human translators out of a job. Arle Lommel doesn’t buy it.

“My expectation is that so much stuff needs to be translated, that there's always going to be a role for the human, especially when you’re introducing new technologies and new topics,” Lommel said. “Machines are always stuck at some point in the past in terms of what they can do. Language is changing, and you need humans because the machines can only go off what’s been done before, not what is new.”

Maybe you find that convincing or maybe you don’t. But either way the pace of change matters.

If the introduction of neural translation software in 2017 had put every translator out of a job in 2018, that would have been a huge hardship for them and their families. On the other hand, if steady progress in machine translation technology leads to a gradual decline in translator employment over the next 10 to 20 years, that’s a more manageable situation. That will give older translators time to reach retirement age and younger translators time to switch careers.

If you’re worried that your own profession will get automated by AI software in the coming years, I hope that the story of translators helps to reduce your anxiety level. It seems very possible that AI software will start to automate other white-collar jobs over the next five to 10 years. But the key word here is start.

The first iterations of the software likely won’t be very good, leaving plenty of tasks in human hands. Early AI systems in your profession may increase demand for human workers to oversee and correct the work of automated systems. Even if that doesn’t happen, job losses are likely to play out over the course of decades rather than years, giving alert workers plenty of time to adapt.

Subscribe now

I expect that fans of Full Stack Economics will enjoy Understanding AI. For example, this morning I published an article about the likely impact of artificial intelligence on the labor market:

In the early 2010s, Silicon Valley thought leaders looked at the early success of companies like Airbnb and Uber, extrapolated wildly, and concluded that software was going to transform the entire economy. Today, AI thought leaders are looking at the early success of ChatGPT and Stable Diffusion, extrapolating wildly, and concluding that AI software is going to transform the economy and put tons of people out of work.

To be clear, I do think AI is going to be a big deal. I wouldn’t have started an AI newsletter otherwise. But as with the Internet, I expect the impact to be concentrated in information-focused industries and occupations. And most of the American economy is not information-focused: It’s focused on delivering physical goods and services like homes, cars, restaurant meals, and haircuts. It will be hard for AI to have a big impact on these industries for the same reasons that it’s been hard for Internet startups to do so.

In previous weeks, I wrote about the copyright lawsuits that threaten to upend the AI industry and about voice cloning technology that’s so good my mother couldn’t tell the difference. I’ll write more about the economic and policy implications of AI in the coming weeks.

Right now, Understanding AI is an experiment. If it proves popular, I’ll continue doing it indefinitely, which would mean keeping Full Stack Economics on the back burner. If Understanding AI isn’t so popular, I’ll switch back to doing Full Stack Economics full time in a couple of months.

Here’s what this means for you as a paying subscriber:

• Monthly subscriptions have been on hold since the start of February. I’ll resume them if and when I return to writing Full Stack Economics.

• Annual subscribers will get a pro-rated refund if I decide not to resume full-time work on Full Stack Economics. I expect to make this decision in a couple of months.

Again, if you’d like to follow my work at my new newsletter, Understanding AI, please click here to subscribe. Thank you so much for supporting my work.

As a D.C. resident and parent of three young children, these rules have never made sense to me. D.C.’s main child-care challenges are high costs and limited supply, not low quality. Requiring teachers to get college degrees will only make these problems worse. Higher costs could even force some parents to take their kids out of formal child-care centers in favor of home-based or unlicensed options that are likely to be of lower quality.

So why are D.C. officials pushing it? I recently listened to the 2020 book Head, Hand, Heart and it gave me a new way to think about the debate. The author, British journalist David Goodhart, argues that Western democracies—especially the U.S. and the U.K.—have developed an unhealthy fixation on college as the sole yardstick for worker quality.

College students develop cognitive skills that will serve them well in white-collar professions like finance, engineering, or marketing. But Goodhart also identifies two broad job categories where cognitive skills are less important.

One is hands-on professions like electricians, plumbers, welders, or workers on an assembly line. For the most part, you don’t need to go to college to do these jobs well. For example, in most states you can get a plumber’s license after a few years working under a master plumber. The rules are similar for electricians.

I suspect that a system like this would be a better way to improve the skills of child-care workers (who fall into Goodhart’s third category of care workers) than forcing all of them to get a college degree. Allowing child-care workers to upgrade their skills on the job would certainly impose less of a financial burden.

But the drive for child-care workers to be college-educated isn’t just about upgrading their practical skills. It’s also about raising their prestige. Americans consider it to be almost axiomatic that jobs that require a college degree are “more professional” than those that don’t. So if someone says that a profession doesn’t require a college degree, that’s almost taken as an insult. But it shouldn’t be.

The evidence for college-educated child care workers is weak

Advocates say there’s ample evidence that young children benefit when their care providers have college degrees. I’ve spent the last couple of weeks reading some of this research, and I came away unimpressed. There is a body of scientific research suggesting that college-educated teachers are more effective in the classroom. But these findings come with some important caveats.

First, most of the research has focused on classrooms at the preschool level or above. I found very few studies looking specifically at child-care facilities for children under the age of three.

Second, the arguments for teachers to get formal education change as you move from older to younger students.

It’s easy to see why you’d want, say, a high school physics teacher to have a college degree in physics. You need an in-depth understanding of physics to teach it well. But this kind of rationale gets less compelling at lower grade levels. Most adults—even those without a college degree—understand math and reading well enough to teach them to a kindergartener. And they definitely understand colors and animal noises well enough to teach those to a toddler.

So at the preschool level, research on teacher quality tends to focus not on subject-area mastery but on classroom style. For example, a 2003 literature review found that preschool teachers with a bachelor’s degree in early childhood education “were less authoritarian, and were rated higher in positive interaction and lower on punitiveness and detachment.”

I personally like the idea of child-care workers being more engaged and less punitive. But it’s not clear how much this ultimately matters. Indeed, it’s not clear that any aspect of preschool quality has lasting impacts on a student’s academic performance. Back in 2018, Vox’s Kelsey Piper dug into the research and found something surprising:

There’s a sizable body of research suggesting that kids who go through intensive education at the ages of 3 and 4 don’t really come out ahead in terms of academic abilities. By kindergarten much of their advantage has receded, and by second grade researchers typically can’t detect it at all.

On the other hand, there’s an equally substantive body of research suggesting that early childhood education produces a profound, lifelong advantage. Kids who enter intensive preschool programs are less likely to be arrested, more likely to graduate, and less likely to struggle with substance abuse as adults.

Piper’s hypothesis is that most of the benefits of preschool come not from its curriculum but from its function as child care. Reliable child care makes it easier for the child’s parents to work full time and advance in their careers. The resulting economic stability provides big benefits later in a child’s life.

While I haven’t found any research exploring this question, this seems likely to be even more true for children aged three and under. It may not matter very much what happens in the classroom as long as parents know their kids will be safe, happy, and have their basic needs met.

And importantly, the youngest children are the most labor-intensive to care for. The federal government recommends that preschools have 10 or fewer children per teacher. For two-year-olds, the recommended maximum is six children per adult, while for babies under a year old it’s just four.

If you’ve ever spent time with small children, this makes perfect sense. Older children are more capable of sitting still, following directions, and doing independent work. Babies and toddlers need constant supervision and attention.

One consequence is that raising the standards for teacher quality is much more expensive at the lowest age levels. Suppose it costs an extra $15,000 to hire a college-educated teacher. In a 15-student kindergarten class, that works out to $1,000 per child. In an infant classroom with four babies per adult, it’s $3,750 per child.

So at the kindergarten level and above, it’s probably worth paying extra for college-educated teachers even if the per-child benefit is fairly small. For babies and toddlers, on the other hand, the benefits would need to be substantial to justify the cost.

Yet I suspect that, if anything, the average toddler gets less value out of having a college-educated teacher than a kindergartener would. On the other hand, attention-hungry toddlers seem likely to benefit more from a low student-teacher ratio. So hiring a larger number of less educated, lower-paid teachers seems like the right approach.

Not every job needs a college degree

But let’s say I’m wrong and the caregiving style of daycare workers really does matter. It still seems pretty unlikely that two years of college coursework is an efficient way to impart the necessary skills. A child-care worker doesn’t need an in-depth understanding of why punitiveness and detachment are harmful—they just need hands-on instruction on how to be less punitive and detached.

This is especially true when you consider the demographics of child-care workers in D.C.: they are overwhelmingly foreign-born and many have limited English skills. Even if they can find the time and financial resources to pursue a college degree (and to be fair D.C. is offering some scholarships), they may not have the necessary language skills or other academic prerequisites. 

But the goal of many advocates isn’t just to improve the skills of child-care workers, it’s also to improve their social status. This is made explicit in one of the reports most commonly cited by advocates of D.C.’s approach, a 2015 report called Transforming the Workforce for Children Birth Through Age 8: A Unifying Foundation.

In Chapter 9, the authors draw a connection between educational requirements and prestige, complaining that “society has tended to perceive working with younger children as less demanding and prestigious work, a perception that runs counter to the science of child development and early learning.”

Taken literally, this is a non-sequitur. Whether a profession is or should be prestigious isn’t a scientific question. I think that what they really mean is that the science shows that the quality of early childhood education is important. And because it’s important, they reason, child-care work should be prestigious and well-compensated. And the way to do that is to require child-care workers to have a college degree.

But if the goal is to raise the compensation of child-care workers, it would be better to just do that. Indeed, D.C. recently set up a system to subsidize the pay of child-care workers. This should improve their lives without raising the cost to parents and without shoving anyone out of a job.

At the same time, it’s worth pushing back on the seemingly relentless drive to raise educational requirements for a range of occupations. Not everyone has the opportunity or aptitude to go to college, and the skills students develop in college aren’t needed to do most jobs across the economy.

When employers mindlessly demand a college degree based on a vague idea that college-educated workers are higher quality or more “professional,” it limits the opportunities of non-college educated workers—a majority of the population—for no good reason.

Last month, Pennsylvania Gov. Josh Shapiro announced an initiative to relax college degree requirements for a range of government jobs. “My view is that if you’re qualified for the job, that you should get the job here in Pennsylvania,” Shapiro said. This is a worthwhile reform that will allow for more upward mobility in the state. I hope D.C. policymakers are paying attention.

Economists have been anticipating a recession because the Federal Reserve has been aggressively raising interest rates over the last year in an effort to bring down inflation. On Wednesday, the Fed raised rates another 0.25 percent and signaled that more rate hikes were likely in the future.

The campaign against inflation seems to be working. Last month the Bureau of Labor Statistics announced that prices rose 6.5 percent in the year ending December 2022. That’s the lowest inflation rate in more than a year, though it’s still well above the Fed’s 2 percent inflation target.

Big rate hikes often often lead to job losses, and in the second half of 2022, the labor market really did seem to be softening. The pace of job growth steadily declined between July and December. However, even December’s job growth figure was above the long-term average. And January’s gain of 517,000 was far above historical norms, suggesting that the job market is still getting tighter even as the Fed raises interest rates.

Indeed, I think the chart above actually understates how impressive that 517,000-job figure is. 

The economy gained about 600,000 jobs per month, on average, over the course of 2021. So last month’s job growth doesn’t look that unusual when you compare it to the previous couple of years. But we have to remember that the job gains of 2021 came shortly after employers laid off a historically unprecedented 22 million people between February and April of 2020. Indeed, I started this chart in October 2020 because otherwise the huge labor market swings between February and September would make the rest of the chart unreadable.

A lot of that 2021 hiring was employers staffing back up to normal levels as the economy reopened. But we’re now three years past the start of the pandemic. Job growth today is mostly not pandemic rebound hiring. So to get a sense of just how impressive that 517,000-job figure is, it’s helpful to compare it to the pre-pandemic years.

Between 1990 and 2019, there was just one month when the economy gained more than 500,000 jobs: a 530,000-job gain in May 2010. Even if you adjust for population, there were only two other months, both during the 1990s boom, when the economy added jobs at a faster pace than it did last month.

Recent headlines have given a somewhat misleading impression because high interest rates have disproportionately affected the high-profile technology and media sectors. There have been a lot of articles about layoffs in Silicon Valley, and that’s reflected in data showing that the information sector lost 10,000 jobs, on net, between November 2022 and January 2023. 

But those layoffs have been dwarfed by robust hiring in the leisure and hospitality sector—which gained 192,000 jobs between November and January—and health care—up 111,000 over two months.

Here’s another way to look at the health of the job market:

This shows the ratio of workers to population for those between the ages of 25 to 54. This is a useful indicator because it excludes those over 55 (who may be retired) and those under 25 (who may be in school).

By this measure, the economy hasn’t quite rebounded to the January 2020 peak of 80.6 percent, but it’s very close at 80.2 percent. The 0.4 percent difference, combined with the ultra-low unemployment rate, makes me think there are still a few people being kept out of the labor market due to COVID.

Still, it’s remarkable how quickly the economy rebounded from the April 2020 low point. It took less than two years for the prime-age employment to population ratio to rise from less than 70 percent in April 2020 to 80.1 percent in March 2022.

In contrast, it took a full decade for this measure to rise from 74.8 percent in December 2009 to 80.6 percent in January 2020. If the recovery of the 2010s had been as robust as the post-COVID recovery, we would have returned to the previous peak eight years earlier—in 2011 rather than 2019.

Slow wage growth isn’t all bad news

The least impressive part of today’s jobs data is wages. According to the Bureau of Labor Statistics, average wages rose by 0.3 percent between December and January, and by 4.4 percent over the last year. With the inflation rate well above 4.4 percent, this means most workers’ real wages have actually been falling.

That’s not great, but there are a few silver linings.

One is that wages are likely to be a lagging indicator. It takes a while for employees to notice they have bargaining power, demand higher wages, and for the resulting raises to go into effect. But if inflation continues to fall while the labor market stays strong—as seems increasingly likely—we may see the opposite process occur over the next couple of years. Inflation may fall back down to the Fed’s 2 per percent target, while wages continue to grow at a faster pace.

We’re also continuing to see a decline in wage inequality.

Tight labor markets are good for all workers, but they tend to be best for workers at the bottom end of the wage scale. The economic boom of the 1990s produced strong wage growth for low-paid workers, while the recessionary periods of the early 2000s and the early 2010s saw low-wage workers falling further behind highly compensated workers.

Since 2015, gains for low-wage workers have once again been exceeding those higher up the income scale. And the pandemic seems to have widened the gap.

As a result, workers at the bottom end of the income distribution have seen their wages almost keep pace with inflation over the last two years, while higher-paid workers have seen their real wages fall significantly behind. 

