Machines and Chaos

9 min readNov 20, 2022

Is it possible to systematically benefit from the unpredictable? In his book Antifragile, Nassim Nicholas Taleb argues that biological systems improve from chaos while mechanical systems are harmed by it. In this blog post, we’ll explore how different biological and mechanical systems respond to chaos and seek a deeper understanding of Taleb’s claims.

A table from Taleb’s book comparing the mechanical to the biological with respect to various attributes — Table 2 from Antifragile

What is a machine?

An art sculpture that is a mechanical interpretation of a peacock — “PFAUENMASCHINE (PEACOCK MACHINE)”, 1982 by Rebecca Horn

To differentiate biological from mechanical, we must define “machine”. Is it something created by humans to make their lives easier? Or does it exist naturally?

Do you think the dictionary gets it right?

machine: a mechanically, electrically, or electronically operated device for performing a task (Merriam-Webster)

machine: an apparatus consisting of interrelated parts with separate functions, used in the performance of some kind of work (Dictionary.com)

While we typically consider machines as cold and hard, the concept goes beyond these narrow definitions. Computers, corporations, political systems (you’re just a victim of the machine, man!), or even living organisms themselves could be considered machines depending on how you define this elusive concept.

What does it mean to be living?

Maybe we’d have better luck defining a “biological system”. Put differently, we seek the meaning of Life. Merriam-Webster makes a weak effort:

life: the quality that distinguishes a vital and functional being from a dead body

Wikipedia tries a little harder:

life: a quality that distinguishes matter that has biological processes, such as signaling and self-sustaining processes, from that which does not, and is defined by the capacity for growth, reaction to stimuli, metabolism, energy transformation, and reproduction.

Three solar hula dancer figurines on a shelf in the sun — These self-sustaining machines dance as long as there’s sunlight

There are machines capable of each of these processes in some capacity. A traffic light can “signal”. A solar hula dancer can “self-sustain”. There are Self-building homes can “grow”. Factorio allows construction of self-expanding factories. Reality lags fiction but NASA has a project exploring this idea. A car can “metabolize” gasoline. Manmade self-replication is rare if not nonexistent but a few projects exist.

While there are parallels between mechanical and biological systems, nature does seem to have mastered the above skills in a way humanity has not. And, as we’ll see, the dictionaries have left chaos out of the picture.

What is chaos?

Before we attempt to explore system interactions with chaos, we should finally also attempt to understand chaos itself. This time, the Oxford English Dictionary offers a clear and thought-provoking definition.

chaos: unpredictable, apparently random behavior exhibited by a dynamical system governed by deterministic laws, typically considered to consist of frequent instability, aperiodicity, and the occurrence of widely diverging outcomes corresponding to small changes in the initial conditions of the system

In other words, chaos is a name for behavior we can’t predict. If you believe that the universe is governed by deterministic physical laws (and thus, that we have no free-will) then true randomness does not exist. For example: the weather should be predictable, but the equations that best describe it are so sensitive to immeasurable changes in initial conditions that “the flap of a butterflies wings could cause a hurricane halfway around the world”. Systems that exhibit this property are called nonlinear.

While chaos is perhaps only a perceived phenomenon due to our inability to predict complex behaviors, it’s real to us. If it can’t be predicted, we can only treat it as random. Thus, our machines must adapt to it.

Biological Systems and Chaos

Neither of the prior definitions of “life” mention evolution or hormesis, which Taleb uses as the key distinguishing factors in his dichotomy. Hormesis is the process by which a system is strengthened by exposure to stressors. For example, exposure to small amounts of bacteria can improve the immune system by giving it a chance to practice identifying and fighting foreign invaders. Taleb’s observation that biological systems love chaos and are thus “antifragile” is based on this concept.

The antifragility of biological systems is one of their most fascinating properties. As described by Taleb, this special benefit from chaos relies on the fragility of smaller component parts and the natural selection or regrowth that results from their failure. This concept extends to non-living systems as well. The fragility of individual companies supports the antifragility of the economy, for example. As entrepreneurs try new ideas the whole system learns from their failures. But does it also extend to machines?

Mechanical Systems and Chaos

Taleb would say that mechanical systems do not similarly benefit from chaos because mechanical systems are designed only to function in predictable environments. When they are exposed to chaotic conditions, they are unable to function properly. However, some mechanical systems do benefit from chaos. A self-winding mechanical watch, for example, relies on the chaotic movements of the wearer to wind itself up and keep running.

The internals of a self-winding watch — Self-winding watches capture energy from the movement of the wearer

One could argue that this isn’t truly chaos. The wrist will generally move within a predictable range of forces and patterns. And, if the watch is dropped or jostled too much, it can be damaged or broken. However, isn’t the same also true of biological systems? All living things eventually die of overexposure to chaos. Even small imbalances such as eating too much sugar and skipping exercise can wreak havoc on the body. As we’ve seen, the effect of chaos on systems of both types varies with the “amount” and “type” of chaos.

Categorizing Interactions with Chaos

Taleb characterizes the outcome of system interactions with chaos into three categories: they can be harmed by it (Fragile), they can remain indifferent (Robust), or they can benefit from it (Antifragile). I propose a fourth category for systems that harness chaos but are not repeatedly improved by it. We can’t call these systems antifragile, but calling them robust seems like a disservice. Let’s explore various systems in these categories.

Machines that are harmed by chaos (Taleb: Fragile)

A rusty engine of some sort in disrepair

This is the largest category. Harm can manifest as complete failure or deterioration and happens by processes such as corrosion, erosion, fatigue, or overload for physical systems. These are often called failure modes. Information systems are victim to corruption of state as a broad generalization of a different set of failure modes.

