Connected Issue #2: Killing Procrustes
I. The Procrustean bed
Quoting from this paper, titled The Procrustes’ Bed and Standardization in Education, this is the legend of Procrustes.
“Procrustes was a crook from Attica, who took the lives of many innocent people passing by his fortress gate. Most of his victims were travelers who were new to the area and unfamiliar with him. With disguised hospitality, Procrustes approached travelers passing by his residence and made a generous offer of a free night’s stay. Tired from their long walk, travelers usually accepted the offer with gratitude. Unfortunately, none of those who stayed at Procrustes’s house left the place alive.”
“The shadow of death fell on each guest as the night fell. As soon as they were in bed and asleep, Procrustes would tie them tight and measure their height. The unit of measure Procrustes used was the length of his bed. If the guest was taller than the length of the bed, Procrustes cut off the parts of his limbs that reached over the length of the bed. If the guest was shorter than the bed, he was stretched out until he fit the full length of the bed.”
The story of Procrustes is instructive of the demerits of forcing arbitrary standards at scale and an exaggerated demonstration of the psychopathy involved in doing so.
II. Is scale always good?
A recurring theme in my thinking is the notion of scale - what breaks when you scale something? How does it differ if it’s a process, an organization, a clique, even your hair? Why is scale taken for granted in a Connected world?
III. How do you scale?
Using leverage. Simply put, leverage means getting more out of less. Naval, in his famous tweetstorm, How to Get Rich Without Getting Lucky, says the following about leverage:
“Fortunes require leverage. Business leverage comes from capital, people, and products with no marginal cost of replication (code and media).”
He then goes on to distinguish between permissioned leverage (capital, people) and permissionless leverage (code and media) and how the latter has zero cost of marginal distribution making it an ideal tool for exercising leverage and achieving scale.
Scale and size are taken for granted in a Connected world whether business or politics and the key to growth is leverage and zero marginal cost of production and distribution.
Buzzwords such as unicorns, economies of scale, hypergrowth, blitzscaling, going viral etc. have taken center stage in our narratives. The unspoken, underlying assumption that scale and fast growth is good can be seen everywhere from influencers on social media trying to achieve virality to startups aiming for unicorn status through VC funding.
This phenomenon of favoring scale can be seen in domains outside of tech too. The general trend observed is that most of our institutions have been growing in size whether schools, farms, or firms. What are some downsides of scale? Is there a difference between good scaling and bad scaling? How do you evaluate the trade-offs?
A slight detour first.
IV. The Idea of Legibility
Let’s try to understand the idea of legibility using some visual aids. Three pairs of images before we go technical.
A checkered square and a random dot diagram
The maps of Brügge and Chicago
(from Slate Star Codex)
A natural, illegible forest and a scientific, legible forest
These three pairs of images should give you a feel for what we mean by legibility.
The simplest definition of legibility is how clearly an object can be parsed by an observer. Legibility is affected by both, the nature of the object and the ability of the observer to see. In essence, legibility is a measure of the ability of the observer to “see” a particular object.
Legibility can be defined as the highest resolution at which an agent can parse a subject.
An observer’s ability to see, let’s call it resolution, depends on both its viewing apparatus, which I call the legibility function and the object’s complexity signature, i.e. whether it is low or high dimensional, dead or alive, static or dynamic. This gives us 4 types of interactions depending on the resolution and complexity signature of the object.
Two interactions I will focus on in this post are:
High dimensional, complex objects interacting with observers having low resolution (Type A)
Low dimensional, static objects interacting with observers having high resolution
The other two types of interactions (high-high and low-low) are less interesting to me for the purpose of this discussion but might feature later in the series.
The interaction map
A good example of a Type A interaction is that of the state and its people. It is also the inspiration for the idea of legibility as introduced by James Scott in his book, Seeing Like a State. The state, in order to manage its people and resources, introduces measures and standardization, thereby imposing legibility on an illegible object, its populace.
The classic example Scott uses is that of modern cities being literally “legible” to humans in the form of grids and aesthetically pleasing geometry but poorly designed for anything else. Another good example is that of legibility applied to forestry where governments slash illegible forests and indulge in “scientific” forestry, only to realize how damaging it is in the long term.