A final upside of slow wage growth is that it should make the Federal Reserve feel more comfortable ending its campaign of rate hikes in the coming months. The Fed’s nightmare is a wage-price spiral where employees start expecting big raises to offset anticipated inflation, and employers automatically pass the costs on to consumers. If this feedback loop became entrenched, as it did in the 1970s, it could become much harder to bring down the inflation rate.

But the latest jobs report continues to show no sign of that happening. Wages are lagging behind inflation, they aren’t driving it. That should give Fed officials more confidence that inflation will continue coming down even if they bring their rate hikes to an end. Which makes it less likely that they’ll tighten too much and bring about the recession economists keep warning us about.

In recent weeks I’ve been thinking the most about generative AI software like ChatGPT and Stable Diffusion. These technologies have advanced rapidly and I expect them to have large and unpredictable effects on the economy in the coming years.

I want to spend the next few months reporting on these technologies, the companies behind them, and their economic and policy implications. And while I expect some of you will find that interesting, it’s not the product you signed up for. So today I activated Substack’s “pause paid subscriptions” feature. This means that if you’re a monthly subscriber, you won’t be charged for the next few months.

In a couple of weeks I’m going to launch a new “pop-up” newsletter focused on AI. It will be my focus for the next few months and I hope you all sign up. Later in the spring, I expect to switch back to posting here at Full Stack Economics with a fresh crop of story ideas inspired by my AI reporting. I’ll restart paid subscriptions at that point.

If you’re an annual subscriber, Substack will automatically extend your subscription by the length of the pause. But if you’d rather have a pro-rated refund now, please reply directly to this email and I’ll take care of it.

Thanks for supporting my work!

Fare-jumping and high-trust societies

Harry Morse writes: “I'd be interested in a longer write-up about fare jumping and high-trust societies.”

Harry is referring to a tweet thread I did on Wednesday that went a bit viral. Helen Andrews, an editor at the American Conservative who lives here in Washington DC, observed that she’s been seeing a lot of people jumping over the fare gates on the DC subway. I’ve noticed the same thing—it’s way more common than it was before the pandemic. I tweeted that “I think a lot of liberals underestimate how much this kind of behavior erodes the norms of social trust and solidarity that make societies work well.”

Let me try to unpack my thinking about this.

The smooth working of a modern society depends on people mostly following the rules even in the absence of strict enforcement:

• Most people don’t shoplift even if they think they could get away with it.

• Most Americans fill out their tax returns honestly even though the IRS would be unlikely to catch them if they cheated.

• Most people keep the promises they make in contracts even though the courts don’t provide a cost-effective remedy for small-dollar breaches.

When norms of trust and cooperation weaken, it adds friction to a lot of everyday interactions. A relatively low-stakes example is that if shoplifting becomes more common, stores start putting more merchandise behind glass, inconveniencing shoppers. If shoplifting gets really bad, companies start shutting down stores altogether.

I learned about a more profound example from Matt Yglesias, who last year wrote that “Italy and Greece are dominated by small, closely held businesses with family-centric management that are reaping huge economic gains by cheating on their taxes. Even the best-run of those companies tend not to expand or professionalize because to do so successfully, they’d have to actually pay what they owe.”

In a 2019 study, economists Bruno Pellegrino and Luigi Zingales blamed Italy’s slow economic growth on its culture of small, family-owned firms. While this culture seems superficially counterproductive, they argue that it may be an “optimal response to the Italian institutional environment” because “among developed countries, Italy stands out for its patronage-based banking sector, its inefficient legal system, and the diffusion of tax evasion and bribes.”

This is one of many reasons I found it alarming when the US elected Donald Trump president. Trump is not only notorious for cheating on his taxes, he has been totally shameless in defending his behavior, arguing during the 2016 campaign that his implausibly small tax bill “makes me smart.” Trump is also notorious for stiffing contractors and then daring them to sue him, knowing that the costs of taking him to court would exceed the amounts he owes them.

Overall, American society is still a relatively high-trust society. If you heard about someone you know cheating on your taxes, you’d probably be embarrassed for them, you wouldn’t think “that makes them smart.” But while those norms are pretty durable, they aren’t infinitely so. The more the public sees people flouting the law and facing no consequences, the more cynical they’ll become and the worse our collective behavior is likely to get.

So this is the kind of thing I had in mind when I tweeted about the erosion of social trust. As many, many people pointed out to me on Twitter, there are distinctions you could draw between cheating on your taxes and jumping over a turnstile. Tax cheats are probably more wealthy than turnstile jumpers, and they cost the government far more money. And to be clear, turnstile jumping isn’t anywhere close to being the most important example of eroding public trust in the US.

However, it is one of the most in-your-face examples of antisocial behavior, at least for those of us who live in urban areas. My neighbor might cheat on his taxes but I can’t see him doing it. The highly visible nature of turnstile jumping is problematic because this kind of behavior can be contagious. If I see someone jumping a turnstile while I pay for my ride, I might feel like a sucker and wonder if I should jump the turnstile myself.

Will lax enforcement of subway fares turn the US into Italy? Obviously not. Will it have an impact on the margin? It’s hard to be sure. I have no idea how you’d go about investigating this kind of question empirically. But intuitively it seems like constantly seeing other people ride the subway without paying is going to make the average Washingtonian at least a little bit more cynical about the society they live in. And I bet that bleeds over into other antisocial activities like cheating on their taxes.

The real threat of AI centralization is law, not hardware

Scott writes: “After reading a lot into ChatGPT and thinking back to other machine learning projects like autonomous taxis, it seems like modern machine learning tools have really gotten to a point where the only viable way to stay competitive is to have a gigantic data center on par with AWS or Azure. Do you think there is any hope for more localized machine learning software, or do you think it will all be consolidated to cloud services for the foreseeable future?”

I think it’s worth distinguishing between training neural networks and using them. Training a neural network tends to be an extremely resource-intensive process—so much that the most advanced neural networks are likely to require the resources of a large data center for the foreseeable future.

But it’s often much less resource-intensive to use a neural network once it’s been trained. Text-to-image software like Stable Diffusion and DALL-E, for example, have versions you can download and run on a PC.

For several years, big tech companies like Apple and Google have been working to add specialized hardware to smartphones to accelerate neural network operations. This allows some AI software to be run locally on your phone, allowing it to function without an Internet connection and without the privacy worries that come with uploading personal data to the cloud. I expect this silicon to get more powerful over time.

With that said, I think the biggest centralizing factor over the next few years is likely to be access to training data rather than computing hardware. What we’ve found over the last decade is that neural networks get better and better as you throw more data at them. Image generation networks, for example, are trained by looking at millions of images of existing works of art. A small startup—to say nothing of an individual—just doesn’t have the resources to collect a large enough data set to compete with the big guys.

So I suspect the big question for neural network centralization won’t be where the software runs but how it’s licensed. For example, one of the leading text-to-image networks, Stable Diffusion, was trained using a publicly available data set of images harvested from across the web. Copyright holders have sued Stable Diffusion arguing that they needed a license to train their networks using their copyrighted images.

As far as I know (I’ll be reporting on this more in the coming weeks), there’s no caselaw to tell us whether they’re right. And what the courts decide will have profound implications for how this industry evolves. If the plaintiffs win, then anyone wanting to train deep neural networks for image generation will need to negotiate with a bunch of copyright holders for permission to use their images. Google, Facebook, Apple, and OpenAI (generously funded by Microsoft) can probably afford to do that. Smaller organizations probably can’t. Moreover, these licensing deals would probably require ongoing royalty payments, which would preclude anyone from releasing the trained networks under a permissive open source-style license.

So if plaintiffs win these lawsuits, I think the likely consequence would be to centralize control over these technologies. It might stop being possible to download large open source models and run them on your PC. And while it wouldn’t become impossible for startups to build these systems, the barrier to entry would get a lot higher.

Statehood for DC, Puerto Rico, and other territories?

Steven von Christierson writes: “Hi, perhaps you could address the pros and cons from the perspective of the territories of the United States and the US government of remaining a territory or of becoming a state. This might include DC, Puerto Rico, the Marshall Islands, Guam, and any others I’ve missed.”

I think the pros and cons differ among the territories you mention. DC is part of the continental United States, DC residents pay the same federal taxes as residents of other states, and there are already a lot of federal laws that treat DC as if it were a state. So making DC a state wouldn’t make a huge difference in terms of substantive economic policy. Rather, the main issues are about sovereignty and representation.

DC residents, myself included, think it’s unfair that we don’t have voting representatives in our national legislature. We also don’t like the fact that Congress sometimes overrules the decisions of our elected representatives in ways that wouldn’t be allowed if DC were a state.

Opponents mostly focus on the fact that we’d get two senators. It does seem a little silly for DC to have the same number of representatives as California. But Vermont and Wyoming have fewer residents than DC and they each have two senators.

The issues are different for Puerto Rico, a predominantly Spanish-speaking island with 3.2 million residents. DC residents overwhelmingly want statehood, but voters in Puerto Rico are much more ambivalent. I haven’t followed this debate closely, but I think one factor is that you can imagine Puerto Rico eventually becoming an independent nation in a way that’s inconceivable for DC.

Puerto Rican statehood would also have more significant policy consequences. Many Puerto Ricans are not required to pay federal income taxes, and many federal programs treat territories like Puerto Rico differently than states. For example, the matching formula for Medicaid gives territories like Puerto Rico less money than they’d get if they were states.

I think the most important impact, however, might be in changing how Americans think of the island. When Hurricane Maria devastated the island in 2017, Americans didn’t feel the same sense of urgency about it as they would if the same thing happened to Florida or Texas. Racism probably plays a role here, but I think a more important factor was that a lot of Americans just don’t know that Puerto Ricans are American citizens.

If Puerto Rico was the 51st (or 52nd) state, then every school child would learn about it the same way they learn about Hawaii and Alaska today. Over time, I think this would subtly but profoundly change how people think about the island, and make mainland citizens more concerned about what happens there.

I think the considerations for America’s smaller territories, like Guam and the US Virgin Islands, are similar to those for Puerto Rico. However, they have the added problem that they’re really small. The next largest territory after Puerto Rico is Guam with around 150,000 people. That’s about a quarter as many people as Wyoming. I don’t think it makes sense to say DC is too small to be a state with 700,000 people. But it’ll be a harder sell to give an island with 150,000 residents two senators.

Don’t invest in art

Atul Mori writes: “How is fractional investment in art through companies such as Masterworks? Does the management make most of the money? In general, is art a good vehicle for investment?”

My Full Stack Economics co-founder Alan Cole wrote a great piece last year on just this question. In a word: no. Art, and collectibles in general, tend not to be good long-term investments.

When you invest in stocks or bonds, the value of your investment can go up in two ways: the asset itself can become more valuable, and it pays interest (in the case of bonds) or dividends (in the case of stocks). Companies also sometimes buy back their shares, which gives an extra boost to share prices.

In contrast, if you invest in a piece of art, all of your returns have to come from rising asset values. And while there will always be a few pieces of art that become super famous, there’s no reason to think the average piece of art (or even the average “blue-chip” piece of art, whatever that means) will appreciate more rapidly than the economy as a whole. So I would expect an art portfolio to underperform a portfolio of stocks and bonds even before you account for the quite high fees charged by Masterworks.

Incidentally, this same point applies to other assets that don’t pay dividends, including gold or Bitcoin. Over the long run they are very likely to under-perform a traditional portfolio of stocks and bonds.

I’ve got two favors to ask. First, please ask me questions for a forthcoming “mailbag” post (like this one). Questions about the economy are of course welcome, but I’m also happy to share my thoughts on other topics—personal finance, the future of artificial intelligence, my political philosophy, life in Washington DC, or whatever else you’re interested in.

Please use Substack’s comment feature to pose your questions so that it’s easy for me to keep track of them. But if for whatever reason you don’t want to post a question publicly, sending me an email directly works too.

Second, I’d be grateful if you could click here and fill out a 10-question survey. Paying subscribers like you make my work possible, and I want to make sure I'm delivering the best possible product. So letting me know what’s been working and what hasn’t would be a huge help.

Thanks as always for supporting my work.

A week later, the BLS announced that prices actually fell slightly between November and December. The annual inflation rate for December was 6.5 percent, the lowest it’s been since October 2021.

Then on Wednesday, we learned that the situation is similar for the producer price index, which measures costs faced by businesses. It fell a surprisingly large 0.5 percent between November and December, while the annual inflation rate for producers was 6.2 percent. And on Thursday, the Labor Department announced that there were 190,000 new unemployment claims filed last week—the lowest number of new claims in several months.

In short, we’re seeing two salutary trends at the same time: inflation is coming down, while the job market continues to boom.

This wasn’t supposed to happen. When the Fed began raising rates last March—and especially after the Fed started raising rates in large 0.75 percent increments in June—a lot of economists assumed that this would push the US economy into a recession. By August, even Fed Chair Jerome Powell seemed to be expecting the worst. In a widely-discussed speech, he warned that the Fed’s rate hikes were likely to lead to “slower growth” and “softer labor market conditions,” and bring “pain to households and businesses.”

He didn’t use the word “recession,” but everyone knew what he meant.

But it’s starting to look like inflation might come back down without a recession or significant job losses. Last month’s 223,000-job gain was the smallest in two years, but it was still above the roughly 200,000 new jobs we need each month to keep pace with population growth. And today’s unemployment rate of 3.5 percent is as low as it’s been at any point since 1969.

It’s too early for Powell to take a victory lap. The inflation situation is fluid, and the inflation rate could start to creep back up in the coming months. But it now looks like there’s a real possibility that the recession everyone was dreading in 2023 won’t happen. We might get an “immaculate disinflation” instead.

A complex inflation picture

You might object that the annual inflation rate is 6.5 percent—that is, prices in December 2022 were 6.5 percent higher than in December 2021. That’s obviously much higher than the Fed’s 2 percent target. However, this is a case where the annual inflation figure is misleading.

You might remember back in August when Joe Biden said there had been no monthly inflation in July, and people made fun of him for it. His claim was technically accurate—prices in July 2022 were actually slightly lower than they’d been in June—but it was misleading because the month-to-month inflation readings for June, May, and March had been scorchers.

The situation now is quite different. Once again, prices fell slightly between November and December. The difference is that this came after several other months of low or negative inflation. Indeed, the annualized inflation rate between June and December of 2022 was just 1.9 percent. If the next six months are like the last six months, the Fed will achieve its 2 percent inflation target.

The big question, of course, is whether we should expect the next six months to look like the last six months.

