In a previous essay, I argued that all formal models of complex adaptive systems are necessarily incomplete. But this tells us nothing about our ability to use imperfect models as blueprints to control complex adaptive systems. After all, doesn’t an imperfect model simply imply that our control of the system will be imperfect? And maybe imperfect control is still a lot better than no control at all? Unfortunately things are not quite so simple. Control based on a necessarily imperfect model operates primarily by forcing a Procrustean simplification upon the complex adaptive system. As I argued in the summary essay, the territory becomes the map. Unfortunately, this transformation via simplification is also an incomplete one. It is more accurate to say that the territory almost becomes the map. Moreover the gap between the system and the control model gradually increases after the initial simplification. The “adaptive” element in the complex adaptive system causes the agents within the system and the system itself to gradually adapt to the control regime so as to make it less and less effective.

The idea that simplification and formalisation is the essence of modern control was introduced by James Scott in his seminal book ‘Seeing Like A State’. By attempting to make the system “legible” system managers “transformed the reality they presumed to observe, although never so thoroughly as to precisely fit the grid”. Rigid control based on any model, however complex, will simplify the system. The very act of making a system legible is a simplification. Some of the best examples of this phenomenon can be seen in the history of colonial intervention by bureaucratic colonial powers attempting to control their illegible colonial populations. Scott himself noted how colonialists in Asia and Africa, by demarcating tribes and ethnicities for the purpose of administration, engineered and accentuated hitherto vague cultural differences within the population1. The model of control was a blueprint for change, a self-fulfilling prophecy. As many researchers other than Scott have discovered, this pattern of “manufacturing what you cannot discover” was a common feature in colonial interventions2.

Why Procrustean Control?

In Greek mythology Procrustes was an inn-keeper who changed his guests to fit the bed (instead of changing his bed to fit his guests) by either stretching those who were too short or amputating those who were too tall. As Nassim Taleb notes3, “Modernity is a Procrustean bed, good or bad — a reduction of humans to what appears to be efficient and useful.” James Scott has shown how modern control is the act of fitting a Procrustean bed not just to humans but to all complex adaptive systems that human beings wish to control. But the metaphor of the Procrustean bed still leaves us with a question – why do system managers adopt the strategy of Procrustean control? What exactly is it about simplification that is so appealing?

To some extent, control itself is the aim of simplification. It is much easier to control a simplified legible system than it is to control a complex, illegible system. But for the most part the simplification of fitting the Procrustean bed is not an end in itself but a means to an end. In modern economic systems, this end is most often increased output or increased efficiency. For example, the simplification of German forests that James Scott analysed was primarily aimed at increasing the timber yield and productivity of the forest. In many cases, stability itself is the aim of the simplification but more often stability and increased efficiency are intertwined and complementary aims. For example, the suppression of fire in many forests was seen as desirable aim by itself but it certainly did not hurt that fire was seen as a wasteful and inefficient phenomenon that reduced the timber yield of the forest.

The Sledgehammer Beats The Scalpel

The first step in the process of simplifying the system is the simplification of the goal. Maximising economic well-being gets simplified to maximising GDP growth which in turn gets simplified to maximising aggregate demand and spending within the economy. The next step is best described as follows: Use a sledgehammer, not a scalpel. If you want to increase the yield of a forest, change the forest to a monoculture of trees of the species that has the highest yield. If you want to increase aggregate demand within the economy, the manager (i.e. the government) can simply increase spending. In social systems where reflexive positive-feedback processes abound, the ideal approach may even involve targeting a variable that is typically the effect of the goal rather than its cause. This is the essence of Greenspan Put monetary policy – asset prices typically go up when economies grow. But spending and aggregate demand also go up when asset prices go up via the wealth effect. By focusing on engineering an increase in asset prices, the central bank can set in motion a self-fulfilling process where increased asset prices lead to increased spending which in turn feeds back into increased asset prices.

A brute-force approach is sufficient to achieve effective control in the short run and it is of course easier to adopt a simpler model to mould the system to. But that is not all – the simpler the system is made and the more blunt the control strategy, the more effective control is in the short run. The simplification translates into increased efficiency, reduction in slack and diversity and thereby improved short term system performance. In other words, the sledgehammer beats the scalpel in the short run. In most cases, this process of Procrustean simplification is not even intentional. In an uncertain world, the often spectacular short-term success of Procrustean control ensures that it is preferred over other approaches. If a system manager were to test the Procrustean approach on an uncontrolled system against other options, he would almost certainly conclude that the Procrustean approach was the most effective.

How is Procrustean Control Different?

To understand the history of modern control, we also need to understand how the errors caused by Procrustean simplification are different from the errors caused by simply having a “wrong” model of the system. James Scott’s argument in ‘Seeing Like A State’ is often criticised as simply being a rehash of arguments made by Friedrich Hayek4 on the dangers of central planning and its inability to take into account the dispersed knowledge of all agents within the system. Scott’s argument, as described above, is very different from that of Hayek. The problem with modern control is not so much that the model does not represent reality but that reality adaptively fits itself to the model (imperfectly) and in the process fragilises itself. Still if we are only referring to some residual radical uncertainty created by the simplification, how different are the consequences compared to the Hayekian case? Nassim Taleb for example has shown how Procrustean simplification causes fragility by exposing the system to damage from consequential rare events with catastrophic consequences.

Addictive Initial Success Followed By Failure

But there is more to the effects of Procrustean control than the Hayekian or Talebian arguments. It is not just that control causes fragility. It is that the process of control sets in motion a systemic transformation that causes increasing fragility and is almost impossible to reverse without undergoing an interim systemic collapse. Almost all complex adaptive systems that are subjected to Procrustean control go through an uncannily similar progression of events that best resembles the process of drug addiction. And what causes this progression is the “adaptive” element in the complex adaptive system, such as gaming/innovation by agents within social systems and homeostatic mechanisms within biological systems.

I laid out this progression of the simplified system in a section titled the ‘Control Treadmill’ in the summary essay. What makes this progression fatal is not that Procrustean simplification fails in its task but that it succeeds spectacularly at first and then fails. The comparison of the system evolution to an increasing addiction is not just a vague analogy. Just like an addict, things get better to start with and then get steadily worse. The control process is characterised not by progressive ineffectiveness but by initial success followed by unavoidable failure. It is this progression from success to failure that is the primary reason for loss of system resilience. Fragility arises not from the failure of high-modernist control but from its misleading successes. For example, simplification of German forests initially caused timber yields to increase appreciably and it took almost a century for the systemic deterioration to manifest itself.

