👾 Econ 05 | The Services Hare and the Goods Tortoise?

In the last article I touched upon the fact that the historical expectation was for manufacturing to be automated first and cognitive areas such as professions like doctors and lawyers coming only much later; but it has turned out to be the reverse so far. Let’s explore this further to identify why this might be so.

It’s perhaps easy to, at least in hindsight, speculate why this made sense. Manufacturing is a physical phenomenon, it’s tangible, it’s visible. Thus it’s easy to spot functions that get automated. But there are many challenges when it comes to manufacturing. 

Any change to the physical setup, say of a manufacturing plant, will have to be tested in the real world before it can be assessed as viable, despite any amount of computer-aided simulation. In any case, there is the inevitable matter of high capital investment. If certain materials purchased as part of trying out such a setup don’t fit or work, they might have to be discarded as scrap, which imposes a cost and inhibits free experimentation.  

A related concern, when it comes to industrial environments, especially the more complex ones, is safety - primarily safety of the human operators working amid such setups. Our current set of human values and ethics do not permit us to, and we can agree rightly so, experiment with human life and limb before concluding that a newly devised industrial production setup is not safe enough. (You might say this goes without saying, but I think it’s worth making this explicit given it was not always so, and given this helps us demonstrate the contrast with the digital world.) 

The world of services has benefited from developments I shall present under two distinct categories of computing: conventional Software Engineering and of course the now more prevalent and fast-growing area of Artificial Intelligence.  (I’ve discussed this dichotomy in greater detail in my AI Concepts series, in particular in articles 01 and 04.)

Put simply, software is abstract by nature. This means, it doesn’t have the physical constraints of the physical world, it’s not bound by material limitations. Unlike the scenario of the industrial plant rehaul experiment mentioned above, it's a relatively trivial computational cost to try out new ways of solving a problem in the digital world. And it doesn’t have the dangers of rejigging an industrial plant, you won’t lose your fingers if the modified software crashes!

One really important difference between the physical world and the digital world is of course the (near) infinite replicability of a digital artifact - when you share an mp3 file with your friend, you still have it too! But you give your CD (or a cassette tape for those who remember!) to your friend, and you don’t have it any more!

This applies to software code as well. A software library once written can be easily reused in another application, whereas a particular physical engine as part of a manufacturing plant cannot - it of course has to be remade using the same components and raw materials as the original was, and to the same specification and with potentially onerous testing needs. 

This leads to a compounding enabler effect as each software component (libraries, frameworks) thus built can be reused in whole or part to build something on top of it, something more complex and composite, and thus potentially solving a bigger problem, or a set of problems. This standing-on-the-shoulders-of-giants effect can be achieved in a rapid and exponential manner because of the easy replicability of digital artifacts as mentioned. 

At this juncture it must be mentioned that this capability has been significantly enhanced by the ‘open source’ movement, which is a sort of a ‘creative-commons free market’ wherein contributors are able to reuse code written by others further accelerating this compounding enabler effect. (The currency of this free market is karma, but more about that another time!)

Again because of the nature of software development and computer use, thanks to connectivity-enablers such as the Internet, it’s possible in the digital world to work remotely and collaborate from anywhere. This is not the case when it comes to making goods by hand and machine tools.

Another happy development has been the exponential decrease in hardware costs. Processing power has doubled approximately every 2 years over several decades (Moore’s law), there has been a similar exponential growth in storage capacity (Kryder’s law), and indeed computer parts have become smaller (miniaturization). And hardware in general, for these reasons, has gotten continually cheaper.

So, enter AI! What may surprise some people is that the core ideas of machine learning and artificial intelligence go back to the 1950s, but it’s only in the last decade that we’ve had the hardware and computing power advanced enough and their costs affordable enough (just about!) for us to get AI working reasonably well to see the impacts of it as we’ve been documenting on Forward Future. 

In particular the large language models we have been discussing, which generate text to appear convincingly human-like in speech have only become viable because of these advances. (More broadly, this applies to the whole new field of Generative AI - where AI models generate not just text, but image, audio or video, or combinations thereof). 

But with advances in machine learning and artificial intelligence, coupled with decreasing hardware costs and increasing computational capabilities, will the all-singing all-dancing robots sneak ahead of their all-knowing all-thinking text-spitting cousins? We shall find out soon!

—

Subscribe to FF Daily to get the next article in this series delivered straight to your inbox.