And the honest answer is we don’t know. There have been several different factors pushing the inflation rate in opposite directions, and we can’t be sure how they’ll net out in the coming months. But there is some reason for optimism.

The low inflation of the last six months is mostly the result of rapidly falling gas prices. Over the last six months, the annualized “core” inflation rate (which excludes food and energy) was 4.6 percent. If gas prices stop falling and core inflation stays at its current level, then headline inflation will come in significantly above the Fed’s 2 percent target.

On the other hand, one of the biggest components of core inflation is shelter, and shelter inflation has been high and rising in recent months. But as longtime readers know, the CPI’s shelter component lags behind “spot rents” (the rent a new tenant would pay if they signed a lease tomorrow) by about 12 months. And the rate of inflation for spot rents (the red and blue lines in the chart above) peaked about a year ago. This means that we should see shelter inflation rate, as measured by the CPI, peak in the coming months.

And so this means that while the Fed hasn’t fully accomplished its inflation-fighting goals, it looks like the end may be in sight. We can expect headline inflation to stay well above 2 percent for the first few months of 2023, due in part to high shelter inflation. But later in the year, shelter inflation is very likely to decline. If we don’t suffer any more pandemics or geopolitical crises, it’s easy to imagine headline inflation getting close to the Fed’s 2 percent target by the end of 2023.

And while it’s too early to say for sure, it’s looking increasingly likely that we’ll get there without large-scale job losses.

Why haven’t we had a recession yet? Again, nobody knows. People have been extrapolating from the Great Inflation of the 1970s. Between 1979 and 1981, Fed chair Paul Volcker jacked up interest rates to tame inflation, triggering a severe recession in the process. Many economists have naturally assumed that Powell’s aggressive rate hikes would have a similar impact.

But the situation today is not the same as the 1970s. We’ve only had a couple of years of high inflation, compared to the nearly decade-long experience of inflation in the 1970s. Maybe there hasn’t been enough time for inflation expectations to become entrenched the way they were back then.

Or maybe modern economies are just less inflation-prone than the economies of the 1970s. Japan has spent the last 30 years mostly battling deflation, and there’s reason to think Western democracies have become more like Japan over time, with an aging and slow-growing population. Maybe one consequence of this is that it’s relatively easy to bring inflation back under control.

Or maybe celebration is premature and we actually are on the verge of a severe recession. But so far I’m not seeing signs of it.

“I didn’t steal funds, and I certainly didn’t stash billions away,” Bankman-Fried wrote in the newsletter’s first post.

Following from there is a long, rambling explanation of how Alameda Research, SBF’s crypto-trading firm, went bust in the fall of 2022. FTX failed because it had loaned billions of dollars to Alameda, which Alameda proceeded to lose through risky trading.

SBF claims that the value of Alameda’s cryptocurrency holdings fell from more than $100 billion in late 2021 to around $11 billion just before it went bankrupt. And he says that by the end, most of the money was tied up in illiquid or “less liquid” assets that couldn’t easily be converted to cash.

Subscribe now

Over the same period, FTX’s loans to Alameda increased from roughly $3 billion to $10 billion. So when panicked customers started trying to pull their money out of FTX in early November, FTX couldn’t get customer funds back from Alameda.

I have no idea if the figures in SBF’s Substack post are accurate; SBF himself labels them as “JUST AN ESTIMATE.” But even assuming everything in the post is true, it isn’t much of a defense.

SBF stands accused of taking money from FTX customers, giving it to Alameda, and letting Alameda lose it. Going into excruciating detail about exactly when and how Alameda lost the money doesn’t really refute claims that Alameda shouldn’t have had the money in the first place.

The former CEO has been offering unconvincing defenses of his actions ever since FTX declared bankruptcy two months ago. Late last year, SBF did a series of high-profile media interviews, including with the New York Times and ABC’s Good Morning America, trying to explain what happened. When I first saw these interviews, I wondered if there might be something to his story, because he really did seem to be trying to answer the reporters’ questions honestly.

But the more I’ve seen SBF try to make his case, the less sympathetic I’ve gotten. Because the fundamental question is pretty simple: Did he allow Alameda to borrow billions of dollars from FTX without customer consent? At this point, almost everyone thinks the answer is “yes.” And while he’s been far more talkative than most criminal defendants, he hasn’t offered anyone a good reason to change their minds.

How FTX went bankrupt

Understanding exactly what happened to FTX customers’ money is a bit complicated because FTX didn’t just help customers buy and sell cryptocurrencies—it enabled them to make leveraged trades, a business that inherently involves extending credit to customers.

For example, FTX offered margin trading. This is a service in which customers might deposit $1,000, borrow another $1,000, and use the cash to buy $2,000 worth of Bitcoin. In this scenario, a customer’s $1,000 deposit served as collateral for the loan—much as a home serves as collateral for a mortgage. If the value of Bitcoin dropped, FTX would ask the customer to deposit additional collateral. If they failed to do so, FTX would sell the customer’s Bitcoins and use the proceeds to pay off the loan.

This worked as long as the value of the customer’s Bitcoins didn’t fall too quickly. In my example, if Bitcoin’s value fell by 50 percent before FTX had a chance to sell them, then the customer would have been wiped out. If they fell by more than 50 percent, the customer would have been wiped out and FTX would have taken losses.

Another FTX offering allowed customers to bet on the future price of cryptocurrencies (FTX is short for “futures exchange”). If Bitcoin’s value rose, then “shorts” would lose money and “longs” would gain. If Bitcoin’s value fell, the opposite would occur.

Once again, customers were required to post a certain amount of collateral to cover losses their positions might incur in the future. But in cases of extreme volatility, it might not be possible to close a position quickly enough to avoid big losses. If a customer’s losses exceeded their collateral, FTX itself would be on the hook for subsequent losses.

This kind of failure is a theoretical possibility on any exchange that allows leveraged trading. For example, Bloomberg’s Matt Levine has written about how the London Metal Exchange almost blew up after a nickel trader couldn’t make good on a massive short position.

This is basically what SBF says happened to FTX: FTX allowed Alameda to make huge, leveraged bets on the value of volatile cryptocurrencies. Then the value of those cryptocurrencies crashed. Not only did Alameda get wiped out, but FTX—and by extension, FTX’s customers—lost money as well.

While this story might be basically accurate, I don’t think it exonerates SBF the way he seems to think it does.

Special rules for Alameda

One of FTX’s biggest selling points has always been that it has a sophisticated “liquidation engine” that was supposed to avoid these kinds of blowups by automatically closing a customer’s position before it got underwater. In a 2019 post, FTX touted its liquidation engine as more sophisticated and conservative than those on rival cryptocurrency exchanges.

“We designed a system that we think will withstand huge market moves and huge volume,” FTX wrote in 2019.

But according to the Securities and Exchange Commission, FTX exempted Alameda from this liquidation process. The SEC says that in May 2020, “Bankman-Fried directed that Alameda be exempted from the ‘auto-liquidation’ feature of FTX’s spot margin trading services.”

This meant that if Alameda’s investments performed poorly, then FTX—and ultimately FTX’s other customers—would wind up paying the cost. It turned FTX into a giant game of “heads Alameda wins, tails other customers lose.”

The SEC also accuses SBF of telling FTX investors “that Alameda was just another platform customer with no special privileges.” If the government’s allegations are true, then this was a lie and potentially securities fraud too.

So this is the key question SBF needs to answer: Did he in fact direct FTX employees to exempt Alameda from the auto-liquidation process? And did he know that Alameda was getting special treatment when he told investors it was treated like any other customer?

SBF doesn’t even try to tackle these questions in his introductory newsletter post. Instead of addressing them head-on, he wrote more than 2,000 words on issues nobody really asked about. The fact that he has nothing to say about the main accusations against him makes me think whatever answers he might have won't actually do him any good.

Subscribe now

But a recent paper from the Roosevelt Institute says that this is nonsense. The paper is co-authored by Nobel prize-winning economist Joseph Stiglitz and Roosevelt’s Ira Regmi, and it argues that the inflation of the last two years has not been driven by fiscal or monetary stimulus. Rather, they blame the war in Ukraine and the aftershocks of the COVID-19 pandemic.

A reader emailed me to ask what I thought of this paper. In a nutshell, I didn’t find it very persuasive. It’s definitely true that some of the inflation of the last year is due to supply chain disruptions. At a minimum, the war in Ukraine undeniably pushed up energy costs.

Gasoline prices rose by 60 percent between June 2021 and the peak of gasoline prices in June 2022. With gasoline prices accounting for about four percent of the average household’s budget, this meant that gasoline alone contributed something like 2.4 percentage points of the 9.1 percent annual inflation rate for June 2022. Inflation absolutely would not have soared so high last year if the war in Ukraine had not pushed energy prices up so much.

But core inflation—a measure that excludes volatile food and energy prices—was still 5.9 percent in June 2022. That’s way above the Fed’s 2 percent target. So this wasn’t the only factor driving high inflation—or even the main one.

Pandemic-related demand shifts contributed to inflation

I think Stiglitz and Regmi were correct to identify pandemic-related shifts in demand as a significant factor in the inflation of the last two years.

For example, in 2020 people went to fewer exercise classes and bought more exercise bikes. Demand for exercise bikes outstripped available supply, so prices soared. But there wasn’t a symmetrical decline in the prices of gyms and yoga studios. So, averaging across these two categories, prices rose as a result of shifting consumption patterns.

Stiglitz and Regmi argue that dynamics like this were a major factor pushing up prices over the last two years. Another example is in real estate. In 2020, demand collapsed for commercial real estate, especially in big cities. But demand soared for housing in certain parts of the country. Again, residential landlords facing high demand tended to raise their rents, while commercial landlords facing falling demand were less likely to cut their rents. So average rents across the economy rose.

The pandemic also caused some big shifts in the pattern of spending across time. In 2020 and early 2021, a lot of people cut back spending on certain categories and saved the money instead. As the economy started to open up in the second half of 2021, they started to spend down those savings, temporarily spending at a faster rate than the pre-COVID trend.

I think that’s a fine description of what happened in 2021, but I’m not sure I’d draw the same conclusion from this as the Roosevelt authors do. The authors say that policymakers should not have worried about these inflationary pressures, since they were likely to be temporary.

But I don’t think that’s right. A big reason Democrats passed the American Rescue Plan in 2021 was concern that there wouldn’t be enough spending to help the economy recover rapidly from the 2020 COVID recession. And to be fair, I thought this was a reasonable fear at the time.

But in hindsight, it’s clear that much of this spending was unnecessary and probably counterproductive. Consumers were going to spend a lot of money on televisions and furniture in 2021 regardless of what Congress did. If Congress had passed a smaller American Rescue Plan, consumers would have spent less, shortages would have been less severe and prices would not have risen so much.

It doesn’t really matter whether the stimulus spending was the original cause of the 2021 spending binge. Putting extra money in people’s pockets made it worse.

Supply chain problems were overrated

Inflation hawks say that inflation rose in 2021 because Congress and the Fed put too much money in consumers’ pockets. This is known in economics jargon as a demand shock. Doves like Stiglitz and Regmi counter that it happened because the pandemic temporarily reduced the nation’s productive capacity, known as a supply shock.

Here I think the Roosevelt report’s claims are just not backed up by the evidence. Here is the closest they come to empirical support for this claim:

A major explanation for these price increases is the idiosyncratic supply chain factors associated with the disruptions prompted by the pandemic shutdowns and subsequent reopenings. The Federal Reserve Bank of New York’s Global Supply Chain Pressure Index (GSCPI), which integrates various indices that analyze delivery times, backlogs, and inventories to assess supply chain pressures, shows that the sustained and frequent problems are much larger in magnitude compared to historic trends.

The problem with this argument is that demand shocks create supply chain problems too. If people start ordering a lot more stuff, that’s going to create shortages and expose supply-chain bottlenecks. So the fact that supply chains have been struggling to keep up with demand tells us nothing about whether that’s because demand is abnormally strong or supply is abnormally weak.

The way to tell the difference is by looking at the quantities of stuff being produced and consumed. When inflation is caused by a supply shock, we should see economic output running below historic norms. Whereas if inflation is driven by demand, then the economy should be producing at or above historic levels. Check out this chart for example:

In the spring of 2020, spending both goods and services briefly fell below 2019 levels. But then durable goods spending soared far above pre-pandemic levels. In November 2022, the most recent month for which we have data, durable goods spending was still 16 percent above the pre-pandemic trend. Services spending also recovered from the early 2020 lows, but it remains about 3 percent below trend.

The blue line on this chart just isn’t consistent with a supply-shock explanation for inflation. Durable goods inflation happened because demand for durable goods soared and suppliers couldn’t keep up. That’s a demand shock.

The red line is superficially more consistent with the Stiglitz/Regmi thesis. The pandemic certainly forced the closure of many bars, restaurants, gyms, concert venues, and the like. But as Stiglitz and Regmi note in their paper, the pandemic also crushed demand for many of these same services. So it’s not clear how much of the reduction in services spending is due to lower demand versus lower supply.

One of the biggest components of that red services line is housing. Did the pandemic disrupt the “supply chain” for housing? In a word, no:

The pandemic’s impact on housing production is so small that it’s difficult to even see on the chart. New home completions declined by about 9 percent between February and May 2020, but then housing completions set a new record in September. The US produced more housing units in 2020 than in 2019, and then more in 2021 than in 2020. So the soaring housing prices we saw in 2021 were clearly the result of surging demand, not pandemic-related problems with the housing supply chain.

The most commonly discussed supply chain problem concerns the automotive industry, which really has produced fewer cars since 2020 than they did in the late 2010s. But even here the conventional explanation isn’t quite right.

It’s true that car companies have been hamstrung by a shortage of computer chips. But the reason isn’t that the pandemic hampered the production of computer chips generally. To the contrary, the semiconductor industry shipped a record number of chips in 2021.

The problem for automakers is that in the early months of the pandemic they panicked and canceled a bunch of chip orders. Chipmakers re-allocated this spare capacity to other customers, and so when the economy rebounded in late 2020 and early 2021, carmakers found themselves at the end of long waiting lists for new chips, precisely because demand was surging for a wide range of durable goods.

Overall, consumers bought more durable goods than ever in 2021 and early 2022. Which makes it hard to believe that the inflation of that period was caused primarily by supply-chain problems.

While I’m fairly confident that Stiglitz and Regmi’s explanation for past inflation is wrong, I don’t necessarily think they’re wrong that the Fed is tightening too rapidly. Monetary policy was clearly too loose a year ago, and we clearly needed significant interest rate hikes last year to bring inflation under control. But inflation is now coming down, and it’s possible the Fed has now raised rates enough—or even too much—to bring inflation under control.

