AI - a catch up guide to early episodes

Photo credit: Nicolas J Leclercq via Unsplash

Have you ever tried to start a major TV series partway through? You hear everybody talking about it at work, and it sounds thrilling. Then you watch the latest episode, and are baffled by names, places and relationships. Why is this woman so angry with that man? Why are those two factions fighting? And why are those people wandering in the wilderness, apparently disconnected from the rest of the plot? You switch to the series guide on your streaming service and realise that, to catch up, you are going to have to watch the three previous seasons. Perhaps you should just watch that cooking show again.

Trying to understand AI can feel like this. To many people, the appearance of generative AI a few years ago was a sudden, magic and unheralded event, followed by a never ending stream of releases, products and announcements. It’s hard to make sense of the present, let alone look to the future.

I believe that, just as there are handy recap sites on the Internet, which summarise just about every TV series ever made, it’s possible to get to grips with AI without having to watch every previous episode. Instead, we can just pick out some of the highlights of the series called Computing (and let’s remember that AI is a season or a spin-off of that original series). Here’s my attempt at a few moments which explain how we got to where we are today – or at least to the point when generative AI became a public phenomenon. We start right at the beginning . . .

In 1823, the British Government grants £1,500 to Charles Babbage to build the Difference Engine, a mechanical calculator. The first ever government IT project, it goes over time, over budget and is never completed. Babbage is distracted by his Analytical Engine, a machine which can be reprogrammed to solve many mathematical problems using punched cards. He is more than a century ahead of his time.

In 1936, Alan Turing writes the paper On Computable Numbers, in which he proposes the idea of a universal machine, which can use just two symbols to compute any possible mathematical problem, while in 1937, Claude Shannon writes the paper A Symbolic Analysis of Relay and Switching Circuits, which shows how mathematical logic can be implemented using physical electrical circuits.

In 1943 and 1944, Turing’s and Shannon’s ideas come to life. At Bletchley Park, Tommy Flowers builds Colossus, the first electronic digital computer, while in the USA, Howard Aiken commissions IBM to build the Harvard Mark I, the first general purpose computer. Colossus is used to crack the German Lorenz cypher, and the Harvard Mark I is used to perform calculations for the Manhattan Project. Computing is born and raised in the cradle of war.

In 1956, a group of thinkers gather at Dartmouth College, New Hampshire, for a summer research project. It formally uses the term ‘artificial intelligence’ for the first time, and proposes that, ‘An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.’ It takes much longer than a summer for its ideas to come to fruition.

In 1957, Frank Rosenblatt builds the Mark 1 Perceptron, a machine to recognise visual images. Rather than following pre-defined instructions, it alters the strengths of signals in a set of circuits designed to emulate the human brain. In other words, it trains a neural network through machine learning. It successfully distinguishes between a square and a diamond and the letter X and the letter E, but struggles to find backing and practical application. Interest in machine learning and artificial intelligence wanes.

In 1960, Grace Hopper leads a team to define COBOL, a programming language designed to let anyone, not just advanced mathematicians, program computers. It will be used to write more than 800 billion lines of code, and to run organisations around the world.

In 1969, Apollo 11 lands on the Moon, guided by a computer using a new technology: the silicon chip. The chips used in Apollo are three years old, and are already being surpassed by chips back on the Earth. The law defined by Gordon Moore - that the power of chips doubles every year - is coming true.

In 1983, two networks of computers used by the US Department of Defence are connected using a standard protocol known as TCP/IP. The Internet is born.

In 1990, Tim Berners-Lee, while working at CERN, develops a set of standards to enable documents to be linked and shared over the Internet. He calls it the World Wide Web. Within twenty years, there will be 17 billion web pages in the world.

In 1997, the year of Judgement Day as predicted by Terminator 2, Skynet does not come online and the word does not end, but Gary Kasparov seems to think it has. He is beaten at chess by IBM’s Deep Blue. Deep Blue is a rules based expert system - it does not use Machine Learning.

In 2006, the Face Recognition Grand Challenge, is won by algorithms which learn rather than follow set rules. It represents a growing trend for machine learning, which has been experiencing a resurgence since the 1990s, to outperform other techniques.

In 2012, a team led by Andrew Ng and Jeff Dean uses a type of machine learning known as deep learning, in which a neural network has many layers, to emulate an important human activity: recognising cats in Youtube videos.

In 2017, a team of researchers working at Google publishes the paper Attention is All You Need. This paper describes the Transformer model: an approach to artificial intelligence in which deep learning networks are strung together to predict a good word to follow a natural language input. This ability enables models to appear as if they are in reasoned dialogue with a human interlocutor.

In 2022, OpenAI launches ChatGPT. It provides a web interface to a transformer model known as GPT 3.5. This model is a deep learning neural network trained on the content of the billions of pages on the World Wide Web. It is built from software created using languages which share a heritage with COBOL, and runs on computers containing silicon chips which implement the principles defined by Turing and Shannon, and which Babbage attempted to build using rods, gears and cogs back in the 19th century.

There have been many more episodes since 2022. Time will tell which matter as most as some of these earlier episodes. But, as there is no end in sight to the series, and as the plot gets more and more complicated, those of us who have been watching for a while have a duty to help newcomers get started.