You Hate The Products You're An Early Adopter Of Because of Invisible Base Rates and Anthropic Bias
You Hate The Products You're An Early Adopter Of Because of Invisible Base Rates and Anthropic Bias
It's an inevitable complaint: when you started using Google, it tried to give you the best match for the search you entered. Now, it tries to give you a fresh result for a broader-interest variant of something you're searching for. (It’s especially annoying when a major event happens for the second time, and you want to compare it to the first—no, I'm not getting details about FTX's bankruptcy wrong, I'm trying to remember what Investors Overseas Service was called!) This often happens when searching for something technical and obscure that sounds like something nontechnical and more common. And this happens on websites, too; Reddit used to have front-page discussions about the relative merits of Python and Lisp for web applications.[1] Twitter used to have more of a tech scene flavor, too, and when Quora first got going it was basically an oral history of post-dot com bubble Silicon Valley, produced by participants—not a site for viral stories and homework help. This even happened to Amazon! Look at the top 25 books in their 1995 bestsellers list, and there's exactly one work of non-science fiction—at the same time, there are two different Douglas Hofstadter books.[2] Meanwhile, autocorrect insists on fixing words like "cant" or "dint," which are perfectly legitimate words but are now slightly inconvenient to use.
Cory Doctorow has some salty words ($) to describe this tendency, mostly using Amazon as an example. His key point is hard to ignore: it's undeniable that in many respects, the site has gotten worse for sellers—take rates have gone up, Amazon pushes them to use its own shipping infrastructure, and internal-to-Amazon search ads basically force sellers to give up their incremental margin in order to get any appreciable sales volume ($, Diff). Meanwhile, the site is simply less geared towards the interests of nerdy users (like the ones who read Cory Doctorow books) and more towards those of average shoppers who just want a deal.
But this tendency also describes a natural corporate lifecycle pretty well.[3] Early on, a company's in a mostly exploratory process. It does need to pay attention to costs, and Amazon was famous for doing this from the beginning. But the main goal is to maximize total wealth creation upfront and figure out how to capture it later on. It's naive to think that this is an act of generosity, but it's almost equally naive to feel betrayed that it was a temporary state of affairs. A company has no reason to operate at a loss unless it expects to turn a profit later on, and if it's operating at a loss because it chooses to charge low prices, it's reasonable to bet that those prices will move up eventually.
And when those prices move up, it's generally because a service has grown beyond its initial audience of early adopters and become mainstream. This is part of the cultural change that gets observed in most big services over time. It’s caused by the twin curses of base rates:
- First, the more average the average user is, the more likely the best interpretation of their obscure (or clever) search is that it was, in fact, a mistake. The reason Google insists that your search for a niche topic that a news story reminded you of is in fact an attempt to search for that news story is that this is, on average, the right bet to make. The median user who performs your query is making a mistake, which Google is helpful enough to fix![4]
- The second curse of base rates is that even if a small percentage of people are terrible, platform size and time raise the probability of bad behavior to 1. This is why ChatGPT will be reluctant to give you many tips on recreational explosives. Even if you are a very competent person who happens to enjoy explosions, the median user asking for these instructions may have more sinister motivations.[5]
These phenomena tie in with another reason that so many services you use have gotten worse since you started using them: the Anthropic principle is the tautological observation that, to make any given observation, you need to exist first. This was originally an answer to the philosophical or theological debate about whether earth's status as a great planet for living things was evidence of divine intervention. The initial form of the argument went like this: if the universe were not set up in a way that was conducive to intelligent life, there wouldn't be anyone around to notice, so of course we live in a universe that supports life, and on a planet unusually well-adapted to it.
But you can extend this to a micro-level: of course that new hole-in-the-wall restaurant you discovered has gotten way more crowded since you first started going there—you found a good deal, it’s just that now other people know about it, too! And of course that social network you use has gotten less nerdy and elitist over time—nerds are early adopters who tend to type a lot, and non-elites value proximity to elites almost by definition.