Not The Wrong Order But The Absence of Disorder

The Hayekian dispersed knowledge argument is essentially an argument of equilibrium, that the central planner causes the system to seek out the wrong equilibrium. It also holds out the hope that if the system were not centrally planned or controlled, the dispersed spontaneous actions of the system agents would allow the system to achieve a better, more resilient equilibrium. This argument also leaves open the possibility that some day the central planner may be able to codify and gather sufficient knowledge so as to achieve effective control. Maybe someday financial market regulators and policymakers can gather all the data and information they need to control the financial system? But Scott’s argument and my extensions of his arguments see the problem not as the persistence of the wrong orderly state but the folly of trying to achieve perfect control and stamp out disorder5. Dispersed and small-scale action is preferable not for its orderly tendencies but its disorderly tendencies – its inability to even attempt anything so vain as to try and make the system legible.

Procrustean Control And The Subversion of Modern Science

Scott’s arguments and my extensions of his arguments are not only applicable to hubristic control projects managed by the state. Scott was correct in identifying this phenomenon with modernity – the state is the most devastating purveyor of control projects but the modern corporation operates in much the same manner. If anything Scott underestimates the breadth of domains to which his ideas are applicable. The pervasive nature of modern control subverts and diminishes the effectiveness of the modern scientific method.

Universal Conclusions Impossible

Modern control has left very few domains untouched. Due to the extent to which the system domain is modified by the control process, many scientific observations and experiments are in reality operating upon and observing this transformed world. For example, any hypothesis of medical and nutritional science is not evaluated according to its results upon the “natural” human body but upon the modern controlled human body. This problem affects not only the system controller but also the agents within the controlled system. For example, is the superiority of any trading strategy during the Greenspan Put era an empirical “truth” or is it simply the result of the controller’s stabilising policies? Does IQ really tell us anything about human intelligence or does it simply measure the extent to which our intelligence is aligned to the systematic, codified world of modern control?

No Independent Events

As Duncan Foley has observed6, “the philosophical bedrock of statistical inference is the assumption that the future will be like the past, so that we can use repeated observations of past events to infer at least some features of the future”. Even if we reconcile ourselves to the possibility that our scientific conclusions are relevant only within the modern controlled domain, the evolution of the complex adaptive system throughout the process of control means that past events are no longer independent in the statistical sense. The pattern of the treadmill of Procrustean control, initial success followed by increasing failure and fragility, means that errors in the control process are no longer independent and random. The system does not simply perform in a steadily suboptimal fashion or fail in a similar manner every once in a while. The magnitude of errors and the control slippage steadily increases as the system adapts to the control intervention. For example, the era of asset-price focused monetary policy hasn’t just been a series of unconnected recessions and crises but a progression where asset prices become increasingly disconnected from GDP which in turn becomes increasingly disconnected from employment and wages7.

All Goals Are Intermediate Goals

The modern scientific method is often viewed as a means to an end, where the ends are chosen by those applying it. The idea that focusing on intermediate goals is bound to fail due to the subversion of control by agent adaptation is a well-researched phenomenon – Campbell’s Law, Goodhart’s Law and the Lucas critique. But what these ideas ignore is the progression from initial success to eventual failure. The control regime is gamed via agent adaptation but only eventually.

In fact no matter how well the scientific method is executed, every control goal is an intermediate goal. There are no clear final goals within any complex adaptive system. At best we can reduce our goals to a fuzzy understanding. For example, we may seek to achieve economic well-being but at best we can reduce this down to a number of concrete intermediate goals such as the minimisation of unemployment and the maximisation of GDP growth. When the sledgehammer of Procrustean control is used to achieve one intermediate goal, system performance starts to deteriorate by all other barometers other than the single goal that is the object of control. And eventually performance deteriorates even when judged by the single intermediate goal that was the focus of the control regime. So the Procrustean control regime of asset-price obsessed monetary policy results in a deterioration in GDP growth, employment, inequality and technological innovation even as asset prices remain at high levels. And eventually even asset prices falter as the cost of the control regime rises to unsustainable levels. Similarly, the result of education policies aimed at maximising exam performance (which is the focus of Campbell’s Law) leads to a deterioration in performance when viewed against any barometer but exam performance and eventually even a deterioration in exam performance.

The Failure of Control As Proof of Success?

The initial success of the Procrustean control regime validates it. But what about the eventual fragility and systemic deterioration? Unfortunately this deterioration too can be interpreted in many different ways. For example, was the decision to let Lehman Brothers fail in 2008 a sign that the financial system is inherently fragile? Or was it a sign that adaptation to successive rounds of stabilisation and bailouts had rendered the system hopelessly fragile? When a psychiatric patient’s symptoms deteriorate upon withdrawal of medication, is it a sign that the patient is “naturally” ill and needs medication? Or is it a sign that prolonged periods of medication has led to the patient becoming addicted?8 In other words, even when the system begins to falter, the pattern of failure can be used as the justification for diametrically opposed explanations. What the evolutionary viewpoint would view as the failure of the control regime, the local analysis of the scientific method will view as the proof that more control is needed.

Modern Science Is Procrustean

It is tempting to end with the conclusion that Procrustean control subverts scientific understanding. But many have argued that the Procrustean approach is not a subversion but the very essence of modern science itself. As Theodore Porter notes9, “the progress of experimental science is the increasing ability to make and use new things, and at the same time to transform the world that science purports to describe”. The successes and the failures of modern Procrustean control are inescapably intertwined with the successes and failures of modern science.

Some of the most spectacular successes of modern science are successes of Procrustean simplification. For example, the spectacular success of antibiotics is inseparable from its transformation of the human microbiome into a barren landscape. By erasing the microbial population of the human body, antibiotics essentially succeed by adopting a “scorched earth” sledgehammer strategy10. The initial intervention upon an unadapted bacterial population succeeded in spectacular fashion. But as the harmful bacteria have adapted to the sledgehammer, the efficacy of antibiotics have diminished and the ecosystem of the modern human body has been left in an almost irreversibly fragile condition.

