Look! I made a machine do a thing!
Photo credit: Laura Ockel via Unsplash
When I am doing hobbyist computer stuff, I often have moments which can be summarised as, Look! I made a machine do a thing! These moments bemuse family and friends, who are summoned to witness the thing that I have made the machine do, only to see a splurge of text at the command line, or an underwhelming web page. (But look at it - it works!)
The feeling of successfully making a machine do a thing is such an important phenomenon in computing that I think we should give it a name: perhaps LIMAMDAT, or LIMDAT for brevity and ease of pronunciation. The history of computing has been full of LIMDAT moments, from the first decrypts and trajectory calculations in the 1940s, through rocket launches and Moon landings, to the birth and development of the Internet.
It may seem that the world is currently full of LIMDAT moments, as millions, or possibly billions, of people enter prompts into chatbots and get responses from computer systems running LLMs. They all seem to be making machines do things.
However, I think that the relationship between the use of current forms of AI and the normal experience of making a machine do a thing is rather more complicated.
First, I am not sure that simply getting output in response to a prompt is a true LIMDAT moment. Part of the reason that making a machine do a thing is special is that it provides a sudden feeling that you can see through the machine, that you have your hands on the strings that control its behaviour. This feeling is often illusory: in modern computing we are always operating with multiple layers of abstraction, and the machine rarely works exactly as we perceive it to work. Nevertheless, we are justified in feeling that we have exerted an increased level of control, even if we are just pulling the strings that pull the strings.
With modern forms of AI, that feeling of control is rather more elusive. We can enter prompts which get the machine to do a thing, but it is not always the thing we want it to do, and the same prompt, entered moments later, may not result in the same thing. The layers of abstraction are rather more opaque (often to the people who build the models as well the people who use them). With current forms of AI, it seems that rather than making a machine do a thing, we are persuading or asking the machine to do a thing. And that is a rather different relationship.
Second, true LIMDAT moments are often followed by a moment of deflation. When we get a machine to do a thing, we realise that it is the start of the work, not the end of the work. We still need to prove that the machine will do the thing repeatedly and reliably, and will not break at three o’clock in the morning when we are on call.
By contrast, many of the exuberant announcements about new AI solutions, features and upgrades don’t seem to acknowledge this second (and third and fourth and so on) step. Rather, they hope that the excitement of having made a machine do a thing will carry them to success, whether that success comes as fame, notoriety or riches.
It is good to be excited when we get a machine to do a thing. It is good to be excited about AI. It will be even better when we get just as excited about the next stages: industrialisation, productivity and control.
