Hi OU students. I am writing this article to explain some of the basic concepts of AI. I’m not going to tell you how to program or anything complicated, just explain the basics, so don’t worry! For good or bad, AI seems to be everywhere (and growing) today so I thought it would make an interesting article.
The term ‘artificial intelligence’ was coined by American computer scientist John McCarthy in 1955. It is a label that covers a wide variety of man-made intelligences, but this article will focus on what most of us think of as AI, such as ChatGPT. A lot of people don’t like AI and have concerns, which are grounded in reality and are fair enough, but I have a more positive view than many. I like to think of AI as a superhuman companion that can function as a personal assistant (but one that is way cooler and less invasive). AI can do things humans can’t do, in the same way we can do things AI can’t do. Figure out how to utilise AI and you can do anything (not really, but AI can do a lot)!
I’ll walk you through the theoretical basics in a straightforward way, with a bit of basic technical detail by explaining terminology and specific AI concepts, like how ‘neural networks’ form the backbone of large language models like ChatGPT, and how some of these concepts relate to each other.
The building blocks of modern AI: neural networks
Neural networks are the foundational tech in AI, inspired by how our brains process info. Basically, they’re a bunch of interconnected nodes – which can be thought of like neurons in a human brain – organised into three layers: input, hidden, and output. Data (numbers) is put in, then gets multiplied by weights (numerical values assigned to the connections between nodes), added up, and passed through activation functions (basically like a switch that decides if the information a neuron receives is important enough to be passed on to the rest of the network) to decide if a neuron fires, which would contribute to the output.
Data given to the network at the beginning during a training phase is crucial for determining how it acts. Using backpropagation, the network tweaks those weights to minimise errors, learning from examples. It’s like adjusting a recipe after tasting – over time, it gets patterns spot-on. Neural networks are crucial because they’re the foundational basis of modern AI, and what enable everything from simple pattern recognition to the more complex things in most AIs we use.
Machine learning: teaching computers to learn from data
Machine learning (abbreviated to ML) builds directly on neural networks, but it’s broader – it’s about algorithms that improve from data without explicit programming. In supervised ML (which is when the AI, such as an LLM using the neural network model is given ‘labelled data’, which means it correlates to a specific output), you feed this data (e.g., images tagged “cat” or “dog”) and use loss functions (maths formula used to determine how far the AI’s output is from the correct answer) to guide the learning.
Unsupervised ML is when data without labels is given, so the AI has to spot hidden patterns and think more, while reinforcement learning (which is another key aspect of ML) is when rewards are given appropriate to the output – for example, training an AI to play games via trial and error when success is pinged in the neural network as ‘correct’.
Neural networks often serve as the model in ML. For instance, a basic neural network might classify emails as spam using gradient descent for optimisation. So basically ML turns raw data into predictions, powering everyday features like fraud detection in your banking app.
Deep learning
Deep learning is like ML. It goes further because it’s when neural networks work with many hidden layers – hence named “deep.” These layers allow hierarchical feature learning: the early ones catch low-level details (edges in images), then mid-layers combine them into shapes, and deeper ones recognise full objects. So this is basically just a term for neural networks working at a deeper and more complex level with more nodes. This process requires a computer with lots of GPUs and processing power.
Generative AI
Generative AI is a subset of artificial intelligence that builds on deep learning. It is designed to create original content, such as text, images, or code by learning the underlying patterns and structures of existing data. In contrast, traditional AI (often called Discriminative or Analytical AI) focuses on following pre-defined rules to categorise data, identify anomalies, or make predictions based on specific inputs. An example would be technology used for collecting and screening biometric data. While traditional AI is excellent at evaluating what already exists and finding patterns, generative AI is built to “create” something that didn’t exist before. LLMs are an example of generative AI.
Large language models (LLMs): conversational AI
LLMs are the AI most widely used. Some examples are ChatGPT, Claude 3 Opus, Google Gemini and my favourite, Grok. Siri isn’t one. Neural networks are the foundation of LLMs. They use maths to weigh word importance and decide the output which you get in a chat box. Every word is a calculated choice from a neural network. Numbers represent the letters chosen. The whole output is just a maths calculation which is transformed to text on your screen.
AI agents: The doers of the AI world
AI agents aren’t physical robots. They’re specialised software built with programming languages like Python (the most commonly used). AI agents extend LLMs by adding autonomy. They plan, reason, and act. Using a way of acting called ‘ReAct’ (Reason + Act), an agent analyses tasks, calls tools (e.g., web search with APIs), and iterates (self-improves) with feedback loops.
Correlating back, agents leverage LLM neural networks for natural language understanding, but add planning algorithms like tree search. Early ones like Auto-GPT chain LLM calls to handle multi-step jobs, say booking a trip, check weather etc. They’re evolving, but are already very useful for productivity, like automating research or scheduling.
Artificial general intelligence (AGI): the next frontier
AGI is a theoretical term for an AI that functions at human level across tasks. This contrasts with current AIs such as LLMs, which only provide written information, and chess bots which only play chess, for example. An AGI is quite a scary concept in many ways, though it could be extremely useful. Imagine having an app you could open that could do a vast multitude of tasks and seems to have a personality of its own. It could be a chatbot, play games, code and think in ways that would make it fascinating to interact with, and would be much smarter than any of our current AIs.
Technically, it might involve scaling up transformers (a type of powerful neural network that can compute multiple data sets at the same time) with better architectures or hybrid systems. As of early 2026, we’re close with multimodal models, but true AGI – handling unseen domains – isn’t quite here yet. Experts have many predictions for when we’ll see it, but they range from about three or four years until the 2030s. It’ll build on today’s deep learning foundations, and will revolutionise fields like medicine and science.
How AI benefits the average person today
At this point we know AI’s not just theoretical or a fantasy from Terminator or Star Wars. It is on our phones, our computers and even being integrated into cars and public places like the London Underground. Neural network-powered assistants like Siri handle reminders for many of our daily tasks. ML now recommends Netflix binges using algorithms based on data it gets from viewers. Deep learning in Maps predicts routes. Generative AIs such as LLMs like ChatGPT answer your questions and perform your online tasks. Agents oversee these procceses by automating chores, like planning meals, by cross-referencing recipes and your calendar. So, at this point it is a pillar of modern life, for better or worse, and learning how to use it is greatly beneficial!
0 Comments