Note: The next section contains various case studies that illustrate this uncannily similar pattern of failure caused by Procrustean control.

What Is A Complex Adaptive System?

The first question that then needs to be answered is: What is a complex adaptive system? David Krakauer defines complex systems as “systems that don’t yield compact forms of representation”1. In other words a complex system cannot be described by a simple set of equations. Why would this be the case? As Krakauer notes, it is the “adaptive” nature of these systems that leads to this intractability. Agents within the system respond to each set of environmental conditions within a complex adaptive system with a different set of responses and the number of such environments and their corresponding agent responses that need to be accounted for to construct an accurate model of the system is simply too large. But is this simply a problem of impracticality? Could we, at least in theory, construct a model that takes into account all possible environmental conditions and all possible agent behaviours? Although some scientists may argue that such an approach is theoretically possible, there is ample evidence that the critical “adaptive” component of some complex adaptive systems may in fact be unmodelable. There is no better example of this than the problems faced by the economist Hyman Minsky in formalising many of his most important ideas.

Limits Of Formal Modelling Techniques: Hyman Minsky’s Insights

Hyman Minsky’s core insight, that “stability is destabilising”2 is simple enough to explain. Long periods of economic and financial stability encourage a shift from conservative financing (“hedge finance”) to more speculative financing strategies (“Ponzi finance”) and an increase in systemic leverage. The fragile economic system is then prone to be tipped over into a recession by even an ordinary shock. At least in theory, this is an eminently modelable process. But Minsky’s insights go far beyond this description of the business cycle. Minsky was a strong advocate of strong action by the central bank and the government to stabilise the business cycle. But he was also well aware of the long-term damage inflicted by such a regime where all disturbances were snuffed out at source – the build-up of financial “innovation” designed to take advantage of this implicit protection and subvert the regulatory framework, the descent into crony capitalism and the growing fragility of a private-investment driven economy3. This understanding that was also reflected in his fundamental reform proposals4. Minsky also appreciated that the short-run cycle from hedge finance to Ponzi finance does not repeat itself in the same manner across business cycles. The long-arc of stabilised cycles is itself a disequilibrium process (a sort of disequilibrium super-cycle) where performance in each cycle deteriorates compared to the last one. An increasing amount of stabilisation needs to be applied in each cycle in order to achieve poorer results compared to the previous cycle.

It is this long-run adaptation by firms, banks and other agents within the economy that is so resistant to being modelled in a formal manner. This problem was identified by Duncan Foley5 who observed that a model that can “represent change only as the quantitative variation of a given set of variables” cannot capture the essence of the adaptive process which is the generation of novelty, “the emergence of qualitatively new phenomena”. This is not a problem that can be solved simply by adding new “regimes” to the model and the problem is not just that any realistic model rapidly becomes unwieldy. There is a much deeper and more profound problem. For example, how can any quantitative representation of this process account for financial products that have not yet been invented? How can a quantitative model in 1990 account for the role of the yet-to-be-invented CDO in fuelling the financial crisis of 2008? And even more importantly, how can any such model account for the ability of firms, banks and other agents to generate such “innovation”?

Simple Systems vs Complex Systems

When analysed against the criteria of whether all possible environments and their corresponding agent responses can be specified in advance, there are many systems we seek to control that are not complex in nature. One of the most vivid examples comes from the field of war. Although conventional, hierarchical command-and-control warfare is not effective in complex environments such as jungles, mountains and urban settings, it is still effective in executing clear missions in simple environments such as the desert6. Another example comes from the field of computing where the modern computing environment faces many of the same control issues precisely because it has transformed from its prior monolithic, relatively simple form into an interdependent and networked collection of adaptive components7. In fact, complex adaptive systems can often be identified by observing the critical role that disorderly processes play in maintaining system resilience. For example, the disorderly and often unpredictable nature of flooding is a vital factor in maintaining the productivity and resilience of many complex river-floodplain systems. But the same rarely holds for simpler systems such as small temperate streams8.

Misidentifying A Complex System As A Simple System

Although some systems are clearly not as complex as others are, system managers must be careful not to misidentify a complex system as a simple system. A common error is to ignore factors at higher hierarchical levels of the system that subvert control efforts at lower levels. Many systems that are relatively simple when viewed in isolation become complex when viewed within the context of a larger system. Desert terrains may represent a simple battlefield but even simple victories on the battlefield can be negated by the unintended consequences on the larger socio-political system9. Even the simplest technological system often exhibits complex behaviour when viewed within the context of the social and organisational system within which it resides.

Another common problem is that systems that appear to be minimally adaptive in the short run may be strongly adaptive in the long run often with catastrophic consequences for the control effort. An excellent example is the program to control the river Rhine during the early part of the 19th century. In response to persistent floods, the Rhine was narrowed and straightened with its embankments raised and its arteries cut off. It took one and a half centuries for the adaptive transformation in the system to manifest itself as systemic fragility when the Rhine experienced four ‘100-year floods’ in 12 years between 1983 and 199410.

The recession that followed the 2008 financial crisis was the most severe that the developed world had experienced since the Great Depression of the 1930s. Even five years later in 2013, unemployment remains far higher than pre-crisis levels. The most popular explanations for why the Great Recession of 2008 has proven to be so long-lived focuses the two themes. One explanation blames the recession on inadequate demand and attributes the persistence of unemployment to the failure of either the central bank or the government in not providing enough stimulus. Monetary economists blame the Fed for not providing more monetary stimulus and Keynesian economists blame the government for not providing enough fiscal stimulus. The second explanation blames the slow recovery on the severity of the financial crisis, noting that historically most recoveries that follow financial crises tend to be slow and gradual1.

But neither of these explanations can account for the most striking characteristic that defines the post-2008 economic recovery in the United States – the fact that when viewed through the lens of corporate profits the recession is already long over. The benefits of the current economic recovery have flown disproportionately towards corporate profits with wages and employment lagging far behind. Since 2008 we have seen a robust recovery in corporate profits, a modest recovery in GDP and a feeble recovery in employment and wages.

And this disconnect between profits, GDP and unemployment is not a new phenomenon. Since at least 1990, each successive recovery in employment has proven to be weaker than the previous recovery, with profits and GDP recovering faster than employment and wages. The ratio of corporate profits to GDP has now reached an all-time high whereas the ratio of wages to GDP continues to fall. Whatever ails the US economy is not just a cyclical phenomenon but a structural problem and worryingly the condition of the economy is deteriorating with each business cycle.

