DOS prompt to GUI, and back to prompt

My delusional parasitosis started slowly, but quickly accelerated to the point of being debilitating. I could tell it was a parasite, because it was consuming my energy. The main symptom was my vision was being distorted, and I was seeing a different reality -- like Alice Through the Looking Glass. For a while I thought it was toxoplasma gondii -- a brain parasite that is thought to be carried by cats, and found in their faeces, and suspected to cause mental illness. What took me a bit longer to realise was that it was delusional. I don't even have a cat, and I definitely hadn't been around their feces. Rather my delusional parasitosis was closely related to the ear worm, but distinctly different (unless you suffer from synesthesia, which I'd like to have, but don't): the idea worm.

The idea started as something surprisingly simple: Approximations were actually the most accurate way of describing the world. Now this might not seem like the kind of idea worm that would eat through your consciousness and distort your vision, but hang in with me. When I was growing up, I was taught that the perfect description of the world was a mathematical equation. The simpler, the more elegant the equation, the closer you were to true (think Kepler being able to describe the planets as an eliptical equation, think Einstein's E=mc2). An approximation was seen as a fudge.

Approximations have been around forever. In the 5th century BCE Thucydides in the History of the Peloponnesian War, describes the Athenian's working out how high their ladders needed to be to scale the walls of Platea (about 65km north west of Athens) by counting the number of bricks in the height of a wall. The general had several soldiers count the bricks and then used the mode (i.e the most frequent number) to determine the number of bricks high. He then multiplied the mode by the height of one brick. It worked and their ladders helped them scale the walls. A good enough approximation to solve a real life problem.

It's a quick hop from approximations to statistics. The mode of the numbers of bricks high is a common statistical approach. It also a fast hop to go from statistics to machine learning (or if you like AI and its new offspring GenAI). I've spent most of my career in and around statistics, either in insurance and actuarial calculations, in markets and Monte Carlo simulations, or now at Coles where we have thousands of machine learning models predicting the most likely demand for products, such as strawberries, lamb cutlets, or even tempeh. Each model is based on hundreds of variables (the statistical view was in some instances over 10x more accurate than rule based models). All of this adds up to improved availability and to maximise pantry life. But it is only recently that I twigged that this was a fundamental change in what it meant to program. I'm late to the realisation by the way. In 1974 Peter Naur proposed "data science" as a replacement for the term "computer science." It seems finally we are here.

Yet, something different has happened in the last year (besides venture capitalists creating excitement around GenAI to increase valuations of their start ups, as they had done previously with blockchain, virtual / augmented reality etc.). There is serious money pouring into AI: in 2023 $154b USD was spent on AI related development. $21.3b of that was on GenAI. This year it is expected to reach $40b alone on GenAI, and some forecasting it will reach $100b by 2027. It is at this point that you can see why I began to think my brain was infected with a virus from cat faeces — it was a lot of money with no obvious business value. And in case you think it is just money going into idea worms, off the back of the AI craze last year saw the growth in energy consumption in data centres increase by the equivalent of total annual energy consumption of Germany (i.e. the power consumed by 84m people in a highly industrialised state in a single year -- it took Germany a century to do that and data centres a year). Data centre power consumption is forecast to double the growth again in the next year.

Still there was something compelling about GenAI when it first came out. The parasitosis seemed definite when I first used GenAI and was stunned how it could give me clear answers to complex questions. Was I loosing my mind, or had machines finally passed the Turing test? The world's knowledge could be accessed and articulated in clear blocks of natural language and was even able to be a bit quirky. Sure, GenAI hallucinated, but so did I when I tried to remember complex bits of information, and occasionally found myself making up bits and pieces to connect stories in my head. But the first hint it was delusional as opposed to a real parasoitosis was when I asked GenAI to add up a few shopping list items and it gave me different answers. Wasn't addition predictable anymore? Well, you could argue that 20+5, is about 22, or about 27. Or put another way 22 and 27 is definitely more accurate than -175,232. Was 22 really a good enough approximation of 20 + 5?

And then I read The Things We Make: The Unknown History of Invention from Cathedrals to Soda Cans - William S. Hammack. If you haven't read it, then you should. If you don't read, then listen to it as an audio book. If you don't do that, then watch him on YouTube (he's the engineering guy). It is full of fresh examples about how the engineering method asks very different questions to science. Engineering asks: how can we get this to work, versus science: how does this work? Engineering is commercial (what's good enough to make this profitable? For the scientisists out there he gives an example of fluid dynamics in compressors; a problem we don't have accurate maths for, and which only experimentation and approximation enable you to solve). Engineering doesn't have to be explicable to be profitable or usable. The Cathedrals opening is excellent, and the soda cans are fun (why isn't coke in square cans, which would be the most efficient to transport? A great example of engineering practicality).

The Things We Make, made me realise that all of my career had been spent in approximations. Business rules, physics, are all approximations, we just used simple equations as approximations of reality. Suddenly everything was statistics to me. Suddenly AI was just one more approximation. Imperfect as it was, it was a better representation of reality that computer subsystems before. And accuracy has its own cost. At Coles we retrain our forecast model every week, rather than daily because while training daily might be more perfect, it doesn’t yield efficiency worth the resources and time. Core to statistics is the measurement of probable error — which dates back to the 18th century. The degree of error being important in determining the accuracy of your analysis. What matters is the design of your experiments, and the quality of the questions you ask (if you are interested, read Ronald Fisher -- who fundamentally designed how to run experiments in the 1920's).

Ultimately what I kept seeing was how easy it was to use GenAI, even if it was wrong. And the idea worm became very clear: GenAI might not be perfect, but is a better approximation of humans interact with each other than anything before. My thought worm morphed into: is this more like a shift from DOS prompt to GUI (graphical user interface) than automation and productivity? Is GenAI the best approximation of how we think because it is trained on all the content we have created, rather than our poor approximations such as business rules and equations that are just one way of thinking? A shift from GUI to GenAI is a fundamental shift, even if you clear out the rest of the hype that's in the market. Behind the delusional parasitosis is something -- and it’s more than all the cats on vacuum cleaners on YouTube that GenAI was trained on.

I'd love to hear what you think, is it really a fundamental shift in the user interface, or something more?