Everyone is talking about and using AI. Its growth in the last few years has been monumental, and its impact is being felt across all sectors of society as well as verticals.
Within AI, there are a blizzard of acronyms and terms we are just starting to become familiar with. To help you navigate the fast moving waters of artificial intelligence, here are 10 common AI concepts explained.
1. Artificial Intelligence (AI)
AI is the ability for machines to learn, infer and reason. The term artificial intelligence was mentioned at the Dartmouth Conference in 1956. It was organised by John McCarthy, an American computer scientist and cognitive scientist, who co-authored a document coining the term and also developed the programming language family Lisp.
A subset of AI is machine learning, which is the ability to train a machine to learn and adapt without following explicit instructions. A machine learning algorithm is very good at predicting patterns or outliers and became popular in the 2010s prior to deep learning and neural networks.
What are neural networks?
A neural network is a type of machine model that aims to mimic the way a human brain works. It consists of layers of interconnected nodes (“neurons”), where each connection has a weight. Data is passed through the layers and the network learns by adjusting the weights to minimise errors.
Example: A small neural network might take pixel values from an image and learn to recognise whether it contains a car or boat.
What about deep learning?
Deep learning uses large neural networks with multiple layers and are able to learn highly complex patterns from very large unstructured dataset such as text, images or audio. The model is then able to automatically discover rules rather than being defined by humans. It means that a model like ChatGPT can learn reasoning, grammar and writing styles.
Think of a neural network as a tool or framework and deep learning as a highly complex stacked set of tools based on huge amounts of data.
2. Generative AI (Gen AI)
Building on what we’ve learned from AI, Generative AI is a type of artificial intelligence that can create new content including text, images, video, audio and code based on patterns it has learned from existing data.
The important thing to note is that Gen AI is probabilistic: it doesn’t copy existing content that it has learned, but generates outputs that are statistically likely, based on the system’s training.
The output could be any of the following:
- Text: ChatGPT, Gemini or Copilot generating answers, emails, marketing plans, film scripts, personas etc from text prompts
- Images: MidJourney, Firefly or Gemini 2.5 Flash Image (aka Nano Banana) creating artwork, illustrations or concepts from text or image prompts
- Videos: Runway or Kling generating films or video ads from prompts
- Code: Claude, ChatGPT or GitHub Copilot generating or fixing code snippets
- Audio (voice): ElevenLabs or Wondercraft creating voice-based audio such as voiceovers
- Audio (music): Suno creating AI text-to-music production
3. Large Language Model (LLM)
A large language model (LLM) is a type of artificial intelligence model that has been trained on vast quantities of text data to understand and generate human-like language.
These models use deep learning techniques to predict and generate words, sentences, and entire conversations or documents.
In very simple terms, an LLM can be thought of as a very advanced autocomplete system that predicts what text should come next, when outputting content such as text or code.
LLMs are referred to as “large” because they can often contain billions of parameters of data to help the model learn patterns in language. Their training data can come from a variety of sources including books, articles, websites, user generated content and code.
How do LLMs use deep learning?
LLMs use a type of machine learning called deep learning in order that it can interpret how characters, words, and sentences combine and work together.
Within LLMs, deep learning uses probabilistic methods to analyse unstructured data, allowing models to automatically identify and distinguish between different types of content without human input.
LLMs are then further trained via tuning: they can be fine-tuned with new data for high accuracy and customisation, or prompt-tuned using a smaller set of parameters to make it more efficient and less resource-intensive.
4. Agentic AI
AI is already transformational, but agentic AI takes things up a notch.
Agentic AI is an artificial intelligence system that can operate autonomously to achieve a specific goal.
Whereas traditional AI models respond to a user’s prompt, an agentic AI offers a multistep process that can plan, reason and take action in the real world, to solve a problem or carry out a task.
Three examples of agentic AI use cases:
- Book me and my wife return flights in late December, from London to Sydney (and back), with stop-overs in either Dubai or Singapore. Avoid flying out Christmas Day or Boxing Day, and only travel economy class. Also, reserve lactose free meals for me only on all flights.
- Buy me the best running trainers, UK size 11, that accommodate wide feet and a neutral pronation. Most of my runs are off-road and water resistance is important to me.