Machines in this category include:

The cold and hard variety
Most electronics and software (turn it off and on, reset to valid state)
Systems with complex rules that are hard or slow to change e.g. Bureaucracies

Despite that “chaos” is difficult to quantify and the value* of its “effect” is generally subjective, for each type of machine we will imagine a typical dose-response curve for chaos exposure.

A graph depicting a downward trend in effect as a response to increasing chaos — Fragile systems always respond adversely to chaos

*value is also an elusive concept

Machines that survive chaos (Taleb: Robust)

As previously discussed, all machines are harmed by sufficient chaos, but some have a higher tolerance. This is facilitated particularly effectively via self-healing, which, compared to hormesis, only returns the system back to its original state in response to stressors.

Machines in this category include:

Fault-tolerant systems that isolate failures and transfer control away from and restart or replace failed components. For example, resilient distributed software systems or avionics.
Biosynthetic materials that don’t replicate hormesis
Many machines and systems that are designed with redundancy but do not similarly “self-heal” failed components as software systems do.

A graph depicting a flat, then downward trend in effect as a response to increasing chaos — Robust systems can tolerate some chaos

Machines that harness chaos (Taleb does not distinguish this category)

A wall of lava lamps in a corporate office — Cloudflare uses a video feed of this wall of lava lamps as an entropy source for their cryptographic algorithms

These machines utilize chaos to perform their primary function, but are not repeatedly improved nor harmed by it.

Self-winding mechanical watch, WITT Pendulum, inertial generator
Random number generators and any system that depends on them. There are countless uses for randomness including games, cryptography (the backbone of internet commerce), simulation, statistical sampling, and art.

A graph depicting effect as a response to chaos increasing from a low value, then flattening at a high value, then decreasing again as chaos increases — Some systems need chaos to function

Machines that are improved by chaos (Taleb: Antifragile)

A beat up yet cool looking guitar on display — This new “relic” guitar fresh out of the Fender Custom Shop sells for $6K

In this special category, machines are improved by chaotic environments. This is frequently by hormesis, natural selection, learning, and two other effects I will call mellowing and enlivening. Sometimes only a percieved improvement occurs with aging. This is called nostalgia.

Mellowing and enlivening are interesting mechanisms by which chaos reduces or increases some dimensions of deviation to a positive effect. For example, the introduction of chaos as white noise can both mask distractions or ease the maddening effects of silence. Humans and other systems dislike both too little or too much deviation. Too little results in monotony and dullness. Too much results in disarray, clutter, and cacaphony. The tendency of the brain to calibrate to stimuli in this way is called neural adaptation.

Example machines in this category include:

Systems where humans are a component

Systems where the failure of individual components strengthens the collective. Examples: biology and nature, free-market economies, anything subject to evolution and natural selection
Systems that adapt quickly to and learn from failure e.g. startups (“move fast and break things”)
Social media and entertainment: someone will make something viral. More users increases the odds.
Interactive art installations like the Austin Graffiti Park
Some “vintage” systems such as musical instruments. Examples: electric guitar, audio equipment. “Improvement” may be a subjective human aesthetic, or due to mellowing or enlivening. Whether the superior sound quality of vinyl is objectively measurable or the result of nostalgia is debatable.

Systems where humans are not a direct component

Reinforcement learning algorithms such as AlphaStar which learn from their interactions with an unpredictable environment. Other machine learning algorithms such as Generative Adversarial Networks and Stable Diffusion can also benefit from chaos.
Materials which benefit from annealing and computer simulations which benefit from simulated annealing
Netflix infrastructure*. They devised a set of “simian army” components (the original was the famous Chaos Monkey) which introduce random failures into their system to ensure it can sufficiently adapt. This is called Chaos Engineering.
Biosynthetic materials that replicate hormesis
Some zippers run smoother over time, possibly due to mellowing. Engines also exhibit this effect during their early life but are otherwise highly fragile.

*This is only to the extent that the infrastructure can respond positively to failures in the absence of human intervention. Merely surviving failures would only characterize robustness. “Doctor Monkey” proactively scans for and terminates unhealthy instances, inducing an automatic natural-selection effect in addition to allowing the human operators to respond and improve the system, as an example. Admittedly, it is a weak one.

A graph depicting an initial slight negative response to low chaos, then a positive one, and then a large negative one as chaos increases — Antifragile systems have a “hormetic window” in which chaos has a positive effect. Low chaos values can cause atrophy.

Many human-operated systems are also antifragile due to the influence of their operators, rather than their design. As an example, the following characteristics of such systems improve in response to stressors:

Stability — failures and bugs result in design improvements
Security — breaches and vulnerabilities result in design improvements
Efficiency — performance issues and bottlenecks result in design improvements

Can you think of systems that I missed that fall into any of these categories? Leave them in the comments.

Conclusion

Is Taleb’s dichotomy so simple? We’ve observed that biological and mechanical systems can exhibit similar responses across all three of his categories when the effect is measured against increasing vs. fixed amounts of chaos. Yet, we saw biological systems have mastered adaptation to chaos in a way most manmade systems have not. Perhaps our antifragile machines of the future will utilize hormesis, learning, growth, and reproduction as well as their biological counterparts to tackle new challenges. In the mean time, thanks to Taleb, engineers of the present at least have some new words and a new understanding for describing and incorporating these phenomena.

Give me a follow here and on Twitter if you’d like to see more of this content! https://twitter.com/jakespracher

I post about engineering, programming, and startups. I recently quit my Apple job to work for myself. I’m currently doing fractional CTO consulting while exploring startup ideas. I also produce music.