Progressive Deterioration In Pace Of Recovery

Corporate Profits/GDP At All-Time High

Wages/GDP At All-Time Low

Some attribute the declining share of wages in GDP to the rapid nature of recent technological change and innovation and in particular to the increasingly automated nature of many economic activities and advances in labour-saving technology. This is an intuitively appealing explanation. But it neglects that fact that productivity growth across the developed world has been anaemic for the last four decades, a phenomenon that Tyler Cowen has called ‘The Great Stagnation’2. How can we suffer from stagnant productivity growth and technological unemployment at the same time?

Product Innovation And Process Innovation

To understand how technological unemployment can occur during a period of stagnant productivity growth, we need to analyse the nature of innovation more closely. Innovation in the economy can be broadly divided into two categories: product innovation i.e. creating new consumer goods and services, and process innovation i.e. creating cheaper, more efficient ways to produce existing consumer goods and services.

The dynamics of product innovation are very different from those of process innovation. Product innovation typically entails a high degree of uncertainty and trial-and-error. Process innovation on the other hand is typically more amenable to being analysed within a conventional risk-reward framework. Even when investments in process improvements are large, outcomes are less uncertain than they are in product innovation.

Most economists deny the possibility of technological unemployment on the grounds that our wants are unlimited. Therefore if automation and artificial intelligence displace workers from one part of the economy, they will find employment elsewhere in another sector. Indeed, something similar has been occurring in our economy for at least the past two hundred years. The most significant example of this reorganisation was the automation of the agricultural sector in the first half of the twentieth century which resulted in a significant shift of the workforce away from agriculture into manufacturing and services.

But this simplistic innovation ignores two critical nuances of how this reorganisation has taken place in reality. First, economies reorganise back to full employment in the long run not via increased consumption of existing goods and services but via consumption of entirely new goods and services. When process innovation drives down costs and prices, consumers may initially consume more of the existing basket of goods and services. But sooner or later, diminishing marginal utility and satiation sets in and consumption of existing goods and services stops increasing. At this point, new product innovation is needed to provide an entirely new basket of goods and services. Due to the uncertain nature of new product innovation, it is by no means a certainty that the economy will be able to provide the requisite amount of product innovation to maintain full employment, especially in the short run. It is precisely this combination of high process innovation and low product innovation that has made technological unemployment possible even in a period of stagnant productivity growth. In the short run, productivity growth can be sustained by process innovation. However, long run productivity growth requires bursts of product innovation along with persistent process innovation.

The second nuance that is often latched upon by advocates of technological unemployment is the question: what if there are no more economic activities that human beings can perform any better than machines and robots? To answer this question we need to take a much closer look at the nature of innovation in automation and artificial intelligence over the last two hundred years, a task that I undertake later in this essay. But first we need to analyse exactly why process innovation has been so rapid and product innovation so sluggish in recent times.

Product Innovation Vs Process Innovation: New Entrants Vs Incumbent Firms

Due to the uncertain nature of new product innovation, incumbent firms rarely excel in it unless compelled to do so by the competitive pressure exerted upon them by new entrants. Even in industries where entry of new firms is free and unrestricted, incumbent firms rarely come up with disruptive new products. Historically, new entrants to an industry have been responsible for most disruptive product innovation. On the other hand, process innovation being a lower-risk activity is typically introduced by established incumbent firms and not new entrants3.

Incumbent firms have very little incentive to invest significant resources in risky initiatives that aim to displace their existing cash-cow businesses. In many instances, they may face resistance not only from internal departments that feel threatened by the potential success of new products but also from customers who are reluctant to embrace disruptive change. In fact a great deal of the uncertainty in new product innovation arises from the fact that it is almost never driven by the customer. As the old adage goes, customers rarely know what they want unless they see it, a principle embodied by Steve Jobs’ time at Apple Computers. New product innovation requires constant trial and error and most of these trials are bound to fail. Incumbent firms are, quite rationally, primarily focused on protecting their existing source of profits and minimising the risk of failure rather than undertaking speculative risks where the odds of failure are greater than the odds of success.

In the absence of new firm entry, even a competitive industry with many players will focus on process innovation and cost reduction and avoid any potentially disruptive product innovation. When incumbent firms do undertake product innovation, they do when their existing source of super-normal profits is threatened by disruptive products from new entrants. In an environment where product innovation is high, not undertaking new product initiatives is the riskier option. Simply protecting existing revenue streams rarely works out. Despite this, many incumbent firms are rarely able to respond effectively to new entrants, primarily due to organisational rigidity4. New entrants on the other hand face a different set of incentives. Having no existing profits to protect, the lure of capturing such super-normal profits drives their actions far more than the much larger possibility of failure.

In other words, unless incumbent firms face the threat of failure due to the entry of new firms, product innovation is unlikely to be robust. The role of failure in fostering product innovation has sometimes been called the ‘invisible foot’5 of capitalism. Process innovation being a lower-risk endeavour is not dependent upon the threat of failure. Simply instituting a regime where firm owners and employees are incentivised to seek higher profits is sufficient to encourage process innovation. In other words, the positive incentives of Adam Smith’s invisible hand are sufficient to give us a high level of process innovation but disruptive product innovation requires the negative incentives of the fear of failure i.e. an invisible foot of dynamic competition from new entrants.

Process and Product Innovation in the Post WW2 United States

Viewed through the lens of product vs process innovation, the Post WW2 economic history of the United States can be divided into two parts6. The first half from 1945 till the 70s was a period when the pace of both product and process innovation was slow. As Alexander Field has shown, much of the productivity growth in the aftermath of the war came from exploiting product innovation that had already taken place during the 1930s7. The damage done to the industrial base of the rest of the developed world meant that there was very little competition for American goods from foreign manufacturers. Most large American firms were also largely insulated from strong shareholder pressure to improve profitability. This combination of low import competition, low rate of entry by new firms and weak shareholder pressure meant that there was very little process innovation or cost control. It is not a coincidence that many view the 1950s and 1960s as a golden age of economic growth and stability. It was essentially a period when neither firm owners, managers or workers felt the threat of failure or even had the incentive to improve efficiency or control costs. It was a period of stability for all, masses and classes alike.

