Paradoxes of Productivity Growth
I'm a big believer in the idea that maximizing productivity growth—the change in economic output that can't be explained by throwing more money or more physical equipment at the problem of producing goods and services that satisfy human desires—is the most important thing economic policy can focus on. This is not original to me, of course; The Great Stagnation made this case, and it builds on observations and mental models from others. We get our working definition of productivity growth (total factor productivity) from the Solow growth model, which basically estimates growth from labor and capital, and then has a residual because those don't perfectly capture where growth comes from.
Productivity growth is some combination of literal technology and social technology. The former is pretty easy to understand: physical technologies from wheels and pulleys to RFID and EUV allow us to get more results from a given amount of effort. Social technologies cover a wide range of other behaviors that can affect economic outcomes, from high-level ones like trustworthiness and punctuality to more granular ideas like accrual accounting, performance-based compensation, post-mortem memos, and the like. Other elements include general attitudes: at the level of companies and countries, having people in charge who think that the institution they're responsible for will last a long time, but could fail if they make a mistake, will tend to produce better results than the vague hope of retiring before things go off the rails. The concept of productivity growth itself is an instance of productivity growth: just by having a new mental model, you can slice up your statistics on economic growth to figure out how much of it to attribute to the gradual accumulation of buildings, equipment, roads, ships, etc., to general population growth, and to the sometimes-mysterious extra factor that makes the sum of these equal more than their parts.
In an economy without productivity growth, the high-level policy decision is how to allocate output between building up more capital and enjoying the results of what we have. This is a fairly narrow, boring problem, that is either a) a question of estimating long-term interest rates, or b) deciding how much to prioritize the wellbeing of future people relative to the present. Since productivity growth means doing something new and different (or scaling something that works), it leads to more interesting decisions and tradeoffs.
One of those tradeoffs is the paradoxical impact that productivity growth has on inequality:
- Productivity growth increases income and wealth inequality: it affects different industries in different ways; Moore's Law had a big impact on the value of programmers but a small one on the productivity of dentists and accountants. The big productivity-improving technologies, like electricity, internal combustion, and transistors, do eventually have an impact on just about everything. But the immediate gains tend to be captured by people who either master these technologies early or bet on them heavily. And the income impact is dwarfed by the wealth impact: someone who owns an equity stake in a company that's increasing global productivity growth and capturing appropriate profits is making a lot of money, and that money tends to be capitalized at a high multiple.
- On the other hand, productivity growth tends to decrease consumption inequality for a given level of income inequality. In fact, when a country goes through a productivity growth cycle, the people who were rich beforehand will sometimes see their standard of living fall; they could already afford to import cars and gadgets that were luxuries in their home country, but it's hard to maintain the same staff of domestic servants if GDP per capita goes from $2k to $10k. Meanwhile, home appliances, electronics, and media that used to be luxuries become widely available. The deployment of general-purpose technologies is deflationary in the long term, because it means meeting existing needs with less labor and materials.
- On still another hand, the deployment phase for these technologies tends to be inflationary, and to create plenty of jobs, and not just for experts. The mid-century productivity boom, built from the deployment of electrification in factories and homes as well as the rise of the automobile, created lots of jobs, and not just for engineers who understood electric motors and batteries. It meant more work for all of the industries adjacent to the main growth industry, and also subsidized the spread of middle-class norms ($, The Diff).
Of these effects, the first one is the most obvious and tends to dominate discussions about the impact of AI and other new technologies. There's a lot they can do, and a lot of it was done by people who were earning respectable but not amazing salaries. If all of those jobs get replaced by API calls, it's obvious that some people will get very rich and many more will lose their livelihoods.
But in general, that hasn't been the story of technological change: in the very long run, employment rates have been surprisingly stable everywhere we're able to measure them—it's surprising and counterintuitive that over the last two centuries of extreme technological, institutional, and cultural change, roughly 90-95% of people who want a job can find one most of the time. In many countries automation has increased formal employment when the alternatives were either agricultural work, working in the informal sector, or rent-seeking. And in a sense that's even true of rich countries' industrialization; the US was much more corrupt historically, and certainly the late 19th century and early 20th centuries had some egregious abuses of government power for private gain, but after a while it started to get obvious to most people that there was simply more money in positive-sum activities than in negative-sum ones—and that tolerating the negative-sum behaviors was a drag on overall growth.
That actually extended to politics, too; the relationship between labor and capital is an easier one to navigate when the question is "how fast do each of us get rich?" rather than "how can I protect my piece of the shrinking pie?"
The story of economic growth is usually a story of overlapping S-curves in adoption, and the first derivative of that S-curve, which measures the new deployment of a technology, tends to peak and decline. As long as something else is taking off when another technology peaks, that's tolerable; it doesn't avoid recessions, but a recession also forces people to leave declining sectors and join growing ones instead.
