AI

There are different kinds of artificial intelligence approaches such as expert systems, machine learning and genetic algorithms. Machine learning solves problems by allowing machines to “discover” their own algorithms, which is used in speech recognition and computer vision. An expert system has an inference engine and a knowledge base of facts and rules. Neural networks are a form of machine learning modelled on the human brain, with a collection of linked digital neurons (“nodes”) that produce an output based on weights between nodes; these networks can be trained on a dataset. Such artificial intelligence can be used to provide insights into current data.

A more recent type of AI that burst into the public imagination in 2023 is generative AI, based on neural networks, which is capable of generating original content including documents, images, computer programs, essays, art and music. Generative AI is based on foundation models (also called large language models), which are trained on very large volumes of contents such as digital books and swathes of the internet. This type of AI is essentially a predictive algorithm. Ask it to produce an essay on the industrial revolution, a poem about rap music in the style of Wordsworth, a computer program subroutine in Python or an image of a space cruiser and it will do so. They can also critique or summarise an article, debug computer code or spot patterns in data. The current generative AIs are extremely good at languages and pattern recognition, though they have some limitations in various areas. The quality of their responses depends on the data they are trained on, and it is possible to take a foundation model as a basis and then train it on, for example, specific industry or company material, such as financial news and stock market prices.

With AI accounting for around a fifth of all venture capital in 2022, the number of vendors claiming to be AI related is vast. Leading generative AI companies include Open AI with ChatGPT, Google with Bard and the Google subsidiary Deep Mind, Meta (with LLaMA), IBM (with Watson).

Leading generative AIs that focus on images rather than text include Open AI’s Dall-E, Midjourney, Stable Diffusion, Adobe’s Firefly and Leonardo. There are also vendors that focus on the infrastructure surrounding AI, such as Hugging Face, which builds tools for building AI applications, and Nvidia, which provides specialist semi-conductors that are very well-suited to the kind of processing used in modern AIs. Additionally, there are companies that provide AI for robotic automation such as UiPath. There are many, many more.

AI can augment the abilities of human beings in a wide range of applications and jobs, possibly replacing some jobs entirely. AIs can perform as well or better than radiologists searching for tumours in medical images, can solve complex problems such as protein folding in life sciences, and can even outperform most pilots in a (simulated) dogfight between fighter jets. They can write program code or generate product descriptions, or generate marketing images just from a text prompt. They can spot fraudulent patterns in banking transactions or predict when an engine will fail based on sensor data and maintenance records. A wide range of jobs will be impacted by AI in years to come, and the effect on the global economy may be quite dramatic, with for example Goldman Sachs predicting a $7 trillion increase in global GDP in the next decade.