A Complex Systems Koan and the hunt for the W Brian Arthur of Neural Networks
‘Practical men who believe themselves to be quite exempt from any intellectual influence…Madmen in authority, who hear voices in the air, are distilling in their frenzy some academic scribbler of a few years back’. Keynes may have his tongue in cheek here but a consistent source of alpha is knowing which scribblers are becoming voices in the air. As the tech industry shifts from the mechanistic world of traditional software to the much more mysterious one of the neural networks the scribblers will be complex systems theorists. Working out how they become voices in the air will lead to fun and profit.
Mechanistic vs Complex Models
The Mechanistic World of Bits
Since the Scientific Revolution the most useful model of the world has been mechanistic: we understand the world by breaking it down into parts and working out how they fit together. Science is the investigative version of this model, progressing through an ever more granular understanding of parts and fits. Technology is the practical version of it, progressing by arranging the parts into every more useful products.
Of course, all models are wrong but some are useful. In the real, physical, world the mechanistic model has always had to contend with the edge cases, frictions, and scarcity. But software created a ‘world of bits’ that is perfectly mechanistic. Software is a set of logic statements: if it does not do what it is meant to do it is buggy; if your system has become too complex to understand you have kludge or tech debt; there is no excuse for friction and edge cases are a knowledge problem.
The scribbler who first recognised the economic consequences of a perfectly mechanistic, infinitely scalable, world of bits was W Brian Arthur. In his 1983 paper (which found fame as an HBR article in 1996) Arthur contrasted the diminishing returns at scale of the industrial economy, where the frictions, scarcity, etc of the physical world lead to price equilibrium and oligopoly, with the increasing returns at scale of the digital economy in which winners would dominate.
Arthur’s article introduced many of the key concepts, from network effects and lock-in to the asymmetric advantages of small, highly capable teams against big incumbents, that are the theory of tech investing and building to this day. Indeed, everything from burn multiples of commodifying your compliments can be seen as a footnote to the theory of increasing returns. But it all rests on a mechanistic world of bits and neural networks change all that.
The Complex World of LLMs
In the second half of the 20th century it became increasingly clear that some of the most interesting phenomena, from the weather to the stock market, could not be understood mechanistically no matter how expert we became. The study of these phenomena became known as complex systems theory.
Key to the definition of a complex system is that its outputs are based on interactions between parts of the system with itself and with its environment. This means they cannot be described mechanistically: even knowing the parts perfectly will not tell you exactly what it will output. You may be able to know the range of potential outputs but you can only speak about them probabilistically and then observe the system in action. As the cyberneticians put it, every complex system is a set of black boxes interacting.
Neural networks, of course, meet this definition. They are very different from the logic machines of traditional software: to know the weights is not to know, precisely at least, the output; and the smallest changes in prompts can result in massive changes in outputs. And as the application layer replaces programmes with agents driven by models we will move from a mechanistic world of bits to a complex world of models.
The Complex Systems Koan
This complex world of models needs its own W Brian Arthur and I am very much not him. Instead I have cribbed a set of comparisons from my own favourite scribbler on complex systems, Frijtof Capra. In his account you can compare mechanistic and systems thinking by three areas of emphasis:
Object > Relationship
Quantity > Quality
Measurement > Mapping
I don’t quite know how to use this yet, but I repeat it to myself like a Koan and every business I see, founder or investor I meet, I try to work out where their emphasis is. My hope is that this will produce wisdom or at least allow me to recognise the W Brian Arthur of models when he or she appears (there is an irony here that Arthur was both an economist and a complex systems theorist, and wrote his famous paper at the Santa Fe Institute, the home of complex systems research, apparently with Cormac McCarthy as his editor).
Behavioural Change and Why this Matters
It is always tempting at this point to become rather hand wavy about dramatic and far reaching change. And while I do think there will be dramatic and far reaching change I will try and tie this to what we know:
In the tech industry big companies are built (and overthrown) when entrepreneurs are able to use new technology to drive changes in pricing models, business models, and/or customer behaviour. New technology alone is not disruptive as we have seen with cloud and mobile, which, for the most part, favoured the Big Tech incumbents because it increased rather than changed user behaviour.
The shift from traditional software to neural networks is far more decisive. Not just in terms of capabilities but in terms of the way that we need to think how the technology works and is deployed. That leaves room for the kind of change from which new, dominant companies emerge.
If you are the W Brian Arthur of neural network economics, or know who he or she might be, please let me know.