So the case for macroeconomic optimism about AI actually rests on three potential foundations:
- AI might completely rewrite the rules, and lead to a point from which it's pointless to extrapolate from economic history except in the most handwavy ways. That applies to the negative scenarios where the robots kill us all, but also to positive scenarios (where the one useful extrapolation you can make is that the global economy has, several times, suddenly and seemingly permanently switched to a growth rate unimaginably faster than what had come before).
- AI could have a Baumol-style effect where knowledge workers become so much more productive on average that wages for all the non-AI-exposed jobs go up. This will be great for people who work with their hands, and terrible for people who like labor-intensive services like restaurants or healthcare (though every other general-purpose technology has made some previously non-scalable parts of industries scalable).
- AI could accelerate fundamental research in other areas, hopefully identifying other general-purpose technologies. If research is held back by the burden of knowledge, then better tools for ingesting and summarizing information, and for interactive tutoring, can help. On the other hand, the last big advance in information-processing technology has yet to trigger an industrial renaissance, at least in part because if you add "software" to the job title of an "engineer" you can triple their compensation.
It's always important to look at historical precedents when evaluating a new technology, but, frustratingly, the higher the stakes are the smaller the sample size gets. And many of the general patterns are most applicable in retrospect. Overall, productivity growth is good, even if it's not good for everyone, and compared to previous productivity booms, states are larger and redistribution is more popular. These states are also more calcified, and more opaque, though; loopholes and regulatory capture are a bigger risk. Of course, no one looking at version 1.0 of the internal combustion engine, the light bulb, or the transistor truly understood their implications. It's fortunate for us that they plunged ahead anyway.
On the other hand, the "productivity-versus-everything-else" metric is not always so clear-cut. Aspirin and air conditioning both increase output, but have a smaller effect on output per hour. AC reduces the need for siestas or long summer vacations, and aspirin also increases available work hours. But this doesn't show up in productivity statistics, in part because the productivity formula doesn't have some way to track the difference between leisure time and being in no condition to work. ↩︎
It's natural to assume a bias here: doesn't everyone have a selfish incentive to live a good life now and let the physical capital that sustains this rust away? But there's a bias in the other direction. We, the people who inherited the world of cheap and ubiquitous computing and information that makes this newsletter possible, are obviously descended from people who made sacrifices in their present in order to improve the standard of living for future people who would be unimaginably wealthier than they were. The true longtermism is to apply a "risk premium" to the utility calculation of future generations, since their priorities will be very different from ours, but to apply a lower discount to the views of past generations, since we can better judge their priorities. And anyway, we owe them one. ↩︎
In a modern Western context, this sounds like a weird thing to worry about. The US has more private planes than butlers. But historically, it was pretty common for upper-middle-class families to have live-in help. Agatha Christie wrote this about her life just over a century ago: "Looking back, it seems to me extraordinary that we should have contemplated having both a nurse and a servant... But they were considered essentials of life in those days, and were the last things we would have thought of dispensing with. To have committed the extravagance of a car, for instance, would never have entered our minds. Only the rich had cars." Friends who work in lucrative industries in poor countries, or in richer countries with temporary visas for such workers, report that this is still common in some parts of the world. In fact, this gives us yet another potential contributor to the Great Stagnation: if Christie was writing at the tail end of the period where a middle-class family could have multiple servants, she was also writing at the end of a period where their children could have one-on-one tutoring and the attendant ~two standard deviation improvement in educational outcomes. There are promising efforts to scale this sort of outcome while working around the constraints of high labor costs. ↩︎
A Word From Our Sponsors
The financial world is messy and so are the unique processes built to handle it. Documents are unstructured, processes are manual, and decision-making is needed at every step.
Started out of the Stanford AI lab, Tennr’s automations platform enables you to quickly turn those painful, document-heavy processes into straightforward automations.
Firms, funds and, fintechs are building custom AI models on Tennr to:
- Write risk reports, DDQs, and market commentaries.
- Read and organize quarterly statements, cap calls, and reports.
- Handle client onboarding and complex inbound inquiries.
- And building bespoke research pipelines and custom investment models.
Want to learn more? Diff readers like yourself can grab some time here to discover what processes you could be automating.
Some FTX customers and ex-creditors are speculating about restarting the exchange. An operating business is typically worth more than that business's assets would be in liquidation, and there's a sort of 4-dimensional-chess argument that New FTX would be more trustworthy than other offshore exchanges because it would face such relentless regulatory scrutiny.
