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

Three-Toed Sloth