The second half from the 1980s onwards has been characterised by an equally low, maybe lower, rate of product innovation. However, process innovation has accelerated significantly. This is the period often referred to as the neoliberal era. The neoliberal revolution is often viewed as a shift towards more deregulated and free markets but this interpretation is a misleading half-truth. In reality, the neoliberal turn in the developed world was characterised by a dramatic resurgence in shareholders asserting their rights over incumbent firms along with a series of initiatives that sought to mimic the positive incentive structure of markets in domains that had hitherto not been subject to such incentive pressures. However although the invisible hand was unshackled, the invisible foot was left in an even more crippled state than before. Deregulation and privatisation often simply replaced staid monopolies with equally conservative yet shareholder-focused oligopolies. Licensing and patent regimes became steadily more dysfunctional and prevented the entry of smaller firms. The regulatory burden also only served to protect large incumbent firms against new entrants. The earlier regime of stability for all had been transformed into a regime of stability for the classes and instability for the masses.

The Greenspan Put And The Great Stagnation

The neo-liberal economic model amplified the control of shareholders over incumbent firms but it diminished the disruptive competition faced by them, thus incentivising incumbent firms to accelerate process innovation and neglect riskier product innovation. It is obvious how barriers to entry such as licensing requirements promoted this shift. But a much more important driver of this shift was the shift in monetary policy that took place during the so-called ‘Great Moderation’.

The conventional wisdom views the Great Moderation as the golden era of monetary policy when the opaque discretionary policies of the past were replaced by rational rule-based policy. All recessions and unemployment that were the result of a shortfall in demand were countered with monetary stimulus and fiscal policy was deemed to have no role in macroeconomic stabilisation. Deregulated financial markets enabled the impact of monetary policy decisions to flow through to the real economy more effectively than it had in the past. At least until the 2008 financial crisis, almost all mainstream economists agreed that demand management was a task best left to a technocratic central bank to manage. Unfortunately, this account of monetary policy bears very little similarity to the actual policy conducted during the Great Moderation.

Monetary policy during the Greenspan era was based on one simple thumb rule: “support asset prices and the rest will take care of itself”. The best example of this doctrine was found early in Greenspan’s tenure when the 1987 stock market crash was countered with a massive monetary stimulus based simply on the fear of a potential real-economy recession. An even more egregious example of this doctrine was the Fed’s reaction to the failure of the hedge fund Long Term Capital Management(LTCM) in 1998 when it cut rates to support the markets at a time when the real economy was booming. This doctrine of monetary policy is often referred to as the ‘Greenspan Put’ which refers to the impact that this policy had on financial markets and the banking sector. Market participants could assume that any fall in asset prices would be countered with monetary stimulus thus providing the “free” protection resembling a put option. But as disastrous as the impact of the Greenspan Put was on the financial economy, it paled next to the impact the policy had on the real economy.

If you protect a system from the effects of any particular risk, actors within the system will take on more of the protected risk assuming rationally that the system manager (in this case the Fed) will protect them. The Greenspan Put regime drove down the risk of being exposed to broad macroeconomic market risk. Market participants rationally took on more macroeconomic asset-price risk and substituted for the risk they had been relieved of by the Fed with more leverage. Conventional portfolio theory views the asset allocation decision as one of choosing the split between the risk-free asset and the market portfolio. But when the risk of the market portfolio is suppressed, the decision changes to choosing how much to borrow against the market portfolio.

And this is exactly what the financial sector proceeded to do. Far from being a neutral channel of monetary policy from the Fed to the real economy, the deregulated yet too-big-to-fail financial sector that was also protected from new entrants realigned itself to take on macroeconomic risk by lending to housing and large established firms. The attractiveness of this strategy meant that banks shunned lending exposed to non-macroeconomic idiosyncratic risks such as lending to small businesses or new firms. The Greenspan Put doctrine thus triggered a realignment away from the idiosyncratic risk-taking that lies at the heart of disruptive new product innovation. But there’s more to it than just the financial market effect. The doctrine also encouraged firms in the real economy to become as bank-like as possible. No firm took advantage of the new regime like General Electric(GE) did. GE under Jack Welch transformed itself into an industrial firm whose profits came largely due to its financial arm, GE Capital which lent to its industrial customers (amongst others). So successful was this transformation that by the time the 2008 crisis hit, GE had also become too-big-to-fail thanks to GE Capital and was found to be eligible for a bailout.

GE also exemplified how the new regime of amplified shareholder control over firms and the Greenspan Put not only discouraged product innovation but encouraged process innovation. Faced with the constant pressure of meeting quarterly earnings targets from shareholders (most of whom were themselves holding diversified market portfolios), the only innovation for which there was any appetite was low-risk process innovation that could cut costs and enable higher leverage. Once the flab had been eliminated from most large firms but the initial burst of private equity buyouts, leveraged buyouts and hostile takeovers (or the threat of hostile shareholder action against management), process innovation was the logical next step in reducing costs further.

Although the monetary policy of the Great Moderation caused stagnant innovation and stagnant wages, it also provided the temporary palliative medicine that maintained full employment and economic growth by fuelling an increase in household leverage. Despite the absence of real wage growth, households were able to increase consumption by borrowing from banks eager to lend against macroeconomic risk-bearing collateral i.e. housing.

Automation, Artificial Intelligence and Process Innovation

Almost all the important technologies of automation and artificial intelligence are technologies of process innovation that help us produce the same basket of consumer goods and services in a cheaper, more efficient manner. Although the recent wave of labour-displacing technology such as manufacturing robotics and machine learning may seem new, they are in fact simply the latest in a long chain of such technologies going back to the machines in the early eighteenth century that attracted the ire of the original Luddites. Despite the popular perception, replacing labour with a machine rarely involves constructing a machine or an artificial intelligence that can do exactly what the human worker can do. Instead, automation involves redesigning the domain itself such that the work can now be done by a machine, or increasingly now, an algorithm. An excellent example of this was seen during the industrialisation and mechanisation of agriculture where many fruits and vegetables were modified such that they could be harvested by a machine without causing damage, as opposed to constructing a machine that could harvest the existing fruits and vegetables in as careful a manner as humans would. In other words, automation and labour displacement does not require that the machine or robot be able to do exactly what the human being does, or that the machine be able to do the task in the same manner as the human worker. Indeed historically this has rarely been the case. There is very little similarity between the brute-force method by which a computer plays chess and the intuitive manner in which a human grandmaster plays chess. But because the uncertainty within the domain of a game of chess is bounded the computer is able to match or beat the human. As the uncertainty in the domain increases, expert humans tend to outperform the brute-force automated approach. This is true even for more complex games than chess such as Go, let alone more complex and ambiguous real-world domains.