In other words, the more perfectly-matched a site or service is for you, the more likely you are to be a very early user in for some very long-term disappointment. And that disappointment gets elevated because in some cases, the service gets harder to ignore as it grows. If you joined LinkedIn when it was mostly one or two degrees away from the PayPal Mafia, you might be annoyed by spammy messages now, but it's hard not to have some kind of presence on the site if you're a professional.
There are some exceptions, which don't eliminate the value-capture part of the puzzle but which do mitigate the quality decline piece. For example, Hacker News is basically the same site it was a decade ago, which has made it priceless from a branding perspective. But that requires constant, deliberate effort, and sometimes through quick and drastic interventions (the first time the site got widespread external attention was when Paul Graham asked users to submit articles about "the innards of Erlang" instead of fluffier topics, and then asked them to stop because the entire front page was one Erlang post after another). Some publications manage the same problem by keeping a narrow focus on a lucrative demographic; Hedge Fund Alert is not tempted to broaden its focus to meme stocks and scandals when straightforward hard news reporting gets them a sticker price of $4.7k/year from 6,000 customers.
In both of those cases, though, there's a monetization model: in Hacker News' case, it's by acquiring equity, and in Hedge Fund Alert's, it's just by charging a high price. In this case, the barriers to becoming mainstream are part of what people pay for. But that's not strictly necessary: some people make a good product, and refuse to profit by making it worse, simply because they like that outcome more regardless of whether it's the most lucrative.
But many, many people will, when given the opportunity, accept a trade where the product they made gets 1% worse so long as it appeals to twice as many people. And if you take that trade enough times, it might go from a beloved niche thing to a lame mainstream product. And for most people, that's a benefit (unless you were literally a founding user for the site you used to love, there was a time when it was too cool and obscure for you to know about it, and from earlier users' perspective you were part of the decline). For the most part, if you discover something great, you should either brace yourself to like it less over time or hope that it's a labor of love.
Disclosure: Long AMZN.
The first time I recall Reddit users revolting against management was when Reddit gave up on Lisp and switched to Python, back in 2005. This was actually before Reddit had implemented comments, so the active discussion, complete with predictions of doom, happened on Usenet. ↩︎
It's funny to imagine the behaviors of Amazon's userbase in 1995 scaled to the size of that userbase today. Imagine living in a country where young adult fiction does not have nearly the cultural weight of, say, a Bruce Schneier textbook on cryptography. If you spent a lot of time online in the mid-1990s, that's the world you lived in! ↩︎
Indeed, it's not too far off from the human lifecycle: you start out with all of your needs taken care of and all of your decisions made by reasonably competent adults, and every year after that is an unpleasant shock in which more is expected of you and the path ahead is less and less clear. It is, in both cases, part of growing up. ↩︎
The other day I found what I thought was a truly bizarre Google search; I was blanking on Petrarch's name and Googled "renaissance poet," and while most of the results were relevant, Google also surfaced people like Countee Cullen and Claude McKay, who were active in the 20th century. But this was not a bizarre bug after all: some fraction of people doing that search actually meant to type "Harlem renaissance poets," and the people who do this are numerous enough that Google can infer that some fraction of "renaissance poet" searches were supposed to include that word, and will include the correct result. This is a very mild inconvenience for most users, but the key thing to note here is that when Google spots a mistake in your query that you yourself didn't notice, you probably don't register them overriding your stated intentions. Meanwhile, in the very process of writing this essay, Google insisted on not giving me a particular result I was looking for: at some point, I think in a Washington Post article on the tenth anniversary of Washingtonpost.com, Warren Buffett was quoted pointing out that there are few examples of a media property getting a large audience and not finding a good business model. Unfortunately, Google is excessively helpful at finding other things I might have been looking for instead. ↩︎
For what it's worth, the good news is that ChatGPT will give you advice on synthesizing LK-99, though it will nag you a lot about proper safety procedures. ↩︎
Diff Jobs
Companies in the Diff network are actively looking for talent. A sampling of current open roles:
- A company using Web3 to decentralize customer loyalty programs is looking for a founding senior engineer with Solidity experience and an interest in brands and the arts. (Brooklyn)
- A company building the new pension of the 21st century and building universal basic capital is looking for a GTM / growth lead. (NYC)
- A company building ML-powered tools to accelerate developer productivity is looking for a mathematician. (Washington DC)
- A vertically integrated PE-backed company applying a rigorous investment/operations approach to a high-growth industry is looking for an analyst who has banking experience. (Little Rock, AR—no remote, but relocation assistance is possible)
- A company building zero-knowledge proof-based tools to enable novel financial arrangements is looking for a senior engineer with a research bent. Ideal experience includes demonstrations of extraordinary coding and/or math ability. (NYC or San Diego preferred, remote also a possibility.)
