With all the buzz around artificial intelligence (AI) today, it's easy to assume that it's a recent innovation. In fact, AI has been around in some form for over 70 years. To understand the current generation of AI tools and what the future holds, it helps to understand where we came from.
Each generation of AI tools is arguably an improvement over the previous one, but no tool is moving towards consciousness.
Mathematician and computer pioneer Alan Turing published a paper in 1950 with the opening sentence, “Let me consider the question: Can machines think?” Turing went on to propose something called the Imitation Game (now better known as the Turing Test), which states that a machine can be considered intelligent if it is indistinguishable from a human in a blind conversation.
Five years later, the phrase “artificial intelligence” was first used publicly in a proposal for a summer research project on artificial intelligence at Dartmouth College.
From these early stages, a branch of AI known as expert systems developed from the 1960s onwards. These systems were designed to capture human expertise in specialized domains, and because they used an explicit representation of knowledge, they are an example of what is called symbolic AI.
There were many early, widely publicized successes, including systems for identifying organic molecules, diagnosing blood-borne infections, and prospecting for minerals. One of the most notable examples was a system called R1, which in 1982 was reported to have saved Digital Equipment Corporation $25 million a year by designing more efficient configurations for minicomputer systems.
The main advantage of expert systems is that in principle, domain experts with no coding expertise could build and maintain a computer knowledge base. Software components called inference engines then apply that knowledge to solve new problems in the domain, providing some explanation through a set of evidence.
These were all the rage in the 1980s, when organizations clamored to build their own expert systems, and they remain a useful part of AI today.
Enter machine learning
The human brain contains approximately 100 billion nerve cells, or neurons, interconnected in a tree-like (branching) structure. So while expert systems aimed to model human knowledge, another field called connectionism was emerging that aimed to model the human brain more literally. In 1943, two researchers, Warren McCulloch and Walter Pitts, created a mathematical model of a neuron, in which each neuron produces a binary output in response to an input.
Read more: AI will soon be incomprehensible to humans – the story of neural networks explains why
One of the earliest computer implementations of connected neurons was developed by Bernard Widlow and Ted Hoff in 1960. Although such developments were interesting, they had limited practical use until 1986, when a software model learning algorithm called the multilayer perceptron (MLP) was developed.
Adrian Hopgood, Provided by author (not to be reused)
An MLP is a simple arrangement of simulated neurons, usually with three or four layers, where each layer is fully interconnected with the next. The MLP learning algorithm was revolutionary: it delivered the first practical tool that could learn from a set of examples (training data), generalize, and classify never-before-seen input data (test data).
This feat was achieved by attaching numerical weights to the connections between neurons, tweaking them for optimal classification on training data, before being deployed to classify never-before-seen examples.
MLPs can support a wide range of practical applications, provided the data is presented in a usable format. A classic example is handwritten character recognition, but only if the image has been pre-processed to extract key features.
New AI Models
Following the success of MLPs, various alternative forms of neural networks began to emerge, one important one being the Convolutional Neural Network (CNN) in 1998. It is similar to MLPs except that it adds an extra layer of neurons to identify key features in an image, eliminating the need for preprocessing.
While both MLPs and CNNs are discriminative models that can make decisions, typically by classifying inputs and generating interpretations, diagnoses, predictions, or recommendations, generative neural network models have also been developed, meaning they can create something new after being trained on a large number of past examples.
Generative neural networks can generate text, images, and music, as well as generate new sequences that aid in scientific discovery.
Two models of generative neural networks stand out: generative adversarial networks (GANs) and transformer networks. GANs achieve superior results because they are partly “adversarial” – they can be thought of as built-in critics that demand improved quality from the “generative” component.
Transformer networks have come to prominence through models such as GPT4 (Generative Pre-trained Transformer 4) and its text-based version ChatGPT. These large-scale language models (LLMs) are trained on huge datasets taken from the internet, and human feedback further improves performance through so-called reinforcement learning.
As well as producing impressive generative capabilities, the vast training sets mean such networks are no longer limited to narrow, specialised domains like their predecessors, but generalise to cover any topic.
Where is AI heading?
The capabilities of LLM have led to dire predictions of AI taking over the world. In my view, such fear-mongering is unwarranted. Current models are clearly more powerful than their predecessors, but the trajectory is toward improved capacity, reliability, and accuracy, not forms of consciousness.
“The Hollywood dream of conscious machines is not going to come true anytime soon, and in fact there is no visible path to get there,” Professor Michael Wooldridge said in 2017, testifying before the UK House of Lords. Seven years on, his assessment remains correct.
While AI has many positive and exciting potential uses, history shows that machine learning is not the only tool. Symbolic AI still has a role to play because of its ability to incorporate known facts, understanding, and the human perspective.
For example, self-driving cars can be taught the laws of the road rather than learning by example, and medical diagnostic systems can provide validation and explanations for the outputs from machine learning systems against medical knowledge.
Social knowledge can be applied to filter out unpleasant or biased outputs.The future is bright and will see the use of a variety of AI techniques, including some that have been around for many years.
