The A to Z of artificial intelligence

---

D is for Deep learning
Machine learning, which involves a computer learning from examples, rather than following specific instructions, has many sub-categories and deep learning is one. It uses layers of neural networks (see below) to tackle complex tasks. For example, an image-recognition network might have one layer to identify general shapes, while another handles letters.

E is for Ethical AI
AI brings many risks, from data privacy and security to the lack of transparency in how these systems make decisions. Ethical AI is used to explain the processes of designing and deploying AI with a focus on fairness and transparency.

F is for Fine-tuning
‘Training’ is the period in which AI systems analyse a dataset and practise drawing meaningful conclusions. Computer engineers then fine-tune the model to improve results or manage computing resources more efficiently.

G is for Generative AI
Generative AI is the name for any system that produces new content or data that is related to its training data but distinct from it. This could be new writing, artwork, music, video and more. ChatGPT is the most high-profile example, but there are lots of others.

H is for Hallucination
Some AI systems can produce information or responses that they believe to be correct or appropriate answers, but which are false. Some of these ‘hallucinations’ are plausible, which means users can be misled by them. Others are simply nonsensical.

I is for Instrumental convergence thesis
This is the theory that a super-intelligent AI could eventually destroy humanity. For example, an AI designed to produce as many paperclips as possible might take drastic measures to avoid being turned off if humans stop wanting paperclips. Think of the sentient computer HAL going rogue in 2001: A Space Odyssey and killing crewmates to avoid having to lie to them.

J is for Joint probability
The likelihood of two or more events happening simultaneously. Some AI models use this technique to understand how different variables are related.

K is for Knowledge graph
A method of organising information by demonstrating how concepts relate to each other. For example, Paris is a city, which is in France, a country, which is in Europe, and so on. This helps the AI figure out that Paris is to Berlin as France is to Germany, for example.

L is for Large language models (LLMs)
LLMs power a lot of generative AI tools. They use enormous datasets – billions of words – to process and generate word-based content. The vast number of words mean the system is incredibly good at generating convincing sentences and entire articles.

M is for Multi-modal
A ‘mode’ is just a type of data. Multimodal AI can process inputs and generate outputs in more than one type of data, such as text, speech, images and numbers.

N is for Neural networks
The human brain works through connected neurons that form a neural network to transmit electrical signals. AI does something similar using artificial neurons to create networks that process data. Layers of these networks are used in deep learning (above).

O is for OpenAI
The company behind ChatGPT and therefore perhaps the leading player in the AI world now. The firm is heavily funded by Microsoft, but other big tech companies, such as Amazon, Google and Facebook, are competing, too.

P is for Prompt engineering
Instructions for AI interfaces are known as prompts, and they can be more effective when well-constructed. The skill of crafting the most effective prompts, for example by telling the AI to answer as if it is an expert in the topic in question, is known as prompt engineering.

Q is for Quantum computing
‘Classical computing’ depends on ‘bits’ that can be ‘on’ or ‘off’. In quantum computing – an emerging technology – bits can occupy both states at the same time, thanks to quantum physics. This has the potential to massively increase computing power, particularly in AI.

R is for Reinforcement learning
There are many machine learning methods (see deep learning, above). Reinforcement learning trains the computer to find the ideal response by offering rewards for good answers and punishments for bad ones. The computer then tries to maximise its rewards in future.

S is for Supervised learning
Supervised learning helps the AI with a labelled dataset. By identifying patterns in the labelled data, the AI learns how to classify new, unseen data. This is a common technique for training AI to filter spam email, for example.

T is for Transformer
First proposed by Google engineers in 2017, transformers allow AIs to pay attention to whole sentences at once, instead of one word at a time. This increased the model’s ability to understand context and nuance.

U is for Unstructured data
Unstructured data, such as customer feedback messages, are more challenging for computers to analyse than information in spreadsheets and databases. However, AI can draw meaningful conclusions from unstructured data, for example by identifying complaints that suggest a customer will switch to a competitor.

V is for Virtual assistants
From Siri to Alexa, we’re all familiar with AI-powered virtual assistants. As AI improves, we can expect them to be more capable and personalised. Before too long, everyone might have a virtual assistant tailored to their unique interests.

W is for Weak AI
Any AI system specialising in just one task is considered weak AI, including everything from Amazon’s product recommendations to self-driving cars. The opposite is strong AI, which could learn new tasks like a human. Some systems are close, such as OpenAI’s GPT-4, but it doesn’t actually exist – yet.

X is for XAI (Explainable AI)
To deal with bias, hallucinations and other AI problems, we must understand how the machine reached its conclusion – but that can be impossible when the AI has gone through endless learning loops. XAI is designed from the outset to document and describe its decision-making process.

Y is for Yottabyte
If you’ve read this far, I’ll be frank: there aren’t many options for the letter Y. The GPT-4 database is roughly a petabyte in size. There are 1,000 petabytes in an exabyte, 1,000 exabytes in a zettabyte and 1,000 zettabytes in a yottabyte. So, a yottabyte is a lot of data. No AI uses datasets of that size yet, but give it time.

Z is for Zero-shot learning
A learning model where the AI can handle tasks on which it wasn’t explicitly trained. For example, researchers are exploring the use of zero-shot learning in identifying rare diseases, for which there is little training data for the system to use.

Many of these definitions are over-simplified, but the good thing about today’s generative AI tools is that, if you ask nicely, they’ll answer any further questions you have.