This post started with my frustration at seeing a lot of smart, well intentioned people getting trapped. It originally bore the sensational headline “Capitalism Makes Us All Slaves to Efficiency,” but there were enough reasonable ideas in here that I decided to tone it down a bit ;)
A framework I’ve been using to diagnose patterns of failure I have observed across a variety of domains and scales is a sort of systems thinking: within metrics driven, self-optimizing systems, systemic or structural failure might trap even rational actors. Instead of attributing failure to the poor decision making of an individual within a system, we can identify properties of the system they are acting within that might tend towards a particular failure mode.
The basic failure mode I’ve been observing recently is a tendency to get trapped in local maxima because it is easier to identify and measure metrics for short term wins than for long term investments.
i. A few examples of optimization leading to local maxima
The Entrepreneurial State: Debunking Private vs. Public Sector Myths. The private sector is practically incapable of funding research projects whose results don’t return profits in under 5 years.
Overpowered Metrics Eat Underspecified Goals. Organizations that attempt to scale by using metrics to align a large number individuals and teams towards a single goal flounder (even when individual agents successfully hit metrics), if metrics are not perfectly aligned with the broader goal.
A Whole New World. On an engineering team, investing in infrastructure improvement is nearly impossible to justify whenever the time to realize benefits is longer the performance review / KPI cycle. (cf. the disincentive for a politician to undertake any project that is unfinishable before their re-election term).
Waiting for Superman. In K-12 education, it’s nearly impossible to scale metrics of teacher efficacy without very short term measurement of student performance on standardized tests that don’t map very well to the long term value of education.
On Liberal Arts Education. Once prestigious liberal arts universities are Whartonizing and trending towards vocational training for white collar information work / middle management. We’re losing sight of the goal of equipping our citizenry with the skills to participate in a democracy.
E Unibus Pluram / Panem et Circenses. The consumer internet, like all other mass media before it, uses attention as it’s best proxy measurement for customer value. As a result, distribution channels and content tend towards baby photos and sensational clickbait.
ii. Proxy metrics
In each of these cases, I think the system optimizes for a short term metric that is an inaccurate proxy for the actual goal of the system for one or more of a number of reasons.
We optimize for proxy metrics, because we don’t have a more accurate way to measure goals. In education, we seem at best capable of measuring test scores, not actual aptitude. In media, we use time spent as a proxy for customer value.
We optimize for proxy metrics, because we require tighter feedback loops. Tight feedback loops are excellent tools for optimizing efficiency (eg, of understood production processes, operations, or distribution funnels). But if we’re dogmatic in adhering to a fixed feedback loop periodicity, we preclude ourselves from undertaking larger scale projects (the space program that reached the moon does not get funded by private markets, an engineer can never justify certain classes of infrastructure investment, etc).
We optimize for proxy metrics, because we prefer predictability and are risk averse. Tight feedback loops can be safer because they hedge against potentially catastrophic failure modes (this is why we have fixed term lengths for elected officials, regular feedback on baseline standards for schools / teachers / students). But feedback loops that are overly tight can lead to complacency and stagnation. Silicon Valley venture capitalists develop the lean startup “release early, release often” mantra as a way to de-risk investments and avoid investing significant capital without gauging market demand, resulting in narrowing / more incremental innovation. Individual creatives might bias towards smaller projects to de-risk the possibility of investing months or years in work that never gets released to customers (eg, singles, blog posts, or features vs albums / long form / epics / infrastructure / more ambitious projects).
iii. Flywheel feedback loops
Metric feedback loops clearly lead to optimization on the production side of systems and markets. But, the most efficiently self-optimizing systems are those with compounding feedback loops on the consumption side. These happen when use of a product begets more use of a product, thus coopting consumers into alignment with the producer’s goal metric. The result is a flywheel effect.
When entire industries optimize to achieve a flywheel effect, it can result in epidemic. When the media industry optimizes around the production and distribution of content designed to trigger dopamine response (newsfeeds of baby photos, fake news), or when the food industry similarly optimizes around empty calories (sugar), they’re exploiting the same biological factors that lead to addiction, and they result health epidemics similar to those that result from drug addiction.
iv. Metrics driven systems are not agnostic
Something I think prevents us from fully acknowledging (and then working to correct) some of these systemic failures is a shirking of responsibility by the designers and operators of the systems, writing off failures as non-issues via the argument that all of the actors within the system are making choices driven by (often rational) free will.
This argument breaks down because the designers of the system should be culpable not only for actions they take within the system (for example, as distributors), but also for the natural tendencies of the system. Eg, a supermarket is not blameless for selling popular products if the natural equilibrium state after ‘agnostic’ optimization is selling only popular, unhealthy products like Oreos and Coca Cola.
v. Escape via external rule changes
In failing systems like these, it’s really hard to find a solution, because often all of the actors within the system are correctly optimizing given the rules of the system.
One example I can think of where we actually seem to have extracted ourselves from a local maxima situation is cigarettes: we’ve drastically curbed usage of a highly addictive and profitable product through some combination of taxes and propaganda. Both taxes and a Surgeon General’s Warning label are rule changes to the system that were mandated by an external actor (the government).
It’s difficult to place all of our faith in these sort of solutions, however, because we cannot trust that the government will always be strictly external to the system: when some of the actors within a system have enough at stake, they may hold influence over the government in the form of lobbies, campaign financing, book deals, etc.
vi. Escape from within
Given this caveat about the lobbying power of actors within the system, and given the efficacy of self-optimizing systems in doing harm, it would be perhaps more resilient to look for ways to exploit metrics within existing systems to course correct them.
In the past, I’ve written very optimistically about escaping failing systems from within: this was essentially my thinking behind a soup restaurant as a solution to nutrition and health care epidemics.
The challenge of trying to design a solution within a cancerous self-optimizing system is that the solution is often just LESS, eg, less media. Or, fewer calories. Studies dating back to the 1930s have demonstrated that:
[W]hen the ad libitum food intake of mice and rats was reduced by 30 to 60 percent, the average life span and the maximal life span (the mean survival of the longest-lived decile) increased by similar amounts. In contrast, rats with nearly unrestricted caloric intake (92 percent of the average unrestricted intake) that were kept lean with exercise and weighed about 40 percent less than sedentary control rats with the same caloric intake had an increase in the average life span but not in the maximal life span.
Even with pretty clear evidence that has existed for nearly a century, there hasn’t really been any significant push towards solutions that leverage this kind of information. This is because (a) generally it is difficult to get consumers to consume LESS, and (b) it’s even harder to build a sustainable business on LESS, because a business that succeeds in selling LESS renders itself obsolete. The whole capitalism experiment presupposes never ending growth, MORE.
v. Conclusion / diversification
I don’t think the way out of these local maxima is setting a 0% discount rate, advocating “we all eat porridge,” and only valuing the future.
Likewise, I don’t think the answer is completely abandoning metrics and self-optimizing systems through a complete structural overhaul.
There’s a lot to love about metrics driven systems: they’re fairly meritocratic, they reward intelligence and hard work, they align (at least in some way) the incentives of producers and consumers, and they accelerate really quickly once they start working.
But, using this framework to understand the systemic failure at the root of a lot of wide scale social problems has made me a bit less optimistic that there are clear ways to correct these problems from within (eg, soup restaurants). So, I might diversify my bets by investing a little more in external / rule changing / infrastructural solutions than I would have a few years ago.