In order to deal with the residual uncertainty created by the routinisation of the domain, human beings are often left the task of monitoring and managing these automated systems. But these are not the only jobs that humans have performed in our increasingly automated economy over the last one hundred years. Many routine jobs that have provided avenues of mass employment during the twentieth century have typically been jobs requiring the use of human sensory and motor skills, skills that have proven hardest to automate. This phenomenon is known as ‘Moravec’s Paradox’8 named after the artificial intelligence researcher Hans Moravec who observed that those skills we typically identify with intelligence (e.g. rational decision making) tend to be the skills that are easiest to replicate via an artificial intelligence (a combination of data and algorithms). But those skills that even a baby possesses, such as the ability to move around complex environments and pick up a variety of objects, tend to be the hardest to replicate in a robot. In a way some of what separates from the machines is what unites us with the animals. So despite the increasing automation of manufacturing and services, humans retained jobs that required routine sensory and motor skills. Even the most automated supply chain required truck drivers and restaurants still required waiting staff. Even within the routinised domain, humans and machines were complements not substitutes.

But this era where Moravec’s Paradox shielded many routine jobs from being automated away is rapidly coming to an end. A critical element of the recent success on this front has again been the acceptance that matching or beating human performance levels does not require an artificial replication of the human method. The robotic vacuum cleaner does not operate like a human vacuum cleaner. Nor does the self-driving car drive like a human driver. But in many cases, a simple algorithm combined with large amounts of sensory data is enough to match or beat the average human’s performance.

The implications of this are simple yet frightening. The routine jobs that provided avenues of mass employment in the twentieth century are increasingly a thing of the past. It is very easy to envisage a future where the vast majority of work in large, established parts of the economy is almost fully automated with human beings simply performing the role of monitoring and managing the system during extraordinary circumstances.

The Future Of Human Employment: The Near-Automated Economy

If robots and machines can perform every economic activity as well as humans can, then the very aim of full employment seems obsolete. The problem then would appear to be one of distribution, of how the returns from technological capital are distributed within the economy.

As of now however, we are nowhere near this point. What has proved amenable to being performed by machines are activities that can be performed adequately by some combination of an algorithmic process and data. This obviously leaves out artistic endeavours or economic activities that require the generation of novelty (such as creating new products). But it also leaves out any human activity that requires artisanal expertise. At least as of now, human creativity still seems to be an essential component of many economic activities. Many of you will object that I am placing too much faith in the inherent superiority of human creativity. And who is to say that the same results achieved by intuitive human expertise in “creative” activities won’t be matched by a machine sooner or later?

This objection arguing for the inevitable perfection of artificial intelligence technology ignores the fact that the imperfect human contribution itself has a positive economic value in many cases. Even if a machine creates works of art that are objectively equivalent to those of Jackson Pollock, does any one doubt that Pollock’s paintings will still command a significant premium? The same could be argued for much of the organic, local food industry whose appeal arises in part simply from the fact that the food is not produced by machines in a distant corner of the world.

This does not imply that the entire “human economy” will reorganise itself to resemble a pre-Industrial Revolution artisanal craft economy. The collapse in the cost of goods and services produced by the automated economy will mean that purely artisanal products will remain a luxury good. What will provide the majority of human employment will be the ‘near-automated’ economy where a small, yet critical, proportion of creative human endeavour is combined with a largely automated process. And it is this near-automated economy that has received the greatest fillip from the last ten years of the algorithmic revolution and the collapse in the economies of scale and scope that it has brought about (in stark contrast to automation through the twentieth century which led to increased economies of scale and scope).

In many economic and artistic fields, near-automation has been a reality for a while now. Even the smallest writers, artists and musicians can now create niche products and produce and distribute them efficiently across the world. They are able to do this precisely because although the design and initial creation remains a human activity, the replication and distribution are automated. 3-D printing and robots that can operate in smaller spaces are just a couple of the many technologies that enable a small-scale manufacturer to compete with much larger firms. The largely algorithmic nature of the production process means that a small manufacturer can simply create the design code for a product and hire a manufacturer in China to produce even small batches of this product without suffering a significant cost disadvantage vis-a-vis large firms.

Even if there remain many activities that human beings can and will continue to perform, there is no denying that large parts of the economy will be almost fully automated. It is imperative that the economic returns from these sectors are not unfairly concentrated in the hands of a few. Else, human job options in the economy of the future will be largely restricted to being servants or court jesters to the rich. It is easy to advocate simple measures of taxation and redistribution to redress the inequality of the current system. But redistribution will do nothing to restore the innovative dynamism of the economy. And moreover, we can do a lot better. The same measures that will help us transition to an innovative, resilient near-automated economy will also give us a fairer society with more equal opportunities than our current neo-feudal economic system.

Transitioning To The Near-Automated Economy

Principles

Earlier I described the 1950s and 1960s as a period of “stability for all” and the neoliberal era as a period of “stability for the classes and instability for the masses”. The essence of my policy proposals is simple: All economic actors must be subject to the disruptive, disorderly forces of competition i.e. disorder for all, masses and classes alike. Many of us would prefer that we somehow turn back the clock and recreate the imagined stable utopia of the 50s and 60s. Even if this were feasible, constructing an economy where firms and all their stakeholders are provided with perfect stability is a recipe for stasis and stagnation. It is a solution that, at best, enables us to share s static economic pie in a more equitable manner. Moreover, even this outcome of a static pie is not certain. Dynamic competitive tension and the threat of failure due to disruptive innovation at the level of the firm is not just important to expand the size of the economic pie. It also helps maintain the resilience of the system against unexpected shocks by either enabling the system to maintain critical functionality or to rapidly reorganise to an effective state after systemic collapse.

