The Uses of Friction

Plus! Market-Making; Poaching and Equity Currency; China's Covid Economy; The Cost of AI; Friendshoring; Diff Jobs

The Uses of Friction

Software success stories can often be summed up as a relentless quest to eliminate friction:

All of these companies have created billions of dollars of market value by greasing assorted skids. And their user numbers and revenue demonstrate that this has been a rewarding activity. And yet, there are some compelling examples of companies who find better results when they create friction instead.

The first and most annoying category of this is adversarial friction: companies going out of their way to make some user action more annoying because the action transfers money in a way that company doesn't like. Expensify makes it easy to submit even a small business expense for reimbursement. Concur makes it hard. If the way to judge expense management software is ease of use, Expensify and its newer cohort of competitors easily win. But if the goal is to reduce corporate expenses, Concur has something going for it.1

Plenty of companies combine a low-friction signup process—free trials (just pay for shipping or installation!) with a convoluted cancellation process involving phone calls, faxes, proof of residence change, etc. This model tends to show up in a few cases:

This is adversarial friction for a reason: everyone involved sees a fairly fixed pool of money and the company has a larger total interest in ensuring that they get the biggest piece. They either don't have to worry much about customer switching or don't expect to exist very long.2

A middling version of where friction works is in finance, specifically lending. Lenders have numerous opportunities to lower friction in their process by, for example, not carefully checking borrowers' claims about their creditworthiness, lending to less creditworthy borrowers, not checking their ability to pay at all, or continuously reducing the interest rates charged. Every one of these choices will increase conversion rates, but with obviously bad side effects.

A more subtle problem a financial company can run into is improving conversion rates by making it easier for borrowers to agree to things they might be surprised by. This retrospective is a great example: offers interview prep, and chose the very reasonable business model of charging people only once they successfully landed a job. They refined the pitch for this to make it simpler, eventually replacing an e-signed contract with a checkbox and a terse explanation of the terms. This was still positive-sum on both sides—users get interview help, and only pay if they get the kind of job that a) they're aiming to get, and b) they struggled with enough to justify wanting interview help. But people really don't like accidentally signing up for financial obligations; if their eyes glazed over on a key part of the signup flow, they'll feel like they fell for a scam.

The other side of low-friction the coin is something that might be called collaborative friction. One classic example is in hiring, where some companies have a deliberately punitive process for either recruiting new employees or for how challenging the initial work will be. Nothing builds esprit de corps like shared suffering and a shared sense of excellence. But most employers can't do this. For the few that can, the process doesn't start with building a ludicrously challenging hiring process—it starts with being an obviously desirable place to work with no need for a hiring process at all. This can happen with businesses that naturally generate high revenue relative to headcount, like prop traders, or companies that have many cofounders who are close to multiple networks of skilled people looking for jobs. Early Google and PayPal both had this: in addition to having founders who came from good schools and had friends they already wanted to hire, both companies were well-funded at a time when competitors weren't. (And neither has quite the same strict hiring standards they had in the old days.)3

But this also happens with user interfaces. Replit's CEO says that redesigning an unnamed feature to introduce more friction led to a "70% increase in conversions". And Ali Abouelatta has a very illuminating thread about a specific example of this with DuoLingo: when they asked users to pick a goal for how many days in a row they'd want to use the app, and when they were forced to pick one instead of having a predetermined selection, they chose more aggressive goals and had higher engagement.

It's notable that both of these are learning apps: DuoLingo teaches people languages through quick drills (subject of a Diff piece here ($)) and Replit teaches them how to code (subject of a Diff post in the future). This is telling: when people talk about the virtues of friction, they're not usually talking about the immediate benefit of doing some specific task the hard way rather than the easy way, but the long-term benefit of being used to doing hard things. You can manually run bubblesort if you really want, but the main point is to cultivate the ability to look at a problem without many shortcuts, like learning a new programming language or learning a new human language, and to be willing to actually put in the effort.