Eric had recently earned his MBA and was working at a small consulting company in the Midwest. He wanted to change jobs, but had signed a sweeping noncompete agreement with his employer that prohibited him from working for a competing company for two years anywhere in North America.

Unfortunately, the agreement wasn’t specific about which companies were considered competitors. Eventually, Eric took a job at a software company that helped customers solve some of the same kinds of problems as his old consulting firm. He hoped that software was different enough from consulting that he wouldn’t get into legal trouble.

Subscribe now

But a few months after switching jobs, Eric got a cease-and-desist letter from his old firm.

Eric was lucky enough to get some pro bono legal advice through a family connection. The lawyer told him he’d probably win in court, but the contract was so vaguely written that it was impossible to be sure. And while the lawyer was willing to help him negotiate a settlement, Eric would have to pay hefty legal fees if the case went to trial.

For months, Eric worked at his new company with the threat of a career-destroying lawsuit hanging over his head. With a mortgage and a young daughter, he couldn’t afford to lose his job. But he also couldn’t really afford to fight the case in court.

“It was the worst time of my life,” Eric told me. “I was constantly worried.” Eric said he battled depression, lost friends, and “was not a great spouse during that time period.”

Eric ultimately reached a settlement. Its terms were confidential and prohibited him from disparaging his old employer, which is one reason I’m not using his real name in this article. He says that he wound up taking a three-month break from his new job “to make the problem go away.”

Eric’s story is far from the worst abuse of noncompete agreements I’ve heard about. Back in 2021, I wrote about three nurses who were ordered to stop practicing their profession in Wyoming. One was forced to take a new job across state lines that paid less and required a two-hour commute; she burned through her savings paying for extra child care. Another woman near retirement age was forced to take a grueling job as a traveling nurse.

None of this would have happened in California, where courts have not enforced noncompete agreements since 1872. Now the FTC wants to make the whole country work like California.

“I wholeheartedly support that,” Eric told me.

Silicon Valley’s secret weapon

Defenders of noncompetes argue that these contracts are necessary so that businesses feel comfortable investing in their workers. One of them is Alden Abbott, the FTC’s general counsel during the Trump years and now a scholar at the Mercatus Center.

“Companies could say ‘we're not going to invest in specialized training or pass on trade secrets to employees, because they might just go off and use that against us,’” Abbott told me. Abbott worries that banning noncompetes could lead to companies investing too little in developing the skills of their workers.

While that worry seems reasonable in theory, it’s hard to square with the existence of Silicon Valley. California’s refusal to enforce noncompete agreements really has promoted a culture of job-hopping in the state’s technology sector. But it does not seem like this has been a serious impediment to the development of Silicon Valley or its ability to innovate—quite the contrary.

In a famous 1999 article, law professor Ronald Gilson argued that California’s lack of noncompete enforcement was actually an important factor in Silicon Valley’s success in the late 20th Century. Drawing on earlier research by UC Berkeley’s AnnaLee Saxenian, Gilson compared Silicon Valley to the tech corridor along Route 128 outside of Boston. Until the 1970s, both of these regions were major centers for the fledgling computer industry.

But in the 1980s, Silicon Valley pulled ahead. And Gilson and Saxenian argued that Silicon Valley’s job-hopping culture was a major factor.

Because Massachusetts law allowed businesses to limit employee mobility, many engineers spent their careers at a single company. In contrast, California law gave engineers the freedom to spend a few years at one tech company before moving on to another one. This job churn was surely frustrating to Bay Area employers, but it also promoted the rapid spread of ideas from one firm to another.

California’s approach has also been a boon for startup creation. One of Silicon Valley’s first big successes was Intel, a company founded by veterans of another semiconductor pioneer called Fairchild Semiconductor. Fairchild, in turn, was founded by a group of engineers who left a company founded by industry pioneer William Shockley.

That tradition continues to the present day, with people regularly quitting jobs at Google, Facebook, or Apple to start new companies that might wind up competing with their old employers. Startups are free to poach talented engineers from the technology giants, allowing the startups to grow rapidly if they have promising ideas. 

Again, this is undoubtedly frustrating for industry incumbents who are constantly losing talented engineers—indeed, several tech giants got in legal trouble for maintaining a no-poaching agreement with one another. And some economists have theoretical models where easy job-switching harms innovation by discouraging companies from investing in new technologies or in employee training. But it’s hard to take these models very seriously given the spectacular success of Silicon Valley.

Indeed, I suspect that these models are backwards: that the rest of the country is being held back by the enforcement of noncompete agreements. If that’s right, then it would be a good thing for the U.S. economy if workers outside of California were as free as California workers to hop from job to job, taking their skills and knowledge—but not the employers’ trade secrets—with them.

The FTC faces an uphill battle in the courts

The most important question about the FTC proposal is whether it will stand up in court. Federal law prohibits “unfair methods of competition” and gives the FTC authority to enforce that prohibition. To a layman like me, it seems plausible that a noncompete agreement could be an unfair method of competition. But Abbott, the Trump-era FTC lawyer, is skeptical.

To understand this debate, it’s helpful to know a little bit of history. In the 1980s, there was an intellectual revolution that narrowed the scope of antitrust law. Scholars like Robert Bork argued that some business arrangements that might seem superficially anticompetitive can actually be beneficial in practice, and so should be allowed under antitrust law. These arguments were embraced by the Reagan Administration and the Supreme Court, and as a result they have been the law of the land for the last 30 to 40 years.

The FTC’s new rule on noncompetes is part of a broader effort by the Biden administration to roll back the Bork revolution. Biden laid out his vision for competition policy in a 2021 executive order, and he appointed the antitrust hawk Lina Khan to chair the FTC around the same time. Last November, Khan’s FTC published a new policy statement laying out its expanded vision for Section 5 of the Federal Trade Commission Act, which provides the legal foundation for Thursday’s new rule on noncompete agreements.

Defenders of the antitrust status quo don’t like the direction Biden and Khan are pushing antitrust law, and arguments over the noncompete proposal is best understood as a front in that larger war.

Opponents’ core argument is that the FTC doesn’t have a history of regulating noncompete agreements, and that it’s not appropriate for the agency to start regulating these agreements without explicit authorization from Congress. Christine Wilson, a Republican member of the FTC, wrote in her dissent that the FTC proposal was a “radical departure from hundreds of years of legal precedent.” 

Wilson also pointed to the Supreme Court’s landmark decision in West Virginia v. EPA this year, which held that Obama’s Clean Power Plan exceeded the EPA’s authority under environmental laws. The high court held that regulatory agencies needed explicit authorization from Congress before they tackled major new policy questions.

Wilson and Abbott argue that similar reasoning applies to the FTC’s new noncompete proposal. They say that the FTC does not have a history of regulating noncompete agreements, and so it needs explicit approval from Congress to do so.

But advocates counter that Congress deliberately gave the FTC broad powers to define and regulate anticompetitive conduct. The law prohibits “unfair methods of competition” without defining this phrase in any detail. Presumably Congress wanted to give the FTC some discretion to deal with new issues as they came up.

In our phone conversation, I asked Abbott about the story of the Wyoming nurses who were forced to quit their jobs and move out of the state to find work. Didn’t that seem like it could be an unfair method of competition?

Abbott didn’t think so. The nurses’ noncompete agreement “affects the nurses and harms them but it may not affect competition in the marketplace,” Abbott told me.

I pointed out that the term “noncompete agreement” literally has the phrase “noncompete” in it. That has to be a clue that it affects competition right? Abbott didn’t buy it.

“That's not the way antitrust laws define competition,” Abbott told me. “Antitrust laws talk about a relevant market, competition between two steel companies or two utility companies, or whatever.” Antitrust laws didn’t apply to the relationship between a nurse and her employer because that was a “vertical relationship,” he said.

I’ll confess that this did not make very much sense to me. But I do think there’s a good chance that Abbott’s view will prevail in court. While there is a lot of grassroots interest in more aggressive enforcement of antitrust laws, on both the left and the right, there has been little sign that the Supreme Court’s six-justice conservative majority is ready to rethink the prevailing antitrust orthodoxy.

Subscribe now

On August 2, 2011, with just days to spare, Obama and House Republicans reached a deal to raise the debt ceiling and cut spending by more than a trillion dollars over 10 years. The fiscal drama unnerved Wall Street: The rating agency Standard & Poor’s downgraded U.S. debt on August 5, and the S&P 500 fell 17 percent between July 29 and August 8.

Now another Republican majority is taking power in the House, and some Republican legislators have signaled their intent to once again use the debt ceiling to force spending cuts. Dylan Matthews at Vox has reported that a “2011 standoff looks unnervingly possible” in 2023, and some pundits are warning of the “Next Debt Ceiling Crisis.”

Breaching the debt ceiling would be disastrous for the economy. Indeed, even the possibility of a debt ceiling crisis could shake market confidence and tip the nation into a recession.

But the experts I spoke to expected 2023 to play out differently from 2011. While many Republicans still talk a big game about fiscal responsibility, the politics of the issue have shifted dramatically. 

Subscribe now

Andrew Biggs, a senior fellow at the American Enterprise Institute, told me that Republicans face a fundamental tension between their constituency and their ideology. “Republicans are historically the party of small government. But Republicans are increasingly the party of people who benefit from Social Security and Medicare.”

And these programs constitute nearly half of federal spending, so it’s hard to significantly cut the deficit without touching them.

So while Congress will certainly have some heated arguments in the coming months, especially around discretionary programs that have modest price tags but high culture war salience, don’t expect a repeat of the dramatic showdowns we saw 12 years ago. A prolonged government shutdown—to say nothing of another debt ceiling crisis—wouldn’t play to Republicans’ advantage, and Republican leaders know it.

Republicans’ two fiscal factions

When John Boehner took the speaker’s gavel in 2011, he led a caucus that broadly agreed on the need for dramatic deficit reduction. In contrast, Kevin McCarthy—or whoever takes the speaker’s gavel—will preside over a Republican caucus that’s internally divided over fiscal issues.

On the one hand are fiscal hawks who want deficit reduction to be a top Republican priority in 2023, just as it was in 2011 and 1995. 

Inflation has made this narrative especially salient, as Republicans are able to blame higher prices on Democrats’ fiscal irresponsibility. Danny Weiss, a senior fellow at the Century Foundation and former chief of staff to Speaker Nancy Pelosi, told me that so long as inflation stays high, Republicans have a strong incentive to focus on the deficit.

But there’s also a significant faction of fiscal moderates who draw their inspiration from President Donald Trump, who vowed not to cut Social Security and Medicare during the 2016 campaign and presided over a large increase in the deficit.

“This Republican class is very different than the one elected in 2010,” Ben Ritz of the Progressive Policy Institute told me. A decade ago, Tea Party Republicans viewed cutting government spending as a top priority. In contrast, he said “this class is all in on the Trump culture war.”

Shortly after the 2010 elections, House Republicans voted to ban earmarks for the upcoming Congress. In contrast, Weiss noted, House Republicans recently voted to keep earmarks, opening the door for members of Congress to sponsor pet projects in their districts.

At the same time, nearly 90 percent of Republican members of Congress have taken a “no new taxes” pledge. And even fiscal moderates often lampoon what they see as excessive spending by the Biden administration. So there isn’t a stark divide between fiscal hawks and doves in the Republican caucus—it’s more of a spectrum.

Crucially, neither McCarthy nor Senate Minority Leader Mitch McConnell seem to be fiscal firebrands. McCarthy is what Ritz called a “political politician,” meaning he’s focused on checking off narrow party priorities but has little appetite for larger structural reforms. McConnell, too, has shown little concern about the deficit; he threw his support behind the recent $1.7 trillion omnibus package.

Subscribe now

Cutting Social Security or Medicare is a nonstarter

Republican disunity is particularly evident when it comes to two of the biggest government programs—Social Security and Medicare. At least three Republican senators—John Thune of South Dakota, Rick Scott of Florida, and Ron Johnson of Wisconsin—have expressed interest in entitlement reform. And 16 House Republicans have written an alternative 2023 budget that seeks to “avert a debt crisis” in part by raising the Social Security retirement age and by reducing benefits.

But they’re opposed by other Republicans—including Sens. J.D. Vance of Ohio and Josh Hawley of Missouri—who have denounced cuts to these entitlement programs.

Meanwhile, Democrats have become more unified in their support for generous Social Security benefits. Biggs, who worked on George W. Bush’s 2005 Social Security reform attempt, pointed out that back then “you had moderate Democrats, particularly in the Senate, who were willing to fix Social Security through tax increases and benefits cuts.”

In contrast, 90 percent of House Democrats today have “co-sponsored legislation to increase Social Security benefits across the board, funded by pretty substantial tax increases,” Biggs said. “As far as I know, there are no House Democrats who have said they’re open to benefit reductions.”

So the prospects for entitlement cuts have gotten worse on both sides of the aisle over the last 10 or 20 years. And even 20 years ago, those prospects were not very good, as the failure of Bush’s reform proposal demonstrated.

“Republicans will talk about cutting Social Security and Medicare,” said Bill Gale, a budget expert at the Brookings Institution. “But they were in power for several years and didn’t do anything in that regard, so it’s mainly opposition talk.” 

Why a debt crisis is unlikely

Both Biggs and James Capretta, another AEI senior fellow, emphasized the poor optics of a Republican House threatening to provoke a debt ceiling crisis to gain leverage in budget negotiations—especially if they’re pushing for unpopular cuts to Social Security and Medicare. “It’s a completely losing strategy,” Capretta told me.

So if there is a major budget confrontation this year, it will likely be over discretionary spending, just as it was in 2011, according to Marc Goldwein of the Committee for a Responsible Federal Budget.

The Republican agenda for nonentitlement spending, though, is difficult to pin down. Ritz gave three examples of Republicans’ current fiscal priorities: Greater defense spending. More resources for border enforcement. And perhaps cutting funding for IRS enforcement. Notice that all three of these actions are likely to increase the deficit.

That’s not to say Republicans have no desire to cut discretionary spending. In fact, there are so many tiny discretionary spending programs, Biggs told me, that Republicans will have no trouble finding programs that “hit conservative hot buttons on race or sexual identity.” Weiss also predicted that formerly nonpartisan science funding may come under increasing scrutiny from Republicans, like money for the National Institutes of Health.

But these ideological fights are not very well suited to a debt-ceiling showdown. If Republicans want to argue that the deficit is such a big problem that it merits risking a financial crisis, they’ll need to propose spending cuts large enough to significantly reduce the deficit. But that’s hard to do if cuts to Social Security, Medicare, and the military are off the table.