In this first type of interaction, making something legible is the act of transforming a high dimensional object to a low dimensional entity based on the capabilities of the agent. It is an act of abstraction in service of superficial characteristics and control.
Some abstraction functions are better than others at capturing the essence of the subject but more often than not, legibility manifests itself as an act of reductionism that brings with it, loss of information and deleterious consequences. In this sense of the word, legibility can be interchangeably used with standardization or abstraction.
For those interested in a richer discussion on legibility, see this post by Venkatesh Rao called A Big Little Idea Called Legibility. This is one of those ideas that can change your worldview.
For the Type B interactions, consider the phenomenon of industrialization. The idea of using machines to make goods from raw materials is based on legiblizing the the final product and using standardization or moulds (sometimes in the literal sense of the world, e.g. plastics and injection moulding) to produce them inexpensively at scale. This type of interaction enables the production of a wide variety of products based on the capacity and imagination of the legibility function being employed.
V. Legibility and Scale
The premise of this essay is the following: in the first kind of interaction, scale imposes legibility and in the second kind, legibility enables scale.
VI. Scale imposes Legibility (Type A interactions)
When an observer with a low resolution legibility function interacts with complex or high dimensional objects, the need for scale imposes legibility, i.e. scaling things requires that you adopt a low-resolution worldview, thereby skipping small but important details.
The notion that scale imposes legibility can be seen everywhere in a Connected world. The capitalistic economy treats all its workers as the same, expecting them to work for the same duration in a similar manner. The bloated school system treats all children the same, failing to recognize that learning preferences may vary. It also judges them on a score that is assumed to encapsulate all of the child’s learning. Universities admit students based entirely on an entrance exam score, using it to determine the candidature and potential of a student. The big fast food restaurant you go to can only make 5 types of burgers because it needs to sell a thousand each hour. The choice of doing things at scale in these examples leads to the imposition of legibility and loss of diversity.
The following paragraphs from A Big Little Idea Called Legibility are particularly illuminating. Observe the theme of centralization and scale leading to imposition of legibility and subsequent failure.
“For every complex problem there is an answer that is clear, simple, and wrong.”
“The book is about the 2-3 century long process by which modern states reorganized the societies they governed, to make them more legible to the apparatus of governance. The state is not actually interested in the rich functional structure and complex behavior of the very organic entities that it governs (and indeed, is part of, rather than “above”). It merely views them as resources that must be organized in order to yield optimal returns according to a centralized, narrow, and strictly utilitarian logic.”
“The book begins with an early example, “scientific” forestry (illustrated in the picture above). The early modern state, Germany in this case, was only interested in maximizing tax revenues from forestry. This meant that the acreage, yield and market value of a forest had to be measured, and only these obviously relevant variables were comprehended by the statist mental model. Traditional wild and unruly forests were literally illegible to the state surveyor’s eyes, and this gave birth to “scientific” forestry: the gradual transformation of forests with a rich diversity of species growing wildly and randomly into orderly stands of the highest-yielding varieties. The resulting catastrophes — better recognized these days as the problems of monoculture — were inevitable.”
This is both a feature and a bug of scaling. Low marginal cost of production and distribution means you need tools such as a metric, measure or process that is cheap to replicate and infinitely repeatable across a broad set of subjects but it imposes legibility and comes with loss of information, diversity, and complexity.
VII. Legibility enables Scale (Type B interactions)
When an observer uses a high resolution legibility function on an object, it can then use that information to parse a whole class of objects that are similar in nature, i.e. if you can take a previously illegible object and make it slightly legible, you can scale your process to object that are similar. In such cases, legibility enables scale.
The idea of a low marginal cost of production and leverage flows directly from the assumption of standardization and legibility. Once you’ve invested sufficient time and effort to understand and make legible an object, the marginal cost of working with it in the future drops to zero.
VIII. Tension between legibility and scale
To make this idea legible, I’ll use a few examples that are illustrative of this tension between scale and legibility.
Clothes made at a tailor shop make your body size/shape fully legible and are a perfect fit. The clothes you buy at a store can only fit you so well because they’re standardized to select sizes. Manufacturers made average body size and shapes legible to industrial machines that make clothes, thereby achieving scale. Subsequently, the need for scale then imposed legibility on the people who lie outside the mould.