The attrition rate for online courses is 96% according to one study ($, FT), and that makes it highly sensitive to small interventions: even if friction makes a course harder, if it increases the variance of users' commitment levels, that's more likely to flip someone who's in the supermajority of future-dropouts than the tiny minority of likely completers.

But does this matter for people outside of the education business? Probably: many companies are partly in the education business. Enterprise software companies have to learn about their customers' processes and then convince those customers to change those processes so that the new product is tightly-coupled with the company's growth. Customer education is part of consumer-facing businesses, too; Google's growth has been a very long process of adding to the list of things people naturally Google. For investors, one lesson is that the newer a company’s product is, the less of a bad sign slow growth is, as long as it’s steady. It’s relatively easy to acquire users for a better replacement to an existing category, given enough marketing spending—but that also applies to future competitors. Whereas a company that pursues a higher-friction model isn’t just acquiring temporary users who will jump to the next low-friction replacement; it’s acquiring loyalists who will stick around for longer.

For an obviously-better product, reducing friction means reducing the effort it takes someone to test out the product, realize it's superior, and never go back. But for something in a completely new category, where learning is part of the process of using it, a little friction can be a useful thing.

Thanks to Stephen Pimentel for this tweet, which inspired this piece.


A Word From Our Sponsors

When you give to charity, how much impact will your donation actually have? This question can be hard, if not impossible, to answer. What programs does the charity implement? What is the evidence that those programs are effective? Can the charity absorb additional funding and use it productively?

GiveWell is an independent nonprofit that spends over 40,000 hours each year searching for the charities that save or improve lives the most per dollar. GiveWell reviews academic evidence, models cost-effectiveness, interviews charity staff, and forecasts future charity spending so that any donor can make a high-impact donation.

Over 110,000 donors have used GiveWell to donate more than one billion dollars. Rigorous evidence suggests that these donations will save over 150,000 lives and improve the lives of millions more.

All of GiveWell’s research is free for anyone to use or review. GiveWell allocates your tax deductible donation to the charity or fund you choose without taking a cut.

First-time GiveWell donors from The Diff will have their donations matched up to $100 before the end of the year or as long as matching funds last. Go to to claim your match.


I had two podcast appearances last week, on Spencer Greenberg's Clearer Thinking and Dwarkesh Patel's The Lunar Society. Both nominally about FTX, but the latter also about the pursuit of power, the social impact of drugs, and the benefits of monasteries.


As this newsletter has noted many times, making a market without an edge is equivalent to writing put and call options: when volatility is low, a market maker collects easy money by buying low and selling high, and when the market is moving in a specific direction, the market-maker systematically buys high or sells (short) low, and loses money a lot faster than they made it. As we get more information about FTX and Alameda, we now have an example: Alameda lost $1bn by assuming an FTX customer's position in the cryptocurrency MobileCoin ($, FT).

As a reminder, the way this came about was that FTX and Alameda had a unique relationship: FTX indicated to customers that it would be a more stable exchange, with fewer wild price swings, because it had an emergency liquidity provider in the form of Alameda. But emergency liquidity provision is a form of falling-knife-catching: the more extreme a price move, the less likely it is to be random noise.

It's interesting to speculate here that, in the short term, this $1bn loss was actually worth it to Alameda, since at least some of that loss would otherwise have been absorbed by FTX customers. By keeping the FTX sales pitch alive, they kept FTX growing, meaning that the exchange could still raise money. But this amounts to a Martingale bet: any time some financial product has a given level of volatility and an investor promises to use it to deliver lower volatility, they're on the hook for extreme price moves that they can be confident will eventually happen. Creating the appearance of offering the same returns and lower volatility than the competition makes it easy to raise money, but that just means that providing that lower volatility is even more expensive. In the end, something breaks, as we saw last month.

Poaching and Equity Currency

Auditing and consulting firm EY has been considering a split between their auditing and consulting units, and taking the consulting company public. Which has led competitor PwC to attempt to poach some of their partners ($, FT). One thing that keeps big partnership privately-held is that the bigger they are, the harder they are to take public—converting everyone's partnership stake into a liquid asset leads to lots of arguments, especially between people who expected their stake to grow over time and the ones who want to get out with the most valuable stake they can. So it makes sense that some partners would be shaken loose, and that these would be some of the most desirable hires—the ones who are confident they'll merit higher ownership stakes in the future.