Disorder also lies at the heart of distributive justice. The collective bargaining power of labour and the share of the economic output that flows to labour increases when the firms that employ them are individually fragile and subject to the constant threat of failure. A competitive free enterprise economy does not equate to the theoretical ideal of perfect competition where no firm earns super-normal profits (also known as rents). It is the lure of super-normal profits that drives the entry of new firms into an industry. No venture capitalist has ever funded a startup that tried to make a market rent-free. A successful new entrant does not extinguish rents, it captures them. What matters is that the incumbent firm earning super-normal profits is subject to the constant threat of losing these profits. The fact that an incumbent firm makes significant super-normal profits does not imply that entrepreneurs and capital as a class do the same. For every successful firm, there are many who fail.

Proposals

Monetary/Fiscal Policy

Apart from the damage that asset price obsessed monetary policy does to the economy’s ability to innovate, doctrines like the Greenspan Put also act as a transfer of wealth from the society as a whole to some of its richest members. Contrary to the popular perception of monetary policy as a neutral macroeconomic policy with a minimal impact on income distribution, real-world monetary policy that focuses on propping up asset prices overwhelmingly favours the rich for one obvious reason that even central banks have now begun to admit9. Most assets are owned by the rich(see table below10). The asset-less poor rarely have a direct stake in the performance of the S&P500. The idea that supporting asset prices is the best way to support the wider economy is essentially a form of trickle-down economics (or as Will Rogers put it: “money was all appropriated for the top in hopes that it would trickle down to the needy.”).

Instead of macroeconomic policy being directed at asset markets, interventions during economic crises should be carried out through measures that provide an equal benefit to all. The best example of such a policy is a ‘helicopter drop’11 where the government simply prints and sends a sum of money to each individual.

Safety Net

Protection against economic risk must focus on providing a safety net for the people and not a hammock for the firms that employ them. No firm should be too big or too important to fail.

Dismantle Entry Barriers

Entry barriers such as licensing, patent regimes and onerous regulatory regimes need to be comprehensively and systematically dismantled. Most of the damage done by entry barriers is not felt by venture-capital funded startups. It is felt by individuals and small businesses in much more mundane parts of the economy – hairdressers, small food businesses etc. In many sectors of the economy, an employee who is made redundant has no option but to seek employment in another one of the large incumbent firms. Even if he wants to, starting his own business is often not an option. This needs to change, especially at a time when collapsing economies of scale and scope make entrepreneurship a viable option in more and more sectors of the economy.

No Safe Assets

A popular explanation for the 2008 financial crisis is that assets earlier deemed safe (such as triple-A rated mortgage backed securities) turned out to be risky assets in reality. Proposed solutions to this problem tend to recommend the creation of truly safe assets by the government to meet this demand for safety. This idea must be rejected. Instead, all assets must be made unsafe. There is no reason why the government should provide any economic actor with the means to preserve purchasing power without taking on meaningful economic risk. Manufacturing safe assets would simply allow those who captured the benefits of the Greenspan Put era to preserve their gains at the expense of the rest of the economy.

Education For The Near-Automated Economy

Since the Industrial Revolution, the economy has become increasingly automated, routinised and algorithmic in nature. But the often incomplete nature of this routinisation combined with Moravec’s Paradox (the difficulty of replicating human sensory and motor skills in a machine) has meant that human labour was typically a complement of the automated system in performing what were still largely routine activities. Deskilling as this process of automation often was(see graph below12), it nevertheless provided avenues of mass employment providing a reasonable compensation, albeit doing routine work. Understandably, 20th century education evolved to meet the demand for workers and employees who could perform in a systematic and algorithmic manner alongside the automated system. Creativity, innovation and deep domain expertise were only required of a few employees in a firm. For the rest, it was far more important to be reliable and efficient. Even many so-called skilled employees in managerial roles were required to align themselves to the bureaucratic functioning of the organisation more than they were required to display any creative insight.

As explained earlier, the era of mass employment in routine work is rapidly coming to an end as Moravec’s Paradox is overcome and more systems achieve near-perfect automation. However, our educational system still remains wedded to the routinised industrial model. In the near-automated economy humans are complements to the machine not within the automated domain but outside it. Their role does not require them to contribute the algorithmic knowledge and actions that the machine can provide. The human worker needs to provide the deep expertise, creative insight and innovation that the machine cannot provide.

It is beyond the scope of this essay to offer a detailed analysis of just how an educational system can prepare its students to meet this challenge. Instead I will just offer a couple of observations that are often ignored in the otherwise comprehensive literature on this topic. First, the process of learning for expertise is itself a disorderly, non-linear path where failure upon facing problems that lie beyond one’s competence is often more effective in stimulating learning than the steady imbibing of facts and techniques followed by the tackling of well-defined, familiar problems13.

Second, performance in many near-automated activities requires us to integrate artificial and human intelligence while still enabling the human to achieve deep, intuitive domain expertise. The most notable successes in this endeavour have come in activities where the fundamental essence of the domain as experienced by the human has remained amenable to intuitive expertise. For example, combined teams of a computer and a human often perform much better at chess than humans or computers alone can14. The fundamental reason why a human Grandmaster remains a valuable component in such a team is that the domain remains relatively intuitive. The chess board remains exactly the same as it has always been.

Unfortunately, this is not the case in many automated domains where, as I earlier described, the process of automation has modified the domain itself in often dramatic fashion to make the domain more amenable to algorithmic control. For example, airline pilots now fly complex computerised machines that provide little intuitive feedback. Financial market traders who used simple models and traded on physical trading floors now monitor the performance of complex black-box models and high-frequency trading algorithms. Therefore, even in domains where the algorithmic complexity of the domain is now unavoidable, students may be better off learning within less-automated domains where deep intuitive expertise may be more effectively achieved.

Conclusion

Innovation and the generation of novelty is a fundamentally disorderly process in all complex adaptive systems – biological, ecological or economic. When we manage a system to achieve perfect order and stability, we end up with stasis and fragility. But this does not imply that we need to embrace chaos and constant failure. To achieve a resilient macroeconomic system, we only need to embrace a little bit of disorder – in our firms, our financial markets and in ourselves.

Like other complex adaptive systems, the primary problem faced by biological systems is the need to explore for new innovations while preserving the old well-adapted functionality until something new and better can be found. This process also needs to be undertaken in an efficient manner with minimal resource usage. Through the metaphor of the fitness landscape (see diagram below) where higher peaks represent fitter populations, the problem can be restated as follows: how does an evolving population move from a hill such as point A to a higher peak such as point B when incremental changes will move it into a valley with lower fitness than point A?1 How can the system evolve in a transformative and radical manner to create what Andreas Wagner2 has called ‘game-changers’ such as photosynthesis and complex nervous systems?