The US Air Force story on “average” cockpit sizes is a perfect example of how top down design of cockpits meant for the average pilot fits exactly no one.
The second example is that of food. The ability to customize food, i.e. make it fully legible for your tastes and decreases as we move from home to small restaurants to large industrialized kitchens like fast food chains. The choice to make things at large scale generally brings reduced variety with it.
The third example is that of education and learning. The reliance on standardized ways of teaching and usage of proxy measures for measuring effectiveness of learning increases as we move from 1-o-1 tutoring to small classrooms to large classrooms in universities and MOOCs. The imposition of legibility on students’ learning preferences is a byproduct of the scale at which the process is being run.
Scott alludes to this tension between scale and legibility in an interview.
Interviewer: Your critique on the engineering of society has been judged as a plea for the free market. Yet you are a self-acclaimed anarchist. Could you explain?
Scott: Some consider Seeing Like a State a right-wing book because I had an occasional good word to say about people like Friedrich Hayek and Michael Oakeshott. My answer to that charge is that I’d like to write a book about the ways in which large capitalist firms rely on standardization in exactly the same way as do nation states. Take a look at McDonald’s and their tools of management and control. The only difference with a nation state is that they have to make the standardization pay in terms of profit.
This point of difference is crucial for incentive alignment and effective outcomes but that’s a separate story.
The full interview is worth reading and can be found here: Seeing Like a Society Interview with James C. Scott
IX. Technology and A Failure of Imagination
If there are processes and interactions that don’t scale to all sizes, the important question to answer becomes one of finding a level of scale that is appropriate.
It is clear that there are trade-offs that need due consideration when choosing between legibility and scale. More importantly, not all scaling is bad. If the alternative to low quality education at scale is no education at all, then it is likely the better option. If the market offers reliable, decent fit clothing to everyone at a reasonable price, is that too bad? Perhaps not.
The examples I discussed to illustrate the tension between scale and legibility all exhibit signs of this trade-off, in this case quantified by cost and utility for consumers. A small restaurant that makes the perfect dish for you or the tailor that makes custom dresses for you are bound to charge you higher for the added effort required to make those. In this sense, the trade-offs between scale and legibility can be resolved if we use a utilitarian lens - ‘the greatest amount of good for the greatest number’.
This defense does not work we look at examples where scale imposes legibility. Even a utilitarian lens doesn’t hold merit. The only answers I find acceptable is that we either lack the technology or imagination to come up with ideas that would enable the representation of a richer, complex reality.
An important point to note here is that not all scaling falls prey to failure modes associated with legibility. Poorly thought out execution of scaling does. This line from A Big Little Idea Called Legibility sums up why legibility, often imposed by a need for scale and centralized gains can lead to failure:
“The deep failure in thinking lies is the mistaken assumption that thriving, successful and functional realities must necessarily be legible.”
This quote also tells us that at times, we even need to question whether the existence or creation of something at scale is necessary.
Broadly speaking, technology is the creation of processes that you can use with little to no extra effort once defined and made legible. The task then for technologists is to come up with ideas that allow us to harness this leverage and legibility (via technology) while retaining maximum information and complexity.
The onus is on us to make processes that don’t substitute illegibility and efficacy for scale and growth that is rubbish. Any new technology that will make legible, previously illegible information has the potential for immense value creation and the challenge for us is to do that without being reductionists.
As it turns out, we’re increasingly making legible, information that resides in the material world to come up with solutions to this problem.
An example from the startup world to illustrate this.
Stitchfix, a startup that makes custom clothing that fits you and matches your taste has been profitable at scale for many years now. Their tagline says it all - “Personal Styling for Everybody”. Here’s a screenshot from their investor presentation I came across in this brilliant article that looks at the company’s unique proposition.
While the idea of “mass customization” has been around for decades, it is only now that we’ve developed the requisite technology to do it in an affordable manner.
X. Legiblizing the whole
What does good scaling look like? How do we make wholes legible?
There are at least two ways to scale well and both have the same end objective - enable a richer representation of underlying reality and not compromising on outcomes.