Nondefense discretionary programs are already a small share of overall federal spending, in part because Republicans have targeted these programs for cuts in the past.

So Biggs believes that negotiations between Biden and House Republicans are likely to have a negligible impact on the deficit—and could even increase it. We saw this in recent budget cycles, Ritz said. The parties “would bid each other up,” Ritz told me.

The Bipartisan Budget Act of 2018, for example, raised discretionary spending caps by $143 billion, allowing Democrats to get higher nondefense discretionary spending in return for Republicans increasing defense spending. This year we’re more likely to see a repeat of the 2018 budget dynamics than a 2011-style standoff. As this week’s rocky start to the 118th Congress demonstrated, any “crisis” seems likely to be one Republicans inflict on themselves.

Aden Barton is an economics student at Harvard. You can follow him on Twitter.

It was the Saturday night before Halloween, and I was ferrying costumed, intoxicated young adults around our nation’s capital. Shortly before midnight, I picked up a normally dressed woman in downtown D.C., who was heading home after a large fundraising dinner at the city’s convention center.

Demand was off the charts, and my passenger told me that Lyft had charged her $59 for our six-mile trip to Arlington, Virginia. I was surprised because that was far more money than I’d earned on any of the 24 trips I’d completed so far that week.

After I dropped her off, the Lyft app revealed my portion of the base fare: $16.52. My passenger added a generous $8.96 tip, for total earnings of $25.48. That made it my most lucrative trip that week. But Lyft raked in much more: $42.48.

A lot had changed since the last time I drove for Lyft as a journalistic experiment in 2014 (and wrote about it for Vox). Back then, drivers took home a standard 80 percent of a passenger’s fare, sometimes with tips and bonuses on top. Earlier this fall, I decided it was time to embed in the driver’s seat of a Lyft once again. It turned out that in recent years, the relationship between driver compensation and passenger fares has become … complicated.

After that Saturday night trip, I started asking my passengers how much they’d paid and comparing it to what I’d earned. The results were all over the map. I had two trips where I got more than 80 percent of the total fare, including tips. I had a couple of others where I earned less than 30 percent. 

I wrapped up my Lyft driving experiment after a week, having completed 100 rides in 46 hours and earning $1,111. Lyft eventually sent me a report showing that passengers had paid a total of $2,139.73 for these rides, so on average I got just 52 percent of what my passengers paid.

Ever since then, I’ve been trying to figure out how Lyft could take such a big cut of passenger revenue and still fail to turn a profit—Lyft says that it lost almost a billion dollars in the first nine months of 2022. The more I’ve looked into it, the more I became convinced that it’s going to take major cost-cutting for Lyft to avoid eventual bankruptcy. I was also left asking one question: If after all these years, Lyft and its competitor Uber’s disruption of the old taxi model still doesn’t make business sense, what was the point?

Subscribe now

An overgrown taxi dispatch service

There’s a thought experiment I find illuminating: Imagine that Lyft announced tomorrow that it was going to shut down its app and lay off most of its staff. Then it was going to open a bunch of call centers and hire people to book taxi rides over the phone instead.

When a passenger wanted a ride, they would call a toll-free number (1-800-GET-LYFT, perhaps) and provide pickup and dropoff locations. The operator would then relay this information to drivers, perhaps in a group chat or over an old-style radio system, and one of them would claim the ride. The operator would provide the passenger with an estimated pickup time and take credit card information over the phone.

Here’s the question I want to ask about this hypothetical version of Lyft: Would it be more or less expensive to operate than the company that actually exists?

Intuitively, you’d expect the answer to be “more expensive,” right? After all, booking a ride on the Lyft app is an automated process. A single server in Lyft’s data center should be able to book hundreds if not thousands of rides in an hour, while costing less to operate than the wages of a call-center worker.

Yet there’s good reason to think that my retro version of Lyft would actually be much cheaper to operate than the high-tech version. I’ve basically described an old-fashioned taxi dispatch company—the kind that dominated the industry before Uber and Lyft came along. Back in 2016, a transportation economist named Hubert Horan crunched the numbers and found that in the pre-Lyft world, about 15 percent of passenger fares (including tips) went to dispatching and other “back office” functions. The other 85 percent went to driver compensation, fuel, and the costs of vehicle ownership and maintenance.

So if Lyft were as operationally lean as a traditional taxi company, the driver would get 85 percent of passenger fares (Lyft drivers pay to maintain and fuel their own vehicles, after all), while Lyft would be able to cover its costs with the other 15 percent. But if my rides are any indication, Lyft is taking way more than 15 percent and still losing a ton of money. It seems that the smartphone revolution didn’t make dispatching cheaper and more efficient, as you might expect. It made it way more expensive.

I’m leaning on my own experience here because Lyft is frustratingly opaque about its financials. Lyft publishes a net revenue figure with driver pay already subtracted out. As far as I can tell, Lyft doesn’t publish its gross bookings, making it impossible to calculate Lyft’s overall “take rate.” The company declined to speak to me for this story or provide data on how it splits fares with drivers.

Uber is more transparent. Last quarter Uber reported ride hailing revenues of $3.6 billion on gross bookings of $13.4 billion, for a “take rate” of 28 percent. That’s still a larger cut than traditional taxi dispatchers take, according to Horan’s math, but it’s much less than the 48 percent “take rate” I experienced as a Lyft driver.

“This is a crappy business”

Last month I argued that rising interest rates were forcing Silicon Valley to undertake a significant culture change. During the 2010s, companies like Uber and Lyft focused on rapid growth without worrying too much about costs. With interest rates at rock bottom, investors were willing to throw billions of dollars at these companies hoping that they’d be the next Amazon or Facebook—companies whose early losses eventually turned into big profits.

Now, with growth slowing and interest rates rising, Lyft is running out of time. The company needs to figure out how to reach profitability at its current scale.

Horan has been arguing for many years that this is impossible because Lyft and Uber have not significantly improved on the business model of traditional taxi dispatchers.

“This is a crappy business,” Horan told me in a recent phone interview. “The only way Uber has ever improved its profit margins is by screwing drivers.”

Like any labor-intensive industry, taxi companies face a basic tug-of-war between workers and customers. Every dollar of driver income comes out of the pockets of passengers. With no real way to raise driver productivity (as might happen in a factory), the only way to raise driver income is by charging passengers more money. But most potential riders aren’t wealthy, so if you charge more, you’ll get fewer customers.

Five years ago, Uber and Lyft aimed to reshape the taxi industry by enabling carpooling at a large scale. Services like Uber Pool and Lyft Line really could have improved the economics of the taxi business, since it would have allowed a driver to earn income from multiple passengers at the same time. But my week as a Lyft driver made me skeptical these services will amount to much: only five out of my 100 rides were shared. That could partly be an aftershock of the pandemic crushing demand for shared rides, but it doesn’t seem like a promising sign.

I think it’s overstating things to say that Uber and Lyft haven’t improved on traditional taxis. At a minimum, Uber and Lyft have dramatically improved the reliability of pickups. I’m old enough to remember calling a taxi in the pre-Lyft days and not knowing if it would show up. I think many passengers are willing to pay at least a small premium for that.

Lyft’s software also has some features that boost driver productivity:

• Because drivers use their personal vehicles, they have the option to head out on the road during periods of peak demand, like Friday and Saturday nights.

• When I was driving, Lyft frequently had a new passenger ready for pickup before I’d even dropped the previous passenger off, minimizing waiting between passengers.

• At the end of a shift, Lyft has a feature that lets drivers only accept rides that will get them closer to home. This helps to avoid “deadheading,” where the driver has to waste time and fuel driving home with an empty car.

In theory, conventional taxi dispatchers could do those last two things, but I bet the combination of large scale and precise GPS coordinates enables Lyft to do it more effectively.

But with that said, I think Horan is right that these are fairly incremental improvements to the basic taxi business model. Which leaves Uber and Lyft with the same challenging economic model as a traditional taxi company.

Subscribe now

Journey to Maryland

The longest and most interesting trip of my week as a Lyft driver came on Sunday, Oct. 30 around 10 p.m.. I picked up a young rapper named Sir E.U in D.C.’s Adams Morgan neighborhood who asked if we could listen to his music on my car’s speakers. With the music playing, we proceeded to Children’s Hospital, where I picked up a second passenger I’ll call Michael. We then drove south into Maryland, dropping off Sir E.U and then continuing to drop off Michael in Waldorf, Maryland, almost an hour’s drive south of the hospital.

During our long drive south, Micheal told me that he’d just left his one-year-old daughter with her mother so she could get surgery for an unusual growth on her neck. He would have liked to stay the night, but he needed to get back to his low-paying retail job on Monday morning.

Michael was 21 years old and told me that he’d been working as a plumber’s apprentice until his car broke down. For a while, he tried to continue his training by taking Lyft rides, but after racking up an $800 Lyft bill in one month he realized he couldn’t keep doing that. So he was stuck in a dead-end job until he could save up enough money to get his car fixed.

Michael paid $61 for his trip to Waldorf, and Sir E.U paid another $19.99, for a total of $80.99. I earned less than half of that—$38.72—for the ride.

If you’re a college-educated professional, you probably take Lyft for a night on the town or to get to the airport. You might assume most of Lyft’s passengers are like you. And I certainly did have a fair number of passengers like that.

But I also had a lot of passengers like Michael: lower-income people who didn’t have a car and relied on Lyft to get places transit couldn’t take them. Michael clearly can’t afford to take $61 taxi rides on a regular basis, but the situation left him with no good options.

It’s passengers like Michhael—even more than Lyft drivers—who would benefit if Uber and Lyft figured out how to make their dispatching service more efficient. Lyft effectively charged Michael around $30 to arrange our ride. I bet if they put their minds to it, they could figure out how to do the job for much less than that. That’s not because it actually costs $30 to connect a rider with a driver—it’s because Lyft needs billions of dollars to cover its massive overhead.

The case for frugality

There’s a wide range of attitudes toward frugality within the tech industry. On the one hand, you have a company like Google that provides its employees with free meals and every imaginable perk. Google has spent billions of dollars on “moonshot” experiments like self-driving cars with little regard for their costs. Google is able to do this because its search engine is a quasi-monopoly that throws off billions of dollars in profit every year.

Google’s workplace culture has had a big influence across Silicon Valley—including at Lyft, which like Google has long offered employees free lunches.

But not every company works like Google. At the other extreme you have a company like Amazon that builds desks out of doors to make a point about the importance of thrift. Amazon had to develop a frugal culture in its early years because it was operating in the low-margin retail industry, competing with famously tight-fisted companies like Walmart.

I think a lot of tech companies could stand to be less like Google and more like Amazon. And that’s especially true for a company like Lyft that is (like Amazon) operating in a low-margin, labor-intensive industry.

Back in April, venture capitalist Marc Andreessen wrote that “the good big companies are overstaffed by 2x. The bad big companies are overstaffed by 4x or more.”

Then last month, Elon Musk put that theory to the test. He laid off around two thirds of Twitter’s 7,500 employees, leading to numerous predictions the site would “break.”

Twitter has obviously had plenty of problems in the last six weeks, but they haven’t included large-scale outages. Twitter’s engineers seem to have been able to keep the site running with a much smaller staff.

And that makes me wonder whether something similar might work at Lyft. Last month, Lyft laid off 13 percent of its employees, leaving it with a workforce of around 5,000 people. Could Lyft make much deeper cuts while keeping the core business running? It’s hard to know from the outside. But I hope the company’s leadership gives it a try. Because dramatically lower overhead would leave more room for Lyft to offer lower passenger fares and higher driver pay—in other words, to prove that all that disruption had a purpose.

Subscribe now

Near the end of the interview, Chris asked me if our conversation had any policy implications. I said that if the government is going to spend more money supporting the parents of young children, as many liberals advocate, it would be better to do this by giving parents cash rather than spending it on subsidized child care.

Chris took the conversation in a direction I didn’t expect, describing this as “basically a conservative position.” He pointed to Republican senators like Utah’s Mike Lee, who has advocated giving parents of young children a larger child tax credit as an alternative to government-supported day care. I don’t think Chris is wrong about the way this debate played out last year. But I do think that giving parents of young children cash rather than child care subsidies makes sense in progressive terms.

Subscribe now

There’s no good reason to prefer day care over at-home parenting

This debate is somewhat analogous to one we’ve had in the K-12 education world for decades. Conservatives have proposed giving parents vouchers that let them choose where and how to educate their kids. There are two big reasons that progressives oppose this idea:

• The public has an interest in ensuring that every child gets an education so that they can fully participate in society as workers and voters. Directly providing schooling enables the government to maintain minimum quality standards.

• Public schools can be an engine for racial and economic integration. With public schools, every family in a particular community sends their kids to the same school whether they are rich or poor, black or white.

As a bit of a right-winger myself, I’ll note that both of these premises are debatable even in the K-12 education context. Public schools don’t always do a good job of educating their students, and residential segregation has led to many communities having de facto school segregation.

But the more important point is that neither one of these arguments really applies to the provision of child care for very young kids. Toddlers don’t need academic instruction, they just need a lot of attention and affection from conscientious adults. Toddlers also don’t form lasting friendships the way older kids do, so it doesn’t matter very much if they have a diverse peer group.

All of which is to say that I think a parent or grandparent can provide a one-year-old with care that’s at least as good—both for the toddler and for society as a whole—as formal, center-based child care. This is especially true because an at-home parent will often be able to provide a more favorable ratio of adults to children than a day care center.

And so given that there’s no compelling reason for the government to put its thumb on the scales in favor of institutional child care, why not let parents choose? Instead of subsidizing day care centers, give the cash to parents and let them spend it on day care if they want to—or on something else if they prefer.

Work is overrated in this context

I think one reason Chris described this as a conservative proposal is that some of the money would flow to traditional families with a breadwinning dad and a stay-at-home mom. That’s true, but the point I made back to him was that it can also promote families with a stay-at-home dad and a breadwinner mom.

But I also think it’s worth interrogating the assumption that it’s a good idea to push mothers of young children into the workforce.

Over the last 30 years, politicians on the right have looked to apply work requirements to a variety of government benefits, including TANF, Medicaid and food stamps. That’s partly to save taxpayers money, but conservatives also argue that work requirements can be good for the recipients themselves, since getting and keeping a job puts a worker on the path to self-sufficiency.

Progressives, of course, oppose these policies. They want to use a carrot—child care subsidies—rather than a stick to push women into the workplace. But the underlying rationale is actually fairly similar: that women need to work to avoid becoming financially dependent on their husbands.

