March 26, 2021


Attention conservation notice: 1000-word grudging concession that a bete noire might have a point, followed immediately and at much greater length by un-constructive hole-poking; about social media, by someone who's given up on using social media; also about the economics of recommendation engines, by someone who is neither an economist nor a recommendation engineer.

Because he hates me and wants to make sure that I never get back to any (other) friend or collaborator, Simon made me read Jack Dorsey endorsing an idea of Stephen Wolfram's. Much as it pains me to say, Wolfram has the germ of an interesting idea here, which is to start separating out different aspects of the business of running a social network, as that's currently understood. I am going to ignore the stuff about computational contracts (nonsense on stilts, IMHO), and focus just on the idea that users could have a choice about the ranking / content recommendation algorithms which determine what they see in their feeds. (For short I'll call them "recommendation engines" or "recommenders".) There are still difficulties, though.

"Editors. You've re-invented editors."

Or, more exactly, a choice of editorial lines, as we might have with different, competing newspapers and magazines. Well, fine; doing it automatically and at the volume and rate of the Web is something which you can't achieve just by hiring people to edit.

— Back in the dreamtime, before the present was widely distributed, Vannevar Bush imagined the emergence of people who'd make their livings by pointing out what, in the vast store of the Memex, would be worth others' time: "there is a new profession of trail blazers, those who find delight in the task of establishing useful trails through the enormous mass of the common record." Or, again, there's Paul Ginsparg's vision of new journals erecting themselves as front ends to arxiv. Appealing those such visions are, it's just not happened in any sustained, substantial way. (All respect to Maria Popova for Brain Pickings, but how many like her are there, who can do it as a job and keep doing it?) Maybe the obstacles here are ones of scale, and making content-recommendation a separate, algorithmic business could help fulfill the vision. Maybe.

Monsters Respond to Incentives

"Presumably", Wolfram says, "the content platform would give a commission to the final ranking provider". So the recommender is still in the selling-ads business, just as Facebook, Twitter, etc. are now. I don't see how this improves the incentives at all. Indeed, it'd presumably mean the recommender is a "publisher" in the digital-advertizing sense, and Facebook's and Twitter's core business situation is preserved. (Perhaps this is why Dorsey endorses it?) But the concerns about the bad and/or perverse effects of those incentives (e.g.) are not in the least alleviated by having many smaller entities channeled in the same direction.

On the other hand, I imagine it's possible that people would pay for recommendations, which would at least give the recommenders a direct financial incentive to please the users. This might still not be good for the users, but at least it would align them more with users' desires, and diversity of those desires could push towards a diversity of recommendations. Of course, there would be the usual difficulty of fee-based services competing against free-to-user-ad-supported services.


To the extent there are concerns about certain content being banned by private companies, those are still there: the network operator, Facebook or Twitter or whatever, retains a veto over content. The recommenders are able to impose further vetoes, but not over-ride the operator.

Further: as Wolfram proposes it, the features used to represent content are already calculated by the operator. This can of course impose all sorts of biases and "editorial" decisions centrally, ones which the recommenders would have difficulty over-riding, if they could do so at all.

Increasing returns rule everything around me

Wolfram invokes "competition", but doesn't think about whether it will be effective. There are (at least) two grounds for thinking it wouldn't be, both based on increasing returns to scale.
  1. Costs of providing the service: If I am going to provide a recommendation engine to a significant fraction of Facebook's audience, in a timely manner, I require a truly massive computational infrastructure, which will have huge fixed costs, though the marginal costs of each additional recommendation will be trivial. It's literally Econ 101 that this is a situation where competition doesn't work very well, and the market tends to either segment in to monopolistic competition or in to oligopoly (if not outright monopoly). As a counter-argument, I guess I could imagine someone saying "Cloud computing will take care of that", i.e., as long as we tolerate oligopoly among hardware operators, software companies will face constant scale costs for computing. (How could that possibly go wrong, technically or socially?)
  2. Quality of the service: Machine learning methods work better with more data. This will mean more data about each user, and more data about more users. (In the very first paper on recommendation engines, back in 1995, Shardanand and Maes observed that the more users' data went in to each prediction, the smaller the error.) Result: the same algorithm used by company A, with \( n \) users, will be less effective than if used by company B, with data on \( 2n \) users. Even when the recommendation engine doesn't explicit use the social network, this will create a network externality for recommendation providers (*). And thus again we get increasing returns and throttled competition (cf.).