But the competing firms don't have a completely straightforward advantage here:

Partners at the Big Four accounting and consulting firm have been told the hiring spree could hold down their share of profits in the coming year, but would bring a competitive edge over the long run...

When a company has publicly-traded equity, opportunistic growth can be good for the stock price, and can be funded in part by issuing more stock. If it takes the form of hiring, this can be even easier, since investors seem to underreact to equity compensation. But for a private company, expansion means cutting partners' payouts, and while they don't see the appreciation in the future value of their partnership, they do see the cut in their annual distribution.

China's Covid Economy

The Economist has a good piece on the difficult situation China is in: since China locked down and other countries didn't, the virus has had a chance to mutate and become more contagious, so an infection today will spread much faster than 2020's did. Since the rest of the world has more herd immunity, China's the only country that would experience R(0) rather than a gentler R(t), a scenario for which their healthcare system is unprepared.4 Another obstacle is that China now has a large Covid economy: "A broker, Soochow Securities, has estimated China’s bill for covid testing at 1.7trn yuan this year, or around 1.5% of GDP." (For perspective, this means that China's Covid testing alone constitutes as much economic activity as the entire GDP of Portugal.)

The Cost of AI

This post has a great overview of the cost, in money and time, of training various AI models. In general, training time rises faster than model complexity, as does price. This means that the economics of AI will be somewhat counterintuitive for people used to other software businesses, and profits will depend in part on how visible differences are between the best and second-best models.


Apple is trying to move key stages of the manufacturing process out of China ($, WSJ). One stage is new product introduction, "when Apple assigns teams to work with contractors in translating its product blueprints and prototypes into a detailed manufacturing plan." Since this is the interface between "Designed by Apple in California" and "Made in [somewhere else]," it's an important stage, and being able to do this well is a competitive advantage; the point at which blueprints become a manufacturing process is also the point at which one particular manufacturer's process gets imprinted on the supply chain.

Diff Jobs

Companies in the Diff network are actively seeking talent! If you're interested in exploring growth opportunities at unique companies, please reach out. Some top current roles:

If you’re 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.

We're especially interested in talking to companies interested in remote contractors with finance and operations experience. If your startup needs half of a Head of Strategy or Director of Finance, please reach out.


This doesn't necessarily mean the companies that use worse software are trying to squeeze their employees. There are two equilibria here: people mostly make responsible and prudent decisions, and the rules are simple and straightforward enough that they're easy to follow—or a company has byzantine rules that have evolved in response to previous abuses, and the employees and company have deep mutual distrust. This kind of situation can arise by accident, as when a company has generous expense policies early in its existence and then tries to tighten the belt later on.


It's worth considering whether such reputational concerns will become more salient now that it's easier to Google people, or to write negative things about a company's owner online as opposed to just complaining about the brand. Perhaps this kind of especially abusive practice will become a specialty of people with extremely common names.


The other obvious hack to get the most out of a brutal hiring process is to have high pay, ideally tied to performance and weighted away from year one. But this, again, is a secondary effect. Step one is to have a business that justifies this kind of compensation.


In theory one solution would be to randomly un-lock-down a randomly-selected 1% of households at a time, let them get infected, and treat the resulting infections without overwhelming the hospital system, and then to un-lock-down another 1-2%, and to keep doing this over a few years to gradually build herd immunity. But the practical difficulties are enormous, and if there's one thing less tolerable than Covid Zero, it's a version of Covid Zero that capriciously fails to apply to some random segments of the population. Many countries essentially did a poorly-measured, ad hoc version of this, as some people completely ignored the pandemic, others started skirting rules early, and others have continued to self-lock-down indefinitely. Ironically, making lockdowns a live political issue and a culture-war signifier did, in fact, slow the spread and help keep hospitals from being overwhelmed. Not that we should try it next time, too.