I have some sympathy for both of these perspectives, but I also think they both underplay the value of parents spending time with their own children. In general, it’s good for adults to work and support themselves. But being the parents of young children is itself a form of work that has large societal benefits. A mom who spends two or three years at home caring for her young children isn’t slacking off.

A better critique for liberals to make of Senator Mike Lee’s proposal is that it’s stingy. Lee proposed expanding the child tax credit to $4,500 per child. That compares to the $3,600 offered by the 2021 American Rescue Plan or the $2,000 child tax credit that existed prior to 2021.

So Lee’s proposal added $2,500 to the existing child credit. That’s not nearly enough to defray the cost of child care in most parts of the US. But if you made the credit much bigger—say, $10,000 per child for the first three years of a child’s life—that would cover a large share of child care costs for the average family. But it would also give them the option to make other child care arrangements (a parent staying at home, leaving the child with friends or relatives) and spend the money on diapers or rent or something else the family needs more.

And while $10,000 per year might sound a lot of money, a public day care facility would likely cost even more on a per-pupil basis. So if you want to spend that kind of money helping the parents of young children, it should be up to parents, not the government, to decide the best use for the money.

Subscribe now

The last time I wrote about the inflation situation, ahead of the Fed’s September meeting, inflation was surging out of control. At that point it seemed clear that rapid rate hikes were needed to bring inflation down. That’s exactly what we’ve gotten, with four 0.75 percent rate hikes since May.

Today the economic signals are more mixed. Some data suggests that the economy is slowing rapidly and may be headed for a recession. But by other measures the US economy is still firing on all cylinders.

Below are nine charts to help you understand this complex and fast-changing picture. Some of them will be familiar from my September story. But with three additional months of data, they may tell a different story than last time.

Subscribe now

1. A good start, but not enough

Headline inflation peaked at around 9 percent in June and has come down fairly steadily since then. That suggests that the Fed’s rate hikes are working and it might be time to stop hiking soon.

However, core inflation—which excludes volatile food and energy prices—has held relatively steady at around 6 percent over the course of 2022. That suggests a less optimistic interpretation: that rate hikes have accomplished little more than preventing further escalation in the inflation rate.

Throwing out food and energy is a bit arbitrary, especially since those are two of the most significant items in many people’s budgets. So the Cleveland Fed computes median inflation, which tries to ignore volatile outliers in a more rigorous way by focusing on the CPI component in the middle of the inflation distribution. The latest reading, from October, shows inflation continuing to increase—an ominous sign for the Fed.

2. Gasoline prices have been falling since June

If I had to pick one chart to explain why inflation has fallen since June, it would be this one. Gasoline prices nationally peaked at $5 per gallon in mid-June. Not coincidentally, inflation peaked at 9.1 percent that same month.

Gasoline accounts for around four percent of the average American household’s budget, and hence it’s around four percent of the CPI. However it has outsized influence on the headline inflation rate because it is far more volatile than other CPI components. Gasoline prices shot up 60 percent between June 2021 and June 2022, which means that gasoline accounted for more than 2 percentage points of that month’s 9.1 percent inflation rate.

But since June, gas prices have been falling, dragging the headline inflation rate down. By November 2022, a gallon of gas was only about 10 percent more expensive than it had been a year earlier. This change alone explains most of the 2 percentage point decline in the overall inflation rate between June and November.

Subscribe now

3. A mixed picture on rents

This chart might be the best illustration of the dilemma the Fed faces right now. I’ve charted year-over-year changes in three different measures of rent: the “rent of primary residence” component of the CPI, plus private-sector indexes from Zillow and Apartment List.

As longtime Full Stack Economics readers know, these measures diverge because they’re measuring different things. The two private-sector data sets measure changes in “spot rents”—rents paid by tenants who sign new leases. In contrast, the BLS measures rents across the entire economy, which includes tenants whose rent was locked in by a lease they signed many months ago. For this reason, the BLS rent index tends to lag the Zillow and Apartment List indices by about 12 months.

The Zillow and Apartment List data show a rapid decline in the inflation rate for spot rents (indeed, on a month-to-month basis prices have actually been falling for a couple of months). But average rents across the economy are still rising rapidly as year-old leases get renewed at new, higher rents.

4. Hiring remains robust

An economy heading into a recession isn’t supposed to look like this. Last month, the economy created a 263,000 new jobs. That was the fewest new jobs created in 2022 (or 2021 for that matter), but it was still far above the average rate of job creation in previous economic expansions. Between February 2010 and February 2020, for example, the economy added an average 190,000 jobs per month.

That’s obviously great news for anyone looking for a job right now. But it could be a worrying sign for the Fed, since it might mean the Fed has not done enough to curb demand.

5. There are still lots of job openings

Here’s another way of illustrating the same issue. This chart compares the unemployment rate (workers looking for work divided by total workers) to the job vacancy rate (open jobs divided by total workers). As you can see, there are typically more workers across the economy than job openings. When the economy gets really strong—as it did in 1999 and 2019—then the two figures come close to parity.

By this measure, the last year has seen the tightest labor market in many decades, with almost two open jobs per unemployed worker.

6. The end of the durable goods boom

One of the most important categories driving the inflation over the last 18 months has been durable goods—products like cars, furniture, and household appliances. For many years prior to 2020, the cost of durable goods consistently fell thanks to globalization and improvements in manufacturing productivity. But in the months after the pandemic, this trend reversed as people shifted spending from services (like bars and restaurants) to durable goods for the homes in which they were suddenly spending more time.

The result was a huge increase in the cost of durable goods. At its peak in early 2022, durable goods inflation reached 18 percent. But now the durable goods boom is coming to an end. Annual durable goods inflation fell to a modest 2.4 percent in November 2022. In a few more months, we might be back to the historical pattern where we enjoy gentle durable-goods deflation year after year.

The concern for the Fed is that services inflation has been rising as durable goods inflation falls. And services account for much more of the average household’s budget than durable goods. However, the services category includes housing, and as we’ve seen the CPI’s measure of housing inflation tends to run about a year behind spot rents. So the services line may start coming down in a few more months.

Subscribe now

7. Imports are starting to fall

When ports became overloaded last year, some people thought they were suffering from new, COVID-related logistical problems. But that was mostly wrong—the main issue was just that Americans were ordering a ton of stuff from overseas. Between mid-2020 and mid-2022, monthly imports (measured by TEUs, a unit of volume) to six major ports were about 18 percent above the pre-COVID norm.

Now that import boom may be coming to an end. Imports in October 2022 were about 12 percent below August levels, and also more than 12 percent below the levels of October 2021 and October 2020. That suggests that demand for tradable goods is softening in the United States.

8. Venture capital funds are shutting their checkbooks

One group of people that is definitely cutting back their spending is venture capitalists. In 2021, US startups raised a record $325 billion in venture capital funding, according to an analysis by Earnst and Young. They started 2022 strong, with $80 billion raised in the first quarter. By the third quarter this had fallen to $38 billion, still a bit higher than the pre-COVID average but way below the 2021 peak.

This isn’t surprising—indeed, it’s a key part of how a Fed tightening campaign is supposed to work. When interest rates rise, investors become less willing to make risky investments, and so less money flows into venture capital. That, in turn, means there are fewer startups renting office space, hiring workers, and doing other things that contribute to higher inflation.

The downside, of course, is that one of the startups that doesn’t get funded could be the next Google or Tesla.

9. Housing construction is still pretty strong

Homebuilders began construction on 1.4 million new homes in October. That was a slowdown from the pace earlier in 2022, but it was still a faster pace of building than occurred in the last few years before the COVID pandemic. And even this cycle’s peak in early 2022 is far short of the all-time record building pace we achieved in 2006.

New homes are another sector of the economy that tend to be highly sensitive to interest rates, since almost everyone needs a mortgage to buy a home. So I’m surprised that home construction hasn’t declined more in the face of rising mortgage rates. The average interest rate on a 30-year mortgage rose from 3.5 percent in January 2022 to 5.5 percent in June and almost 7 percent in October. Yet housing starts declined only modestly, from 1.67 million in January to 1.58 million in June and 1.43 million in October.

That’s obviously not great, but I would have expected the doubling of mortgage rates to have a bigger impact. The big question for the Fed is whether the decline so far is the start of a longer-term trend like the housing collapse that started in 2006, or if demand for new homes is so strong that customers will continue buying despite rising mortgage rates.

Subscribe now

I’m not alone, of course. More than a million people started using ChatGPT in the first few days it was publicly available, and my Twitter timeline has been filled with people posting screenshots of their favorite conversations.

ChatGPT is remarkably versatile. It can explain obscure philosophical doctrines, debug computer code, write and revise recipes, and compose stories on any topic—and in any style—the user requests. It remembers context from one question to the next, so if its initial answer isn’t what the user is looking for, the user can ask ChatGPT, in plain English, to revise it.

ChatGPT has gotten so much attention in recent days that I’m going to assume you are familiar with the basics. If not, I encourage you to read Sam Hammond or Maxim Lott or Jon Stokes. In this post I want to speculate on how this technology might impact the American economy in the coming years and decades.

The Hype Cycle

As a reporter I’ve always tried to be early in identifying and understanding technologies that will have a big economic impact. One thing I’ve learned over the last 15 years is that it’s difficult!

Take self-driving cars as an example. I wrote my first article on the topic in 2008. At that time I was excited about the technology but thought it was still decades from commercialization.

But then the technology advanced faster than I expected, and by 2017 I—like a lot of people—though we were just a few years away from large-scale commercialization. So when I took a job at Ars Technica that year, I made self-driving cars a major part of my beat. Since then, however, progress has seemed slower. Today I’m still optimistic about the technology’s long-term prospects but I’m unsure if it will be commercially significant before the end of the decade.

This pattern is so common that the technology consulting firm Gartner developed a fun model of it called the Gartner Hype Cycle. Self-driving cars reached the Peak of Inflated Expectations around 2018, and we’re now somewhere around the Trough of Disillusionment. With luck, we’ll ascend the Slope of Enlightenment and reach the Plateau of Productivity—where self-driving cars are a normal, boring part of the economy—within a decade or two.

Generative AI systems like ChatGPT and Stable Diffusion are still on that first upward slope—a lot of people are just starting to take this technology seriously, and our minds are starting to run wild with the possibilities.

Right now this technology feels as promising as any new technology I’ve written about over the last 20 years. My early optimism about self-driving cars was tempered by the fact that it seemed to still be decades away from commercialization. In contrast, programmers are already using GitHub Copilot, a variant of OpenAI’s GPT technology, to help them write computer code.

In this respect, the release of ChatGPT feels more like Steve Jobs’s 2007 introduction of the iPhone. The iPhone’s value was immediately clear to anyone who watched his January 2007 presentation. The iPhone quickly set the standard for the smartphone industry, attracting imitators and creating a huge new market for mobile apps.

I think something similar is likely to happen with ChatGPT. Big tech companies like Google and Amazon will scramble to match its capabilities, and a bunch of new startups will be created to apply generative AI to fields like law, accounting, graphic design, public relations, tutoring, and so forth.

And then I’m not sure what will happen next. Maybe a lot of these companies will bump up against the limits of the technology and fail, just as a lot of early self-driving startups failed. Or maybe as with smartphones, the technology will work fine and everything will be up and to the right. Right now I can talk myself into either scenario.

Could ChatGPT do my job?

One way to get a handle on the possibilities is to imagine applying this technology to the industry I know best—journalism.

A few days ago I asked ChatGPT to write an article about last month’s collapse of FTX. I supplied ChatGPT with a few key facts about the collapse. The resulting article was surprisingly good. It was formatted like a news article, had perfect grammar, and pulled in relevant context I didn’t supply in my prompt. For example, ChatGPT accurately noted that Alameda Research was “a major liquidity provider” for FTX.

At the same time, ChatGPT’s piece was far from being the most interesting or insightful thing I’ve read on the topic. It didn’t break any news or offer any new insights. And that’s not surprising because ChatGPT was trained on data that only runs through the end of 2021. It literally didn’t know anything about FTX’s collapse other than what I told it.

Of course, anyone trying to make an AI-driven news site would fix that. They’d have software to scrape press releases, other news sites, blogs, and social media in real time. That could allow the AI to produce timely articles with much richer detail.

If I’m being honest, a significant share of the articles I wrote for Ars Technica—maybe 20 or 30 percent by article count—didn’t do all that much more than summarize documents that were publicly available online. For example, every quarter Tesla would report how many cars it produced during the preceding quarter, and I’d write a 400-word news article basically summarizing the announcement and providing readers with historical context. Usually I would add a few sentences of my own analysis and predictions, but I’m not confident that a suitably trained version of GPT couldn’t do that equally well—or that readers would have missed those parts of the stories if I’d left them out.

For my more ambitious Ars stories, I’d often interview experts, policy advocates, or industry insiders to try to better understand the topic I was writing about. ChatGPT can’t talk on the phone, but it could certainly do text-based chats with experts. So could a computer program do my job?

For what it’s worth, ChatGPT itself is skeptical of this idea. “Using ChatGPT to conduct interviews with sources and incorporate what it learns into a news article would likely result in a poorly written and poorly structured article that does not accurately represent the information provided by the source,” the chatbot told me.

I would like to think this is true! It’s nice to think that my best work does more than just summarize information provided to me by other people—that I synthesize the information and offer original insights in ways that a computer program couldn’t replicate. After using ChatGPT for a few days, I still think that’s probably true, at least for the current iteration of the technology. Will it still be true in five or ten years? It’s hard to say.

Even if humans maintain their monopoly on in-depth analysis, this technology could still have a big impact on the organization of newsrooms. In many digital newsrooms, writers are expected to write a mix of short, newsy articles and more in-depth pieces. Maybe in a decade they’ll delegate the shorter articles to an AI (probably fact-checked by a human reporter) and free up human reporters to do more in-depth reporting.

Rethinking the newsroom

It’s also very possible that instead of replacing human reporters, AI software could work with human reporters and make them more productive.

A fair amount of my time as a reporter is spent writing and re-writing sentences, paragraphs, and even whole sections in an effort to find exactly the right words. There’s also a fair amount of busywork looking up things like titles and affiliations, dates, stock prices, and so forth.

Generative AI has the potential to automate a lot of this work. A decade from now a reporter’s workflow might look something like this:

1. Write a bullet list of the facts that should be in an article, along with some hints about tone and structure,

2. Have a generative AI draft an article (or maybe 5 or 10 candidate articles) based on this prompt, and

3. Read through the article(s), make notes of the things you want to change, and return to step one to refine the prompt.