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.
Elsewhere
The Big Tech Cycle
Andrew Walker calls attention to an interesting phenomenon: from time to time, big tech companies get surprisingly cheap—not "cheap relative to other high-quality growth companies" but cheap relative to the overall market. There are two pieces to this argument: first, tech investors who have been around long enough know that some companies with apparently unbeatable advantages can go through long fallow periods, or die suddenly. There's also political mean-reversion; two of the earliest Big Tech companies were AT&T and IBM, both of which experienced worse results after their size and aggression got the government’s attention.
So one takeaway is that these companies sometimes get cheap because they face idiosyncratic, unhedgeable risks. And one conclusion from that is that Big Tech has its own risk premium: the best-run tech companies will, in an average year, outperform the market. But every once in a while, there's a drawdown that changes both the fundamentals and the narrative. So at least some of the returns investors have gotten from these companies constitute a sort of premium from insuring against the risk that at least one of today's tech giants is the next IBM.
Chinese Consumer Internet Competition With American Characteristics
Shein and Temu have been downplaying their Chinese origins and suing one another as competition between them escalates ($, WSJ). Some of their competition follows a model common in China, where there are more straightforward attempts to monopolize other parts of the supply chain. The US version of this is a bit more genteel; some behaviors have the effect of encouraging a supplier to only work with one company in an industry, but without explicitly saying so. Resolving these disputes through litigation demonstrates that the companies have adopted some US business norms wholesale.
Reallocating
Blackstone's BREIT has been in the news recently because of significant investor withdrawals, which has led to a debate over whether a) they are one of the best-managed real estate companies in the US, or b) they're taking advantage of the opacity of being a non-listed fund to artificially mark up their assets. Cutting against the latter thesis is the fact that they've been liquidating some properties recently to fund more investments in datacenters, and have generally done so at a profit ($, FT). It's not impossible for this to coexist with assets that are mismarked overall—part of what China Evergrande did, for example, was to realize profits on developments where prices had gone up, and to hold in inventory anything that hadn't. That strategy works until the company either runs out of appreciated properties to flip or runs out of cash. But intentionally doing that is a very dangerous strategy, and is probably negative expected value for an asset manager that has multiple funds and thus more reputational risk. There will always be skepticism when a private asset seems to outperform publicly-traded peers, but in this case there's circumstantial evidence that that's exactly what's going on.
Durables
The mid-2010s direct-to-consumer mattress boom is over, as is the pandemic-era furniture-buying spree, but new DTC mattress companies keep popping up, albeit with a more sustainable business model. Even after a sector goes through a period of overinvestment and high competition, it will often be structurally more vulnerable to new entrants: there's an established playbook for launching and growing such a company now, and in a cyclical industry that means supply expands as soon as there's a bump in demand.
Crypto Use Cases
The crypto industry is in one sense far more transparent than the rest of the financial system: if you feel like it, you can trace a given user's transaction history to your heart's content, or follow a given sum of cryptocurrency around the blockchain to see where it's been and where it goes next. But in another sense, crypto is full of mysteries, since there's no transparency into why people use some of its properties. For example: why does anyone want to hold Tether, which is pegged to the US dollar but backed by a company that does not go out of its way to avoid the appearance of being fraudulent. There aren't perfect answers, but there are clues ($, FT): the volume of Tether/ruble exchanges almost quadrupled the day of the Wagner Group insurrection last month. Tether is worse than the dollar in many respects, but if someone doesn't have direct access to dollars, perhaps because their name is mistakenly or quite correctly on a list of sanctioned individuals, Tether is the next best thing.