1. Doing things that don’t scale
In classic Silicon Valley fashion, this seemingly contrarian principle to achieving scale and success is the bedrock of the startup playbook endorsed by Y Combinator, the seed stage venture firm. The phrase “Do things that don’t scale” was first introduced by Paul Graham, in an essay filled with examples of successful companies using this principle to scale well.
The idea behind this maxim, though not spelt out explicitly is that of finding near-perfect legibility functions. As seen before, once a good legibility function is identified, scaling becomes easy. It is finding that legibility function that is hard and non-trivial. By asking founders to do things that don’t scale and providing extraordinary service to customers, he is effectively asking for the elimination of shortcuts and biased legibility functions that are prone to failure at scale.
Another example of doing things that don’t scale comes from Stripe and its recruiting playbook. Recruitment in organizations is arguably the most critical function. It is also a notoriously difficult process to scale as indicated by the plethora of recruiting tools and services available in the market and the near unanimous disdain for them. Outside of narrow domains like tech recruiting for well defined, specific roles, general purpose recruiting tools fail miserably at almost everything.
Why is recruiting so difficult to scale? Because it tries to make legible a person’s candidature whereas the relationship between the requirements of a role and a candidate’s suitability is complex and illegible. How do we scale recruiting without resorting to reductionist methods?
Stripe’s recruiting playbook exemplifies what it means to scale fast without compromising on quality of candidates. It is one of the most respected and sought after companies in the world.
Recruiting at stripe - Almost everyone is involved in recruiting, decisions to hire are made in a decentralized fashion and there is a massive amount of effort that is spent outside of regular channels for recruiting. Stripe, foremost, assumes that people they’ve recruited are 100% to be trusted and understand what it means to be an employee at Stripe. By relying on respective team members and managers to make hiring decisions, they forego the need for overly centralized recruiting practices and processes, allowing them to act fast on offers and admit a bigger pool of candidates than any other startup when interviewing prospects. Since everyone is responsible for recruiting in a spiritual sense, Stripe is always hiring, so to speak.
The big tech companies all rely on standardized recruiting practices - whiteboard interviews, algorithm and puzzle questions, centralized messaging through HR all of which favor a certain class of candidates over the rest. It is this imposition of legibility that leads to all kinds of adverse selection and wasted effort in recruiting. Stripe employees spend a lot of time outside of their work to look for potential candidates by being helpful to anyone on the hunt for a tech job. @patio11 is a classic example of such an employee. By doing so, these people have created a feedback loop that gives them access to the best talent available.
Doing things that don’t scale proves beneficial in the long-term as it avoids adverse selection and failure modes associated with imposition of legibility. This method of scaling works extremely well for Type A interactions where scale usually imposes legibility.
The second way to avoid failure modes associated with legibility is to take a platform approach based on modularization of the product or process.
Modularization is the idea of distilling a complex reality into its basic constituents, packaging it into a platform and allowing end users to choose the final product based on their requirements. This idea of modularization and mass customization has been around for decades and has rapidly caught on everywhere from industrial applications to consumer goods and software applications.
Amazon, Airbnb, Shopify, Spotify, Stripe and many other software companies are based on this idea of modularization and platform-as-a-service. Even semi-industrialized kitchens like Chipotle use this concept to allow for mass customization.
From the Wikipedia entry on modularization: “Modular design inherently combines the mass production advantages of standardization, since modularity is impossible without some level of standardization, (high volume normally equals low manufacturing costs) with those of customization.”
Good modularization is characterized by the following principles (taken from this software engineering course at the University of Zurich):
high cohesion - functions within a module should be closely related
loose coupling - modules should be loosely coupled with each other with minimal dependencies
information hiding - modules should expose only necessary functions while the rest remains hidden from the end user
This method of scaling works well for the Type B interactions where legibility enables scale. More on how technology can enable mass customization in the physical world in this McKinsey report.
XI. Killing Procrustes
Time to finish the story. Procrustes met his inevitable death at the hands of Theseus, a Greek hero who was passing through town. Fittingly, he was killed in the same manner as his numerous victims with his limbs chopped off after being tied to the bed.
Unfortunately, there aren’t any modern day heroes akin to Theseus who can help us with our Procrustean infestation. We do have entrepreneurs though, hacking away, one Procrustean bed at a time.