This is similar to the process that programmers have gone through for decades: write source code, use a compiler to turn the source code into an executable program, and then run the program to see if it works.

This is also how AI art gets created today: an artist writes a text prompt, has the AI generate a batch of images, and then sees which ones they like best and repeats the whole process.

I could see this kind of workflow dramatically improving the productivity of reporters, especially in the (fairly common) case where an article has a few sentences of new information followed by a bunch of background summarizing earlier developments in the story. The AI should already know all the background information and be able to fill it in automatically. The reporter would just have to skim the finished article to make sure the AI didn’t make any mistakes.

But why stop there? Once you have software that can automate the drafting of news articles, a lot of new opportunities open up. For example, there are many events—an election, a baseball game, a hurricane—that unfold in realtime, with new information coming in every few minutes. Traditional news organizations only generate a new story every few hours, but a generative AI could generate new stories continuously. In 10 years you might be able to visit ESPN.com at any time and read a news story about any in-progress baseball game that includes events that happened just a few seconds earlier.

An AI could also generate different stories for different readers:

• Someone who reads a lot about a topic (say the war in Ukraine) could get an article that focuses on the latest developments, while a newbie could get an explainer with a lot of background information.

• If a member of Congress comments on the war in Ukraine, the software could include her quote in the version of the story read by her constituents.

• If a reader has ties to a particular part of Ukraine, stories could provide more detail about how the war is unfolding in that area.

Obviously this is all easier said than done, and it’s hard to predict exactly how the industry will evolve. But the important point is that generative AI is going to open up a lot of new options for how to organize a newsroom. Some might involve replacing human reporters with computer programs, while others might involve using software to enhance the productivity of human reporters. Most likely we’ll get some of both, in a complex and unpredictable mixture.

How generative AI can fail

One of the most interesting responses I got from ChatGPT was this one: I asked it to give me directions from my Washington DC neighborhood to Dulles Airport in Northern Virginia. And if you just glance at the answer it seems about right, listing approximately the right streets in approximately the right order.

But if you pull up a map and actually trace out the route suggested by ChatGPT, you realize that it’s total nonsense. In step two it says to turn left when it should say right. In step three it says to turn right when it should say left. These directions would be useless for someone who actually needed to drive to the airport.

Traditional software is good at solving complex but well-defined problems like “give me directions between these two points on this map,” but hopeless at understanding subjective human concepts. ChatGPT turns this on its head: it’s surprisingly good at grasping the nuances of human speech, but it fails at straightforward computational tasks like finding the shortest route between two points on a map.

Under the hood, ChatGPT uses a fantastically complex statistical model to estimate the likelihood that sequences of words are associated with other sequences of words. This allows it to understand seemingly complex relationships like “Paul Krugman is more neoliberal than Bernie Sanders” or “Alameda Labs was a major liquidity provider for FTX.” But it leaves it helpless when trying to answer questions that require a non-trivial amount of non-verbal reasoning.

For example, just now I asked ChatGPT to give me a prime number between 5,343,532,432 and 5,353,532,432, and it cheerfully responded with 5,353,523,421. This is not a prime number! For starters, it is divisible by 3.

ChatGPT is smart enough to know that 5,353,523,421 is between the other two numbers, which honestly is pretty impressive in its own right. But ChatGPT has no idea what a prime number actually is. (Weirdly, ChatGPT does know how to generate a computer program to determine whether a number is prime. So it seems like this should be a fixable problem.)

Some critics of this technology like to put a lot of stress on this point, arguing that ChatGPT doesn’t “really” understanding what it’s saying and is merely engaging in a sophisticated form of pattern matching. That’s not wrong exactly, but I think it underplays how often human beings do the same thing. From time to time, we all have to reason about abstract concepts without fully understanding the underlying reality—for me this happens every time I take my car into the shop, for example.

The difference is that we have a sophisticated understanding of which tasks humans can do reliably and which they can’t, and human societies have developed habits and institutions to accommodate the limits of human cognition. We don’t let a 7-year-old drive a car, for example, and we don’t let someone perform surgery unless they’ve received appropriate training.

In contrast, we’re just starting to learn about the capabilities and limitations of generative AI. And as we learn more, we might realize that the limitations are much more significant than we thought at the outset.

Again, I think self-driving cars are a helpful point of comparison. It was relatively easy for computer scientists to build a car that could drive 99 percent of trips correctly. This led to a wave of initial euphoria in the mid-2010s. However, self-driving cars ultimately need to drive as safely as human beings, and that’s actually a very high bar: about one fatal crash every 100 million miles. Reaching that level of safety—and then proving that they’ve reached it—is going to be a huge challenge.

We might face similar challenges applying technology like ChatGPT to new industries. Nobody would want to use an AI tax preparer that does their taxes wrong five percent of the time. So until the software surpasses human levels of accuracy, I expect generative AI to create jobs more than it destroys them. Because you’ll still need human beings to figure out what questions to ask the AI and then check that the answers are correct.

In the short run, I think this might worsen the “average is over” problem that Tyler Cowen wrote about a decade ago. Generative AI is likely to increase the productivity of the most knowledgeable and talented people in creative industries, which could allow them to claim a larger share of the work and leave their less talented peers with worse job prospects.

It’s hard to say what might happen in the longer run. Maybe enough white-collar jobs will get automated that the college wage premium will decline and labor will shift to more hands-on jobs from welding to nursing. Or conversely, maybe there will be a bunch of new AI-related jobs and highly educated workers will continue to do much better than everyone else. Right now it’s way too early to predict.

Last February, comedian Larry David starred in a Super Bowl ad hawking the cryptocurrency exchange FTX. In a series of faux-historical scenes, David dismissed some of humanity’s greatest and most useful inventions—including the wheel (“I don’t think so”), coffee (“it’s awful,”), and the lightbulb (“it stinks”).

In the final scene, a man tells modern-day Larry David that FTX is a “safe and easy way to get into crypto.”

“Eh, I don’t think so,” David replies. “And I’m never wrong about this stuff.”

“Don’t be like Larry,” the commercial concludes. “Don’t miss out on the next big thing.”

“Larry” was vindicated last month when FTX filed for bankruptcy and admitted it had lost around $8 billion in customer deposits.

FTX’s campaign to get more people “into crypto” went far beyond Super Bowl ads. Last year, FTX bought naming rights to the stadium where the Miami Heat plays basketball for $135 million. FTX signed endorsement deals with famous athletes like Tom Brady and Steph Curry. And recently FTX CEO Sam Bankman-Fried has spent a lot of time in Washington lobbying for legislation to allow more widespread use of cryptocurrency in the US.

We now know that FTX was a bad way to “get into crypto,” but the larger question is whether ordinary Americans should be getting into crypto at all. Over the last decade, policymakers in the US and much of the world have taken a hands-off approach to cryptocurrency to avoid strangling a supposedly promising technology in its cradle. But cryptocurrency isn’t so new any more, and it doesn’t seem to be living up to the early hype.

In recent years a lot of people have “gotten into crypto” as a speculative investment, as a way to buy illicit drugs or to make ransomware payments. But we’re still waiting for truly mainstream uses like purchasing a cup of coffee.

And since that FTX Super Bowl ad aired, a lot of ordinary people have lost their investments in ways that US financial regulations are designed to prevent. Maybe it’s time for policymakers to stop treating the crypto industry with kid gloves and start rigorously enforcing those laws. If that means a lot fewer people using or investing in cryptocurrency, that wouldn’t necessarily be a bad thing.

Outlaw technology

One of the first times bitcoin was mentioned in a mainstream news story was in June 2011, when reporters started writing about the bitcoin-based drug marketplace Silk Road. The site’s existence drove Sen. Chuck Schumer (D-NY) nuts.

“Literally, it allows buyers and users to sell illegal drugs online, including heroin, cocaine, and meth, and users do sell by hiding their identities through a program that makes them virtually untraceable," Schumer said. He called the Silk Road “the most brazen attempt to peddle drugs online that we have ever seen,” and described bitcoin as “an online form of money laundering used to disguise the source of money.”

I started writing about bitcoin around this time, and back then there seemed to be a real possibility the US government might try to shut down the bitcoin network. US law requires payment networks to follow “know your customer” rules and report suspected money laundering to the authorities. The bitcoin network not only doesn’t do this — it’s designed to make compliance impossible.

In May 2013, the feds shut down a digital payment network called Liberty Reserve that the government said was heavily used for money laundering. As I pointed out at the time, the US government’s arguments against Liberty Reserve could easily be used against Bitcoin. Indeed, I believe this is a big reason Satoshi Nakamoto chose to create Bitcoin using a pseudonym: If he had identified himself, he might have wound up in prison like Liberty Reserve founder Arthur Budovsky.

But early bitcoin advocates convinced US officials that it was better to take a hands-off posture. The network’s decentralized structure meant that it would be almost impossible to shut it down. Stopping US-based bitcoin activity would simply push the network overseas — potentially beyond the reach of US subpoenas.

These arguments persuaded the Obama administration, which signaled in late 2013 that no crackdown would be forthcoming. At a November Congressional hearing, a Justice Department official praised bitcoin’s “many legitimate uses.”

This had a big impact on the bitcoin ecosystem. Venture capitalists poured hundreds of millions of dollars into bitcoin-related startups in 2014, and prominent venture capitalists started to tout the technology.

In 2014, a little-known 20-year-old named Vitalik Buterin launched the Ethereum network with a crowdsale of its cryptocurrency, ether. People copied his approach, and by 2017 there was a craze for “initial coin offerings,” with hundreds of new coins being offered for sale.

Once again, there was a plausible case that most of this activity was illegal. US law requires anyone selling stocks, bonds, or other securities to register with the Securities and Exchange Commission (SEC) and follow complex rules designed to protect customers against fraud. A lot of these crowdsales seemed to be covered by US securities laws, yet very few of them followed the SEC’s rules.

But for the first year or two, the SEC mostly looked the other way, only prosecuting a few of the most egregious frauds. When Biden’s SEC chairman, Gary Gensler, took office in 2021, he had a mess on his hands. His predecessor’s hands-off posture had enabled the development of a sprawling industry that was largely out of compliance with SEC rules and — perhaps not coincidentally — also rife with scams, conflicts of interest, and other problems. Gensler started pressing for stricter enforcement of securities laws, but he faced pushback from cryptocurrency companies, their backers in the venture capital world, and their allies in Congress.1

Launch first, change the law later

Cryptocurrency isn’t the first new technology to fit awkwardly into existing legal frameworks. It’s interesting to compare it to the ride-sharing model that Lyft pioneered—and Uber quickly copied—in the early 2010s.

A decade ago, most big cities had taxi regulations that didn’t allow someone to just start offering people rides in their personal vehicle. Rather than getting cities to change their laws, Uber and Lyft largely ignored the laws and launched their services anyway. They were gambling that they could quickly build a large customer base that would then help them lobby for changes in the law.

Uber and Lyft argued the rules were outdated and propping up a corrupt taxi monopoly. It helped that there was some truth to these critiques, but the main reason their lobbying succeeded was that their services really were better. Customers came to depend on the convenience of hailing a car with their smartphones, and that created a strong constituency to let them keep operating.

Over the last decade, the cryptocurrency industry has been pursuing a similar strategy. They’ve warned that strict enforcement of existing laws could strangle a promising technology in its cradle. And they’ve built a community of bitcoin users and investors who have lobbied for pro-bitcoin legislation.

But I think that — as with ride sharing — the success of this lobbying will depend on whether cryptocurrency actually proves useful. And while the utility of Uber and Lyft was obvious, the value of blockchain technologies is not. The longer things continue this way, the harder it will be to argue that governments need to give crypto more time to prove its value.

Waiting for cryptocurrency’s killer app

Early advocates thought bitcoin would become a mainstream payment system that people would use for everyday purchases. And they also expected it to open up a lot of other new possibilities.

In a 2014 interview, for example, venture capitalist Marc Andreessen imagined a scenario where someone drives their car into a parking garage and then the car uses bitcoin to automatically reserve and pay for a parking space. He also talked about the potential to use bitcoin-based micropayments to prevent spam by charging a fraction of a penny for each email. If ideas like this had panned out (and at the time I thought they might) the early hype would have been justified.

But over the last decade, venture capitalists have poured billions of dollars into cryptocurrency startups exploring ideas like this:

• There were startups to help businesses accept bitcoin payments.

• There were plans to have large companies use blockchains to track their supply chains.

• People tried to use blockchains to build apps that did a better job of protecting user privacy.

• There were plans to use smart contracts to build decentralized autonomous organizations.

Some of these efforts are still around in some form, but after a decade of experimentation, we seem to be no closer to building blockchain-based applications that are useful to ordinary Americans. The applications that have thrived have mostly been tools that let people speculate on the value of tokens. That certainly describes the two most recent cryptocurrency crazes, non-fungible tokens and decentralized finance (DeFi).

DeFi uses blockchains to enable people to directly borrow, lend, and trade cryptocurrencies. While the technology is fun for finance nerds to think about, it doesn’t really work as finance. The conventional financial industry funds the construction of homes, office buildings, factories, and so forth. The decentralized finance industry only seems to “finance” gambling on cryptocurrencies.

And there’s a good reason for that: a core feature of most cryptocurrencies is that transactions are irreversible. If a hacker steals your bitcoins, you generally won’t have any recourse because no one has authority to reverse transactions. But nobody would want the real estate system or the stock market to work like this, because it would mean that hackers could steal your house or your retirement savings.

A technology for freedom

Ultimately, I think the most valuable application for blockchains is the one that hard-core bitcoin advocates have focused on since the beginning: freedom from government interference.

This isn’t that valuable in a developed country like the United States. We’re blessed to have a stable currency, a functional financial system, and a government that largely respects individual freedom.

But it’s not hard to see how bitcoin could be attractive in countries without these advantages. If your country suffers from hyperinflation, then bitcoin’s independence from government control might be quite valuable. Usage of bitcoin may be growing in Argentina for this reason. It’s not hard to see how bitcoin could be useful to dissidents in authoritarian countries.

But for bitcoin to be a useful technology for resisting government overreach, it’s important that people have custody over their own bitcoins rather than relying on intermediaries. When I first invested in bitcoin back in 2012, the existing bitcoin exchanges seemed shady so I printed out my private keys (strings of letters and numbers that control access to bitcoins) on a piece of paper and put them in my filing cabinet. Today there are “hardware wallets” that make this easier than it was a decade ago, but it still requires a certain level of technical sophistication to do well, and there’s still a real risk that your coins could be lost or stolen.

