It is generally accepted that Charles Babbage was the first person to describe what we would recognise as a computer. His analytical engine was imagined as a general purpose calculator which could perform many tasks. Still without the means to build such a thing, Alan Turing took the idea still further in the 1930s and at the time of his death in the mid 1950s, the first such machines were actually being built.

The key conceptual leap is that rather than needing to build different machines for different situations, one needs to build just one flexible machine and program it however you wish. Early computers were the size of rooms, fed by punched cards, but the principle was the key, and computers got faster and smaller over time. Room sized machines became furniture sized and finally suitcase-sized. Home computers became a thing, from the tiny ZX80 which could either display an image or think, but not both at the same time, to early Apple and IBM machines which began using familiar parts, but these just made universal computing more accessible. It was still the same ball-game – if you know how to codify the task, you can get the computer to execute it.

But universal problem-solving turned out just to be chapter one. I remember using a dumb terminal to connect to a mainframe when I was a boy, so networking is almost as old as computing. But local networks are just ways of distributing computing power. The second chapter is the internet, universal connectivity. Now each of the universal problem solving machines in the world is connecting to all of the other universal problem solving machines in the world.

I owned a ZX81, ZX-Spectrum, a ZX-Spectrum 128, a BBC micro, and a Window 3.1 machine. None of them were connected to the Internet. But I remember only a few years into the dot-com boom, a friend wryly observing that if you put a young person in front of computer that can’t connect to the Internet, they think it’s broken. That’s the world we live in today. It wasn’t true in the 1980s.

Graphical user interfaces, with a mouse to move a pointer across the screen, made computers way easier to use, but that’s just a different point of entry. The internet, and particularly the World Wide Web, truly heralded the second chapter in the history of computing. And like the first, it was all about universality. You don’t have to already have access to the data you need. It’s on another computer and your computer can go get it for you.

Online shopping, social media, web apps, reference works, streaming, all flowed through copper wires, fibre optic cables, and through the airwaves. And the computers continued to get smaller. They fit in our pockets. They became ubiquitous. The same device could take a picture, know where on Earth it was taken, upload it to a server and it could be shared with every human alive. Universality had achieved everything.

One frontier remained. Computers had got progressively easier to use in steady increments, from punched cards to commands lines to graphical user interfaces, to touch controls. But creating new solutions to problems was still specialist work. Even when the bulk of the complexity was hidden from view, it was still a minority that messed around with APIs, app-to-app interactions, or detailed customisation.

That changed with the rise of LLMs. Windows 3.0 got 4 million users in its first year. Windows 95 got 40 million. By 2000, Google was delivering 100 million searches a day, by which time it had been available for 3-4 years. ChatGPT got to 100 million active users in two months. It took Facebook four-and-a-half years to get to that number.

Because this is the third chapter, the third age of universality. Now you don’t have to understand a thing about how the universal problem solving machine works, or how it connects to all the other universal problem solving machines in the world. Now everyone can ask it to solve problem in plain English. The machine which can do everything and which is connected everywhere can now understand anyone.

Where does that leave us? It’s hard to know. Could we have predicted the rise of social media in 1996? Possibly. Bulletin boards were well-established. But the idea that we’d pay to be the ones creating stories, videos and images for other people to consume is very strange. Could we have predicted streaming? Possibly. I like to say that I use my TV now as a jukebox – I pick what I want to watch and watch it when I want – rather than as a radio – turning it on and accepting whatever comes out of it. It’s not a big jump from home video to Disney+, just ask Netflix.

So what will happen next? There’s a tension currently (as there so often is) between security and convenience. I get frustrated when the “intelligence” on my phone can’t read my emails and add things to my calendar. These things have been walled off to keep us safe, but those walls are getting more and more porous by the day. What’s clear is that before long, we’re going to be interacting with our devices using natural language (whether spoken or typed) more and more. Graphical user interfaces will soon seem as quaint as commands lines or punched cards.

And we can expect much more autonomy as we decide the intelligences require less and less handholding. We will say what we want to happen, and the systems will do our bidding, without further interaction – weekly, daily, hourly, continuously. And more and more tasks, especially language related tasks will be handed off to these systems.

But herein lies a flaw. It’s just possible that the fourth chapter will be a step backwards, not another leap forward (just as it was once said that if the third World War were to be fought with nuclear weapons then the fourth World War would be fought with bows and arrows). The LLMs we’re all using today have been trained on a couple of hundred years worth of material, all painstakingly handcrafted by humans. That’s changing.

By some estimates, around a third of the text that’s being added to the internet each year is now machine generated rather than human generated. And that means the models increasingly are being trained on their own output. And while so-called model collapse (where the model starts producing gibberish output) is nowhere near a certain outcome, it is possible, and more subtle degradations are more likely.

That’s the future we should be worried about as we blithely instruct the robot in our phone to read our emails and reply to them for us. That we’re poisoning the well with every command we issue.

At least – that’s where I’m intending to leave this little essay. But I haven’t had ChatGPT give me its verdict yet.

ChatGPT: Overall, I’d say the first three-quarters is already very close to publication-ready. The place I’d spend time is the transition into the model-collapse section. At the moment it’s an interesting ending attached to a strong essay. With a better bridge, it becomes the logical conclusion of the essay rather than a separate thought appended to it.

What do you think?

Whither Who?