Normally I'd say there'd also be switching costs to lock users in to the first recommender they seriously use, but I could imagine the network operators imposing data formats and input-output requirements to make it easy to switch from one recommender to another without losing history.

— Not quite so long ago as "As We May Think", but still well before the present was widely distributed, Carl Shaprio and Hal Varian wrote a quietly brilliant book on the strategies firms in information businesses should follow to actually make money. The four keys were economies of scale, network externalities, lock-in of users, and control of standards. The point of all of these is to reduce competition. These principles work — it is no accident that Varian is now the chief economist of Google — and they will apply here.

Prior art

Someone else must have proposed this already. This conclusion is an example of induction by simple enumeration, which is always hazardous, but compelling with this subject. I would be interested to read about those earlier proposal, since I suspect they'll have thought about how it actually could work.

*: Back of the envelope, say the prediction error is $O(n^{-1/2})$, as it often is. The question is then how utility to the user scales with error. If it was simply inversely proportional, we'd get utility scaling like $O(n^{1/2})$, which is a lot less than the $O(n)$ claimed for classic network externalities by Metcalfe's law rule-of-thumb. On the other hand it feels more sensible to say that going from an error of $\pm 1$ on a 5 point scale to $\pm 0.1$ is a lot more valuable to users than going from $\pm 0.1$ to $\pm 0.01$, not much less valuable. Indeed we might expect that even perfect prediction would have only finite utility to users, so the utility would be something like $c-O(n^{-1/2})$. This suggests that we could have multiple very large services, especially if there is a cost to switch between recommenders. But it also suggests that there'd be a minimum viable size for a service, since if it's too small a customer would be paying the switching cost to get worse recommendations. ^

The Dismal Science; Actually, "Dr. Internet" Is the Name of the Monster's Creator

Posted at March 26, 2021 14:03 | permanent link

Actually, "Dr. Internet" Is the Name of the Monster's Creator

(I can't remember if Henry Farrell came up with this phrase, or I did, as the title for a possible joint project.)

See also Linkage

An Appeal to the Hive Mind (Ironically Enough)
Some Blogospheric Navel-Gazing, or, Strange Memories of the Recent Past
Books to Read While the Algae Grow in Your Fur, July 2017
Kill All Normies: Online Culture Wars from 4Chan and Tumblr to Trump and the Alt-Right (Nagle)
Experimental Considerations Touching on the Art of Winning Lotteries
Random Linkage, May 2015
The Presentation of Self in Internet Life
Speaking Truth to Power About Weblogs, or, How Not to Draw a Straight Line
One Roll of the Dice Will Never Abolish Warblogging
Someone Has Found a Way to Make Money from the Internet!

Posted at March 26, 2021 14:01 | permanent link

An Appeal to the Hive Mind (Ironically Enough)

Attention conservation notice: Asking for help finding something that you don't know about, that you don't care about, and that a bad memory might have just confabulated.

I have a vivid memory of reading, in the 1990s, an online discussion (maybe just two people, maybe as many as four) about what online fora, search engines, the Web, "agents", etc., were doing to the way people acquire and use knowledge, and indeed to what we mean by "knowledge". My very strong impression is that one of the participants was linked somehow with the MIT Media Lab, and taking a very strong social-constructionist line (unsurprisingly, given that affiliation). At some point the discussion turned to her experiences with an online forum related to a hobby of hers (tropical fish? terraria?). The person I'm thinking of said something like, the consensus of that forum just were knowledge about \$HOBBY. One of her interlocutors made an objection on the order of, why do you trust those random people on the Internet to have any idea what they're talking about? To which the reply was, basically, come on, who'd just make stuff up about \$HOBBY?

I have (genuinely!) thought of this exchange often in the 20-plus years since I read it. But when I recently tried to find it again, to check my memory and to cite it in a work-in-glacial-progress, I've been unable to locate it. (The fact that I don't recall any names of the participants, or the venue, doesn't help.) I am prepared to learn that, because this is something I've thought of often, my mind has re-shaped it into a memorable anecdote, but I'd still like to see what this started from. Any leads readers could provide would be appreciated.