This created a market opportunity for intermediaries like FTX. FTX told ordinary users that its platform provided users with all the benefits of cryptocurrencies while being “safe and easy.” Obviously, that was wrong as far as FTX was concerned, but bitcoin purists argue that it’s wrong in a broader sense too: that the whole point of bitcoin is to avoid relying on intermediaries like FTX.

If you buy this argument, one implication is that we shouldn't worry too much about excessive government regulation of the cryptocurrency industry. The core bitcoin network is robust — so robust that the US government probably couldn’t shut it down if it tried to. And the most valuable applications of cryptocurrency don’t actually require an extensive cryptocurrency industry, since they involve people taking direct ownership of their own coins.

If US regulators had been enforcing existing US law in a rigorous fashion over the last decade, much of the contemporary crypto industry wouldn’t exist. Many — perhaps even most — cryptocurrencies are probably securities under US law, which means that they shouldn’t have been sold to non-wealthy Americans without first undergoing extensive due diligence.

There are also serious questions about whether leading crypto exchanges are following applicable US regulations pertaining to money laundering and the safekeeping of customer assets. And there are long-standing questions about the solvency of “stablecoins” like Tether.

So far, the SEC and other regulatory agencies have proceeded gingerly in this area. One reason is that the cryptocurrency industry has cultivated powerful allies in Congress. Back in March, eight members of Congress sent a letter to the SEC basically complaining that the agency was investigating cryptocurrency companies too aggressively.

The collapse of FTX is likely to change the politics of this issue and focus the minds of regulators. Perhaps now they’ll make a serious effort to shut down the many shady companies that operate at the margins of US law. 

And hopefully that will mean a lot fewer celebrity endorsements and and a lot fewer Super Bowl ads encouraging ordinary Americans to invest in cryptocurrency. Because most Americans shouldn’t invest in cryptocurrency—both because it’s unlikely to be a good investment and because most Americans lack the technical skills required to do it safely.

Thanks to Josh for asking me to write this piece. Please check out his excellent Substack, Very Serious.

At one point in the interview, Bloomberg’s Matt Levine asked SBF to explain yield farming, an investment strategy that enables cryptocurrency owners to generate earnings using their digital tokens. When SBF finished his explanation—which I’ll describe in more detail shortly—Levine and the podcast’s co-hosts seemed stunned.

“I think of myself as like a fairly cynical person, and that was so much more cynical than how I would have described farming,” Levine said. “You're just like well I'm in the Ponzi business, and it's pretty good.”

SBF responded that this was a “pretty reasonable response” with “a depressing amount of validity.”

SBF now faces accusations that he ran a literal Ponzi scheme, so many people have understandably interpreted this as effectively a confession of his own guilt.

But SBF was actually making a broader point about the speculative dynamics that drove last year’s cryptocurrency boom. And the dynamic he identified didn’t just happen in crypto—something similar was happening in the broader world of venture-backed technology startups during the late 2010s.

The Box and the X Token

Yield farmers deposit cryptocurrencies with decentralized finance platforms and earn returns in much the same way that you can deposit dollars in a conventional savings account and earn interest. In 2021, yield farmers were earning a lot more than you could get from conventional financial products, which made some people wonder if something shady was going on. In that May interview, Levine asked SBF to explain how yield farming worked.

In response, SBF sketched out a “toy model” of yield farming involving a box that could hold digital assets. “For now ignore what it does,” SBF told his Bloomberg hosts, “or pretend it does literally nothing” aside from storing people’s tokens.

This hypothetical box was associated with an equally hypothetical X Token, SBF said. X Token holders get to vote on changes to the box, and if the box ever turns a profit, X Token holders would get a share. Every day, the box creates a bunch of new X Tokens and gives them to anyone who has stored money inside the box.

In SBF’s telling, this leads to a speculative flywheel effect: people put more money in the box to get more X tokens. The flood of cash into the box makes people more optimistic about the box and its associated token, so the value of the X Token rises. As X Tokens get more valuable, people put still more money in the box and the cycle continues.

While SBF’s hypothetical box does literally nothing, that’s not quite true of the real boxes SBF was trying to explain. Certainly the creator of a box doesn’t portray it that way. According to SBF, a box creator will “probably dress it up to look like a life-changing, world-altering protocol that's going to replace all the big banks in 38 days or whatever.”

SBF said he was specifically thinking about crypto platforms like Uniswap, Aave, and Compound that allow people to trade or borrow cryptocurrencies. Let’s take Uniswap as an example. Uniswap is a decentralized exchange that lets people trade one cryptocurrency for another. Rather than using a conventional company as an intermediary, Uniswap runs its code automatically on the Ethereum blockchain.

Anyone can lend tokens to a Uniswap pool to provide greater liquidity and more attractive prices for traders. People who do this are rewarded with UNI, tokens that work a bit like shares of stock in Uniswap. UNI holders have the right to vote on proposals to change how Uniswap works. In SBF’s analogy, Uniswap is the box and UNI tokens are the X tokens. 

Uniswap doesn’t literally do nothing—lots of people use it to trade one token for another. But SBF’s point was that there are lots of platforms out there that employ this kind of token reward scheme. And in many cases the crypto industry wound up with a tail-wagging-the-dog situation where people were participating in new platforms mainly to earn tokens, without caring very much about the underlying functionality or value. The result was a speculative bubble that has been deflating over the last few months.

Yield farming in the real world

As I was pondering this analogy, I realized that I experienced something like it way back in 2014. That year I spent a week driving for Lyft and wrote about the experience for Vox.

Over 50 hours of work I completed 54 rides and earned $594 in fares. At the time, Lyft had a promotion guaranteeing new drivers at least $1,500 for their first week of driving. So Lyft kicked in another $906.

This was obviously not a sustainable business model. But at the time Lyft’s priority was growth, not profits. And Lyft had recently raised $250 million, so it could afford to spend lavishly to attract new drivers.

SBF’s box analogy fits this situation pretty well: Lyft accelerated its growth by offering subsidies to drivers in much the same way that Uniswap accelerated growth by offering subsidies to liquidity providers. In both cases, the subsidies were financed by people who were making speculative bets on the future of the platform. For Lyft that was venture capitalists who held Lyft shares. For Uniswap it was people who held UNI tokens.

An investment like this isn’t necessarily irrational. Lyft investors were betting that there were powerful network effects in the taxi business—that the more drivers on the road, the more valuable the network will be for passengers, and vice versa. Today Lyft is still struggling to turn a profit, but I think it will ultimately prove to be a viable business—albeit not as profitable as early investors hoped.

Other tech platforms pursued similar business models during the late 2010s. One of the most hilarious examples was Doordash, which sometimes offered meals to people below cost. In 2019, one pizza shop owner discovered that Doordash had started offering a delivery service for his restaurant without his permission and was offering a $24 pizza for $16. His friend realized that this was an arbitrage opportunity.

If someone could pay Doordash $16 a pizza, and Doordash would pay his restaurant $24 a pizza, then he should clearly just order pizzas himself via Doordash, all day long. You'd net a clean $8 profit per pizza.

Of course, the pizzas themselves weren’t free—the ingredients cost around $7 per pizza—so the profit per pizza was initially small. But the pizza shop owner eventually figured out that he could just throw uncooked pizza dough into the box, allowing him to earn something like $7 per order.

This is the brick-and-mortar equivalent of yield farming. Lyft and DoorDash were paying people above-market rates (or offering them below-market pizzas) to incentivize more people to join their platforms. And as more people joined their platforms, the value of their stock went up, which allowed them to raise more money and pay more subsidies.

As in a Ponzi scheme, this kind of growth seems unsustainable. But there’s also an important difference: Ponzi schemes involve stealing the money of unsuspecting customers. In contrast, DoorDash and Lyft were perfectly transparent about what they were doing. The companies’ investors willingly dumped billions of dollars into these companies in hopes of earning longer-term profits.

By the same token, the phrase “Ponzi scheme” doesn’t fit many of the speculative cryptocurrency investments SBF was describing with his box analogy. People who chose to buy and hold UNI tokens weren’t being defrauded. They were making a speculative bet on the future of the Uniswap platform in much the same way Lyft investors were betting on Lyft.

With that said, I do think there’s a potentially significant difference: taxi rides and restaurant meals have clear utility to ordinary people. In contrast, platforms like Uniswap seem to mostly facilitate cryptocurrency speculation, a zero-sum activity that produces little to no value for the broader society. It’s hard to see how it can support a thriving cryptocurrency industry over the long run.

To be fair, there is lots of speculation in conventional financial markets too. But conventional financial markets also fund the creation of new businesses, new homes, new factories, and so forth. As far as I can tell, crypto’s “decentralized finance” industry doesn’t do anything like that.

I’ll be thinking about this question more in the next couple of days as I write a longer piece about the state of cryptocurrency regulation. I’d love to know what you think. What do you see as the most significant applications of cryptocurrency? What would we lose if cryptocurrency networks shut down?

Twitter has been in constant turmoil over the last three weeks, as Elon Musk has announced a flurry of layoffs and new products—then quickly suspended some new products that didn’t work out as intended. Rather than write a conventional article with a clear thesis, I thought it might be better to offer some loosely-connected thoughts about how things have gone so far.

1. Let’s get this out of the way: so far, Elon Musk seems to be doing a worse job running Twitter than I expected. I expected things to be chaotic, but I didn’t expect them to get this chaotic this fast.

2. I don’t like the way Musk has treated Twitter’s employees. Layoffs are always painful, but Musk seems to be going out of his way to do it in a haphazard and disrespectful way. This must be an immensely stressful time for Twitter employees—especially those who are supporting families or have immigration visas tied to their jobs.

3. It would be nice to say that Musk’s efforts are doomed to failure because this kind of callous management style never works. And if I had to guess, I’d say it probably isn’t going to work. But I’m not as confident as a lot of his critics.

4. Musk is clearly trying to make Twitter more like those at SpaceX and Tesla, where employees routinely work punishingly long hours. You might think that was unsustainable, but each of these companies has been doing incredible work for close to 20 years now.

5. We’ve known for almost 50 years that programming productivity doesn’t scale linearly with the number of programmers. A major reason is that programmers have to coordinate with one another, and the cost of this grows more-than-linearly with the number of programmers.

6. So I don’t think it’s that crazy to think that a 2,000-employee Twitter with a SpaceX-like work culture could be more productive than a 7,500-employee Twitter with a conventional big company culture.

7. The big question is whether Musk can convince enough people to sign up for this. Reporting so far suggests that a ton of engineers are likely to quit beyond the 3,000-some people Musk fired earlier this month.

8. SpaceX and Tesla are dream jobs for a lot of engineers. Lots of kids grow up dreaming to build rockets, and Tesla is plausibly making a big difference in combatting climate change. It’s not clear that Musk’s Twitter will have that same kind of draw.

9. At the same time, working for one of the world’s leading social media companies will always be more prestigious than the average programming job. I also think liberals might be underestimating the number of engineers who like Elon Musk, broadly share his politics, and might like the idea of working at a tech company that’s not on the woke bandwagon.

10. There may also be some young, ambitious engineers who see this as a career opportunity. Lots of people in their 20s put in long, punishing hours at law firms or investment banks that most people have never heard of. With a lot of senior people quitting, some junior people might be excited at the opportunity to move up through the ranks quickly.

11. On the other hand, there’s a big difference between building a workaholic company from the start—as Musk did with SpaceX and Tesla—and trying to force a big change on an existing workforce. Firing 50 to 75 percent of workers is bound to create morale problems that could deep schisms inside Twitter’s workforce and make it hard for Twitter to recruit new engineers for years to come.

12. In other areas, I think Musk’s tolerance for risk might be an under-appreciated advantage. For example, earlier this week Musk announced he was turning off a bunch of microservices because he claimed most of them weren’t needed for Twitter to work.

13. This was widely mocked online, especially after Twitter’s two-factor authentication stopped working, making it impossible for many people to log into the site. But the problems seem to have been fixed within a few hours, with no lasting damage to Twitter.

14. Unlike a conventional CEO, Musk doesn’t have to worry that his board will fire him if he screws up. So he can try a lot of stuff and then quickly reverse decisions that don’t work out. This allows him to learn rapidly.

15. This might also be the right way to think about last week’s pay-for-verification fiasco. A bunch of people paid $8 for verified accounts purporting to be famous brands and tweeted out funny parody tweets, forcing Twitter to stop selling blue checkmarks.

16. While this has been widely portrayed as a huge blunder, it’s not obvious that this did any permanent damage to Twitter. Obviously it’s not going to help Twitter sell ads over the next month or two. But if the site is working well a year from now, I have to imagine most of these brands will come back.

17. If anything, those parody tweets produced a round of funny stories that kept Twitter in the news.

18. This basic approach—iterating rapidly without worrying too much about failures—has been a hallmark of Musk’s leadership at SpaceX. For a while in late 2020 and early 2021, SpaceX was launching a prototype of its new Starship rocket almost every month. Most of those launches failed, but SpaceX engineers learned a lot from the failures. Maybe the same approach will work for Twitter.

19. A lot of people have mocked Musk’s idea of charging people $8 per month to get a blue “verified” checkmark. And I agree that it’s silly to expect people to pay for a blue checkmark solely as a status symbol.

20. But Musk has made it clear that that’s not all he’s doing. Musk has said that verified users will have priority in replies, mentions, and search, which will provide another incentive for people to pay.

21. In the long run, the way things might work is that anyone can sign up for a free account if they just want to read other people’s tweets and occasionally reply to tweets by their friends. But anyone who wants to build up popular Twitter profile will need to pay $8 per month to become verified, because those without a blue checkmark will be almost invisible to people who don’t already follow them.

22. Obviously, Musk isn’t going to get literally all of Twitter’s users to pay. But if he can get, say, 10 percent of users to pay, that would represent $2 billion in additional annual revenue.

23. Josh Barro has an interesting (paywalled) post suggesting that Musk could monetize Twitter by turning itself into an OnlyFans competitor. In Josh’s telling, the blue checkmark would function not only as identity verification but also age verification.

24. It can sometimes be useful to have polite fictions about what people use a product for. If Twitter formally announced a porn subscription service, many men might be reluctant to sign up for fear of angering their wives. It would also cause friction with Apple’s app store gatekeepers.

25. But if you charge people for blue checkmarks and then later announce that you’re protecting children by limiting porn to verified adults, that has the same practical effect while not ruffling any feathers.

It’s time to do a paid-subscriber-only Q&A post! Please leave questions for me in the comments to this post, and I’ll answer some of them in a future post. Thank you for supporting the newsletter!