From the ZX81 to ChatGPT: code matters

The Sinclair ZX81 turned 42 recently (it was launched on the 5th March 1981). 42 seems an appropriate age to celebrate the existence of the machine which gave many people in the UK (including me) their first proper experience of computing. It may not have been the answer to Life, the Universe and Everything, but it was an answer to the question of how to make computing accessible to a lot more people.

For anyone who is not familiar with the ZX81, it was a small, cheap home computer invented by Sir Clive Sinclair, the founder of Sinclair Research. It had one kilobyte of memory, expandable to 16 KB through the famously temperamental and wobbly RAM pack, and relied on cassette tapes to store data.

Despite its age, and the massive disparities in power, the ZX81 has some notable parallels with technology that we use today. It was a flat, black piece of plastic, much like the flat black devices that occupy our pockets and backpacks – although you had to plug it into a TV to get it to display anything, and nobody who spent any time typing on its keyboard would describe it as touch ‘sensitive’ with a straight face. It was also a personal device: unlike the shared computers at schools, universities and businesses, this machine was typically the cherished possession of one person. And, despite this personal nature, the ZX81 and home computers like it created a crude form of social network, made up of programs typed in from magazines, games on tapes, and heated playground discussions about which machine was best.

However, I think that the most interesting parallel – and potentially the most intriguing difference – is between this ancient computer from way back in the last century, and ChatGPT, the hottest, most heavily hyped technology in the world right now. This parallel may also give a clue to ChatGPT’s remarkable and sudden popularity.

It all starts at the command line. When you boot up a ZX81, the first thing you see is a cursor: a place to start typing commands. What you type is up to you and, as long as the code you write is valid, it will make the computer do something. The program you are supposed to write first is a ‘Hello world’ program, but when I was growing up, the actual first program usually looked something like this, especially when you were showing your friends your new machine:

10 PRINT “<name of friend> is <hilarious insult>”

20 GOTO 10

RUN

Followed by laughter, outrage, and a scramble to hit BREAK and change the text.

The point of this experience was not, of course, the quality of the wit. It was the transformation of the computer from something mysterious and unapproachable to something that could be controlled. You just had to learn the language.

ChatGPT similarly starts with a blank text box. You can type in anything you want. Maybe you ask it to tell you a joke or summarise an article. Maybe you ask it to create a story. And words come out: you get the same feeling of surprise and delight as you get at the ZX81 command line.

And I think that it is this feeling of power, potential and possibility that makes both ChatGPT and early microcomputers so attractive. That blank screen or blank prompt is like a field of fresh snow, waiting for you to walk on it. Many of our experiences of technology today abstract us from this raw computing power: we interact through the icons, buttons and options of user interfaces, not by typing commands and getting results. Most of us are always at a distance from the technology we use to run our lives.

Yet there is a fundamental difference between the ZX81 command line experience and the ChatGPT prompt experience. Within the constraints of the BASIC interpreter built into the ZX81, entering code issues precise instructions to the machine, with predictable results (unless the machine glitches or the RAM pack falls off). The possibility of the software that we can write is unlimited, but it is composed of units of known behaviour.

By contrast, when you enter a prompt into ChatGPT or other LLMs, you don’t know what you’re going to get. It doesn’t take long to realise that the same prompt often yields different results, a novel experience for those of us used to controlling computers with code. Indeed, that’s much of the charm and the whole of the point: an LLM which produced completely predictable output in response to precisely defined prompts would be no more than a complex and indirect PRINT statement.

When new technologies such as ChatGPT emerge suddenly into the public domain, we are left struggling to formulate a response: to know what to think. Maybe the comparison of the ZX81 command line experience and the ChatGPT prompt experience can help us start to figure this out. The common experience of words that make the computer do something reminds us of the wonder and unlimited potential of computing. It puts power at our fingertips. And the difference between predictable execution of code and unpredictable responses to a prompt remind us that the ZX81 and ChatGPT are very different. The logic of the ZX81 is made explicit in hardware: we can discover it by reading the manual, or even by reverse engineering the components. The logic of ChatGPT is implicit within the model: we discover it through interaction. It is the difference between inspecting a machine in a brightly lit room and throwing stones into a dark well to find out what is down there.

Finally, there is one more lesson to take from the ZX81. It has been said that tools such as ChatGPT will render the need for coding obsolete. I believe the contrary: that if we are to make sense of a world with ever more powerful technology, we should understand the fundamentals of that technology. Even though LLMs are models, not programs, the models are built and managed using code running on computers: understanding how computers and code work matters more than ever.

(Views in this article are my own.)