But one of the questions that needs to get answered for this plan is what FTX's true economics looked like, and in particular the risk that they were distorted by FTX's relationship with Alameda. Normally, if an exchange were started by a company that planned to trade on it, as well as on other exchanges, the worry would be that the traders would get special privileges. But FTX was more venture-fundable than Alameda, which created the possibility that Alameda could run deliberately flawed trading strategies in order to attract volume from sophisticated traders, and could use that volume to raise money for FTX from VCs, who would be looking at accurate volume data but wouldn't recognize how much of it came from a related party running deliberately bad trades. (The ideal strategy for this would be something high-frequency with a low negative expected value for each trade. Just running a market-making strategy with deliberate lag would do nicely.) The hard part about actually pulling this off is somehow getting the money intended for FTX back to Alameda so it doesn't run out of money. This would be easier to pull off if the hedge fund, exchange, and exchange customers all happened to share a bank account. (Edit: Matt Levine has a good writeup on this theory from right after the FTX collapse.)
(To be clear, there's no direct evidence that this was happening, at least not yet. And, like many other fraudulent behaviors, it's easy to imagine it starting as something innocuous: a new exchange needs volume, and one way to get that volume is to offer cheap liquidity. Two-sided networks always need someone to kickstart them, whether it's Pierre Omidyar selling a broken laser pointer on eBay or Ross Ullbricht growing some psychedelic mushrooms in order to get his marketplace going.)
One dynamic in the relationship between Google and Apple, recently discussed in The Diff ($), is that Apple has a lucrative audience to show search ads to and Google has a search engine that produces a lot of revenue per user. But another dynamic is that Apple, in practice, controls as much of the stack as is convenient, and almost never fully outsources a part of its business that might produce high margins some day. This piece notes that one way Apple can compete in search is to focus on the parts of search that Google doesn't really do, and has some disadvantages in, like searching within a user's device. Over time, that can easily extend to searching within apps—iOS is already very good at identifying cross-app interactions (like opening a link in Twitter, saving it to Instapaper, sending it to a groupchat directly from Instapaper, etc.). Some forward-thinking companies moved people towards apps specifically because they didn't want to pay the Google Tax for search: Yelp would prefer that people do searches within Yelp instead of Googling and potentially getting sent to a competitor. It's sometimes possible to reduce reliance on a specific search engine, but it's very hard for companies to escape the economics of search in general.
The IPO Window
VCs have been advising startups to hold off on going public, because of recent IPOs' lackluster returns ($, FT). The IPO cycle is closely tied to the overall market cycle, but as the current situation demonstrates, they're not identical; the Nasdaq is only 11% off its all-time high (but 7% off its recent late-July numbers, and near-term changes always loom large in a case like this). For the IPO window to be open, there needs to be some surplus of money and investor attention; for someone to buy shares in a newly-IPOed business, they need to get up to speed on it, and need to be comfortable with a rocky first couple quarters before they and the company figure out what the cadence of growth will look like. Fortunately for companies that are planning to IPO, this kind of sentiment can change quickly.
Markets in Everyone
Robin Hanson has a post revisiting the idea of being able to sell shares in one's future income, this time with the proposed mechanism of making a market in people's future tax receipts. He suggests some ways this could be useful, like automatically giving parents a share of their offsprings' future income. But one of his examples illustrates why this concept in general is so hard: he proposes that taxpayers could buy back stakes in their own future taxes, but this creates a huge adverse selection problem; these assets all need to be priced, and traders will want to trade only if they think they're getting a good deal. This adverse selection is one reason that successful ways of converting future income into present cash are usually debt rather than equity: when there's a wide range and information asymmetry, investors want to invest in the part of the capital structure where that's the smallest problem.
The dollar's value has risen about 6% since July, partly driven by higher US rates ($, WSJ). A higher dollar tends to be a tax on anyone who uses dollars, for borrowing and for pricing trade, but doesn't do most of their earning in dollars. These emerging market borrowers and dollar users have a higher marginal propensity to convert their earnings into investments that grow global GDP, whereas dollar-denominated investors are richer and live in countries where more of the obvious investments have been made. So in global terms, an expensive dollar is a suboptimal tax, which transfers money to people who won't use it as well as the taxpayer.
Companies in the Diff network are actively looking for talent. A sampling of current open roles:
- A systematic hedge fund is looking for researchers and portfolio managers who have experience using alternative data. (NYC)
- A company building the new pension of the 21st century and building universal basic capital is looking for a product manager with fintech experience. (NYC)
- A new fintech startup wants to bring cross-border open banking to LATAM, and is looking for a founding engineer. (NYC)
- A vertically integrated PE-backed cannabis company is looking for someone who has worked with complex operational and financial models in Excel. (Remote)
- A proprietary trading firm is seeking systematic-oriented traders with ML experience—ideally someone who has displayed excellence in DS and ML, like a Kaggle Master. (Montreal)
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.
If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.