Although the task appears insurmountable, the conflict between robustness, innovability and efficiency disappears when we allows ourselves to consider a few key characteristics of real-life genotype-phenotype mapping:

• Micro-fragility: Phenotypic robustness does not depend upon genotypic robustness. There is ample evidence that phenotypes are fairly robust with respect to both genetic mutations as well as environmental change3. Phenotypic robustness by definition implies that the phenotype does not change with most mutations in the genotype. This would seem to rule out the sort of phenotypic variability/evolvability that is required for transformative change – after all, how can new phenotypes be explored if genetic alterations usually have no impact on the phenotype?
• Distributed Robustness: Most biological systems contain a much larger number of genotypes than phenotypes. Even key functions in biological systems can be maintained in many different ways4. This robustness is distributed from the presence of diverse multi-functional components with partial functional overlap rather than the presence of redundant copies of each gene, a property known as degeneracy56.
• Cryptic Variation: Genotypes with the same phenotype form networks and most changes in such genotype networks are “nearly neutral” and are only weakly selected for. Movement across this network allows the accumulation of cryptic, underground variation7 that serves as a reservoir that enables transformative phenotypic change.

Robustness and innovability are therefore not only consistent with each other, robustness is in fact a precondition that enables transformative complex change. The distributed nature of the system’s robustness enables such a system to be maintained in a near-optimal manner without requiring an excessive usage of resources.

Unfortunately, the reality of the flood management program on the river Kosi has turned out to be anything but beneficial. The culmination of the failure of the program was the 2008 Bihar flood4 which was one of the most disastrous floods in the history of the state. So what went wrong? Was this just a result of an extraordinary natural event? Most certainly not – As Dinesh Mishra notes5, the Kosi carried only 1/7th of the capacity of the embankments in 2008 and at various points of time since the 50s, the river had carried far greater quantities of water without causing anywhere near the damage it caused in 2008. The system was thus unable to withstand even modest adverse shocks after prolonged stabilisation.

So what caused this loss of system resilience? As Dinesh Mishra explains6:

By building embankments on either side of a river and trying to confine it to its channel, its heavy silt and sand load is made to settle within the embanked area itself, raising the river bed and the flood water level. The embankments too are therefore raised progressively until a limit is reached when it is no longer possible to do so. The population of the surrounding areas is then at the mercy of an unstable river with a dangerous flood water level, which could any day flow over or make a disastrous breach.

As expected, the eventual breach was catastrophic – the course of the Kosi moved more than 120 kilometres eastwards in a matter of weeks. In the absence of the embankments, such a dramatic shift would have taken decades. With the passage of time, a progressively greater degree of resources were required to maintain system stability and the eventual failure was a catastrophic one rather than a moderate one. The stabilisation did not merely substitute a series of regular moderately damaging outcomes for an occasional catastrophic outcome (although this alone would be a cause for concern if a catastrophic outcome was capable of triggering systemic collapse). The stabilisation transformed the system into a state where eventually even minor and frequently observed disturbances would trigger a catastrophic outcome.

When faced with the possibility of a catastrophic outcome, the managing agency has two choices, neither of which are attractive. Either it can continue to stabilise the system using ever-increasing resources in an effort to avoid the catastrophic outcome. But this option must only be followed if the managing agency has infinite resources or if there is some absolute limit to this vicious cycle of cost escalation that is within the resource capabilities of the agency. Or it can allow the catastrophic outcome to occur in an effort to restore the system to its unstabilised state. But this option risks systemic collapse – it is not just the unprecedented nature of the outcome that we have to fear, but the fact that the adaptive agents of the complex system may have lost the ability to deal with even the occasional moderate failures that the unstabilised system would throw up.

For example, in the pre-embankment era when the Kosi was allowed to meander and change course in a natural manner, the villagers on its banks had a deep
understanding7 of the river’s patterns and its vagaries. The floods sustained the fertility of the soil and ensured that groundwater resources were plentiful. This is not to deny that the Kosi caused damage but because the people had adapted to its regular flooding patterns, systemic damage only occured during the proverbial 100-year flood. This highlights an important lesson in complex adaptive systems: The impact of disturbances cannot be analysed in isolation to the adaptive capacities of the agents in the system. If disturbances are regular and predictable, agents will likely be adapted to them and conversely, prolonged periods of stability will render agents vulnerable to even the smallest disturbance.

The targeted pest developed resistance to the pesticide used and often staged a dramatic resurgence in an environment where the pesticide has often also eliminated its natural enemies. The damage done to pests’ predator species meant that the resurgence is accompanied by an outbreak of many other pests that had not been a significant threat to the crop prior to the beginning of the pesticide regime. Paradoxically, the system that started with a diverse array of crops and a few significant pests was left at the end with a monoculture of crops and a diverse array of pests. Predictably, an ever-increasing dose of pesticides need to be applied. Nicaraguan farmers who needed to spray once or twice during the season in the 50s found themselves needing to spray their crops twice a week by the 80s and eventually cotton had to be abandoned.

“The healthy heart dances, while the dying organ can merely march.” Ary Goldberger.

It would seem obvious that a healthy heart should maintain a constant and steady heart rate, like clockwork. Conversely, a diseased heart would presumably be characterised by an irregular heartbeat. Until recently, expert medical opinion mirrored this seemingly obvious intuition. In 1989, Ary Goldberger and his colleagues showed1 that this conventional wisdom was not only mistaken, it was an exact inversion of reality. The time intervals between the beats of a healthy heart varied constantly in an unpredictable manner. It was in fact the diseased heart that exhibited perfect predictability and regularity, with the intervals between beats sometimes being identical. A healthy state is characterised not by stability or perfect predictability but by variability and unpredictability. In the diagram below showing the heart records of three patients, it is the relatively irregular heart rate of record B that is the only healthy one. The regular patterns of A and C are of patients suffering from severe heart failure.

But just as perfect order and predictability is pathological, so is constant disorder and excessive disturbance. Just as the regular heart-rate patterns of panels A and C signify a heart on the verge of failure, so does the excessively irregular heart-rate pattern of panel D below.