Update, the next day

The hive mind Lucy Keer (with an assist from Mike Traven) delivers:

Specifically, the seed around which this story nucleated in my memory may have been a January 1996 piece by Prof. Bruckman in Technology Review — it has the right content (sci.aquaria!), the right date, my father subscribed to TR and I'd even have been visiting my parents when that issue was current. Only it's not a conversation between multiple people but a solo-author essay, it's not primarily about the social aspects of knowledge but about how to find congenial on-line communities and make (or re-make) ones that don't suck (the lost wisdom of the Internet's early Bronze Age), and contains nothing like "who'd just make stuff up about \$HOBBY?" (In short: Bartlett (1932) meets Radio Yerevan.)

More positively, I very much look forward to reading Bruckman's book (there's an excerpt/precis available on her website).

Actually, "Dr. Internet" Is the Name of the Monster's Creator; The Collective Use and Evolution of Concepts

Posted at March 26, 2021 12:32 | permanent link

Regression, Thermostats, Causal Inference: Some Finger Exercises

\[ \newcommand{\Expect}[1]{\mathbb{E}\left[ #1 \right]} \newcommand{\Prob}[1]{\mathbb{P}\left( #1 \right)} \newcommand{\Cov}[1]{\mathrm{Cov}\left[ #1 \right]} \newcommand{\Var}[1]{\mathrm{Var}\left[ #1 \right]} \]

Attention conservation notice: An 800-word, literally academic exercise about an issue in causal inference. Its point is familiar to those in the field, and deservedly obscure to everyone else. Also, too cutesy and pleased with itself by at least half.
I wrote the first version of this for the class where we do causal inference long enough ago that I actually don't remember when --- 2011? 2013? (In retrospect I had probably read Milton Friedman's thermostat analogy but didn't consciously remember it at the time.) Posted now because I've gone over the point with two different people in the last month.

The temperature outside \( (X) \) is a direct cause of the temperature inside my house \( (Y) \). But every morning I measure the temperature, and adjust my heating/cooling system \( (C) \) to try to maintain a constant temperature \( y_0 \). For simplicity, we'll say that all the relations are linear, so \[ \begin{eqnarray} X & \sim & \mathrm{whatever}\\ C|X & \leftarrow & a+bX + \epsilon_1\\ Y|X,C & \leftarrow & X-C + \epsilon_2 \end{eqnarray} \] where \( \epsilon_1 \) and \( \epsilon_2 \) are exogenous, independent, mean-zero noise terms. We can think of \( \epsilon_1 \) as a combination of my sloppiness in measuring the temperature and in tuning the heating/cooling system; \( \epsilon_2 \) is sheer fluctuations.

Exercise: Draw the DAG.

To ensure that the expectation of \( Y \) remains at \( y_0 \), no matter the external temperature, we need \[ \begin{eqnarray} y_0 & = & \Expect{Y|X=x}\\ & = & \Expect{X - a + bX + \epsilon_1 + \epsilon_2|X=x}\\ & = & (1-b)x -a \end{eqnarray} \] Since this must hold for all \( x \), we need \( b=1, a=-y_0 \).

What follows from this?

Exercise: Build your character by doing the algebra.

So, as long as control isn't perfect, the naive statistician (or experienced econometrician...) who just does a kitchen-sink regression will actually get the relationship between \( Y \), \( X \) and \( C \) right, concluding that external temperature and the climate control have equal and opposite effects on internal temperature. Sure, there will be sampling noise, but with enough data they'll approach the truth.

Exercise: What do you get if you regress \( C \) on \( X \) and \( Y \)?

I have implicitly assumed that I know the exact linear relationship between \( X \) and \( Y \), since I used that in deriving how the control signal should respond to \( X \). If I mis-calibrate the control signal, say if \( C = -y_0 +0.999X + \epsilon_1 \), then there is not an exact cancellation and everything works as usual.

Exercise: Suppose that instead of measuring the external temperature \( X \) directly, I can only measure yesterday's temperature \( U \), again with noise. Supposing there is a linear relationship between \( U \) and \( X \), replicate this analysis. Does it matter if \( U \) is the parent of \( X \) or vice versa?

Exercise: "Feedback is a mechanism for persistently violating faithfulness"; discuss.

Exercise: "The greatest skill seems like clumsiness" (Laozi); discuss.

Engimas of Chance; Constant Conjunction Necessary Connexion

Posted at March 26, 2021 09:08 | permanent link

Three-Toed Sloth