The first time I came across Curtis' thesis in HyperNormalisation hypernormalisation, I thought he'd gone off the deep end. He claims that, starting in the 70's, politicians and capitalists have given up on the messiness of the real world and have instead substituted and sustained a simpler world of their own design. I'd deeply admired Century of the Self century-of-the-self, but HyperNormalisation just came across like a basket of conspiracy theories cobbled together and served up as social critique.

I was also reading Baudrillard's Simulacra and Simulation simulacra-and-simulationaround the same time and found myself, again, irked by puffery. This time, it was from the gratuitous application of terms borrowed from technical fields citation-needed, which seemed to aim more at creating an impression of scientific technique than at conducting any sort of rigorous analysis of modern scientific technique / its effects in social reality (possibly cf: plastikworter). I would not be surprised, however, if this was a deliberate (if annoying) stylistic choice in service of advancing his "post-modern" argument.11. Apparently, some people were so irked by writing like this they wrote an entire book about it fashionable-nonsense.

I will concede in both cases, however, that there is probably a tiny kernel of truth buried deep, deep underneath the overburden of pretentious terms and over-exaggeration. We humans rely on models of the world in order to be able to live effectively in it. There is not such a thing as an un-mediated reality, and our representations of the world become less discerning as the horizon we have to deal with expands outwards (the analysis in Public Opinion public-opinion comes in useful here).

On the state level, such representations—of people, of resources—have to be systematised in a grid of the state's making in order to make legible everything within its confines, to levy taxes and effect conscription (so as to wage wars). The aspects of the world that appear on such representations (e.g. census, cadastral surveys, maps) are highly restricted, privileging those that are of interest to the state, regardless of the fidelity of such representations to the lived reality of the everyperson. seeing-like-a-state

This reductionist tendency is one I have observed in practice of data science. As a younger data scientist, I was baffled whenever more senior colleagues would reach automatically for the absolute simplest tools—linear regression, logistic regression, those techniques you learn about in the first week of introductory machine learning—when conducting analyses of business data. There are so many models out there that would be able to give us much more accurate predictions, so why are we still sticking to these dinky little elementary tools?

Well, I soon learned that outside the confines of academia, people cared very much about the interpretability of such models, often over and above the error metrics academicians so emphasized. People wanted to know more about why the model made a certain prediction—for example, why did the model predict that this customer would keep their subscription, and that one would churn? Knowing the why would enable the business to take action, for example, by ameliorating whatever the root causes of customer churn was.

On the other hand, complex models, while usually more accurate, were less interpretable, with fewer easy explanations of "X factor increases Y outcome". This is known as Bonini's paradox, summed up nicely in The Breakdown of Nations:

Oblivious of their own inconsistency, some of our modernists point out that small states had an easier time of it, being so insignificant in size and population. But this is exactly it! Because they were small, they could not only solve their problems better than their large counterparts; they could do so without the assistance of such brilliant minds as Marx, Schacht, Cripps, or Keynes. They did not need to deal with aggregates which, in large countries, even statisticians can only guess, and whose meaning even experts do not always understand. They could at all times see their economy at their feet—open, surveyable, manageable. They did not need to operate on assumptions which no one on earth can prove, however great his learning and many his degrees.

— breakdown-of-nationsp. 152

Such analyses were not only less actionable to people around the company, they could also be perceived as being less trustworthy—nobody likes a black box with the potential of yielding arbitrarily wild predictions. And so the simplest of models have stayed fast in the toolboxes of data scientists through the years, despite the inadequacy of such simple tools and explanations for the task of modelling the large complex world we find ourselves having to deal with, and despite all the recent advances and hype surrounding deep learning.

The attitude I've just described—the preference for interpretability over accuracy—sits, in some odd ways, on the enlightened end of the spectrum. More insidious is when people unquestioningly hail the application of such models as some magical silver bullet. This applies to all models, and not just those over-hyped specimens from the AI/ML world.

"All models are wrong, but some are useful," is an adage I wish more people understood. I'd also hasten to add, "in limited contexts." Oft quoted is the George Box version:

... all models are approximations. Essentially, all models are wrong, but some are useful. However, the approximate nature of the model must always be borne in mind....

— George Box

Timely and germane is the example of how wildfires have been handled in California.

tl;dr: Firefighters used a model of fire spread that worked only under certain assumptions of "light surface fuel" which failed as fuel loads in forests built up over time due to fire suppression policies

Flammable as it looked, even forests mismanaged like that patch burned until recently in the historical way, at low severity along the forest floor. As a result, the entire field of wildfire science—including every modeling tool with which firefighters make life-or-death decisions and society structures itself in fire-prone areas—is based on that kind of fire behavior. The core mathematics of this science date to the early 1970s, when a Forest Service researcher named Richard Rothermel used small laboratory fires to produce equations expressing the relationship between wind speed, ground slope, and how fast a fire spreads. Rothermel knew his approach worked properly only for wildfire in light surface fuel like that in his lab—and failed to capture what happened when blazes got into treetops and jumped crown-to-crown. But these so-called Rothermel spread equations were applicable to so many wildfires that the Forest Service quickly developed paper-and-pencil ways for firefighters to plug in numbers for wind and slope angle and make reasonable guesses about how fast and in which direction a fire might spread—in a single heading, on a straight line. Eventually that modeling framework was run on cumbersome supercomputers, then on handheld calculators. In the early 1990s, PC-based software finally allowed firefighters to predict fire spread in two dimensions on a map.

1. Modellers usually know the limitations of their models up front—any model requires the acceptance of certain assumptions, the negligence of certain terms. Such limitations need to be communicated clearly, especially to non-technical end-users, lest the model be over-extended in its application. Sometimes, though, when a model becomes too useful, it is no longer within the control of the modeller as to where it gets applied—and that's where the trouble starts. (cf small-is-beautifulp. 196)
2. There is an understandable, if sometimes damnable universalising tendency that undergirds our culture. All models are built on certain sets of assumptions, assumptions that are inevitably an encoding of the modeller's own culture and experiences. Even in physics—the models we developed first, and indeed are the most sure of, are those of the space- and time-scales that we humans have the most direct experience of. To deny the limited applicability of models (and especially those outside of the "hard" sciences) takes ignorance, arrogance, avarice, sloth, or some combination thereof. In California, as in Australia, the accumulated wisdom of the natives in managing their dry and fire-prone homelands was ignored in favour of the forest management techniques the colonizing Europeans brought from their comparatively wetter climes, to horrifying consequences.

And so, from Curtis: yes, state and bureaucratic machinations necessitate simplifying assumptions for their efficient workings. It doesn't mean there is a grand conspiracy starting only in the 70's that's engulfed us all today; if anything, we've been hypernormalised since the birth of the state. Possibly before. From Baudrillard: what he calls the precession of simulacra, I call the precession of models.