- I have back pain and need to see a local osteopath who has experience of lower back pain and pregnancy. Book me an appointment for any time next Friday.
Agentic AI systems typically consist of the following components
- Planning: The AI system analyses a user’s prompt and splits it into smaller, manageable tasks; it then creates a plan to fulfil the task at hand.
- Reasoning: The AI system uses knowledge to determine the best course of action and will aim to identify potential obstacles and adjust its plan accordingly.
- Memory: Agentic AI needs a memory to store previous behaviours, observations and feedback, allowing it to learn from its experiences and enhance its performance for future tasks.
- Tool Use: Agentic AI can, when required, use external tools, such as browsers, applications or APIs to gather information, carry out calculations, or even control devices.
Whilst it’s relatively early days for Agentic AI, the use cases and possibilities are endless. From a business perspective, the potential to enhance operational efficiencies is huge.
5. Tokens
In the context of AI and Large Language Models (LLMs), a token is a basic unit of text that the model processes.
Tokens can be as short as one character or as long as a full word. Spaces, punctuation, and partial words all contribute to token counts.
Here are some examples of tokens
- A whole word (e.g. engineer}
- Part of a word (e.g. un- or -ing)
- A punctuation mark (e.g. . or ?)
- Even whitespace (like a space or line break)
In AI, tokens are created through a process called tokenisation, where text is split into these smaller chunks.
Why do tokens matter?
Model Input & Output
LLMs don’t “see” sentences as we do. Instead, they convert everything into tokens, which are then mapped to numerical IDs the model understands.
Cost & Limits
Most AI services (like OpenAI, Anthropic, Gemini) charge based on the number of tokens processed.
Example: “I love ChatGPT.” might be 6 tokens (I, love, Chat, G, PT.)
If a model has a 50k token context window, that might amount to ~35k actual words it can remember at that time.
To give you a real example, the OpenAI charter equates to 476 tokens.
Performance
Bear in mind that longer prompts = more tokens = more computing processing time and potential cost. And efficient prompt design often means reducing unnecessary tokens.
Taking tokenization a step further, tokens are split into a few categories as follows:
- Input tokens – tokens in your prompt
- Output tokens – tokens generated in the output response
- Cached tokens – reused tokens in conversation history
- Reasoning tokens – in some advanced models, extra thinking steps are included internally before producing the final output response
When working with APIs, AI vendors will charge you based on your token usage, among other pricing variables. Some reasoning models may use more tokens than others.
OpenAI has a fun interactive Tokenizer tool, which allows users to calculate the number of tokens and see how text is broken into tokens.
Here’s an example of OpenAI’s Tokenizer tool in action
When it comes to multi-modal image and video generation, whilst tokens will play a part in the transformer process, pricing is more likely to be measured in computing resource usage and media-centric units, rather than token-units e.g. video length, resolution, complexity.
6. Temperature & Top P
In AI, you will the word “temperature” mentioned fairly ubiquitously.
Temperature is a parameter that adjusts the randomness and creativity of a large language model (LLM) when generating an output.
Temperature values typically range from 0-1.0 but some LLMs increase the range to 2.0.
A low temperature e.g. 0-0.3 will produce more predictable and less random results. The model will choose the most likely words, and outputs will be more factual and consistent.
A high temperature e.g. 0.7-1.0 will create more randomness in results with a larger variety of words – sometimes less common. Outputs will be more creative and varied.
Example of temperature use cases:
Low temperature
- Good for: Factual answers, code, legal documents, medical text, analytics
- Pros: Predictable and very factual responses
- Cons: Can lack creativity and be repetitive
- Best for: Tasks requiring lots of precision, such as technical documents, coding tasks
High temperature
- Good for: Brainstorming, creative writing and marketing copy
- Pros: More diverse and creative output
- Cons: Can result in less coherent and unexpected output or even mistakes
- Best for: Creative tasks, brainstorming, ideation and novel ideas
Let’s try an example:
Prompt: “Write a strapline for a coffee shop”
- Temperature = 0.2: “Fresh coffee served 24/7” (safe and predictable)
- Temperature = 0.6: “Awaken your senses with our award-winning coffee” (creative and more varied)
- Temperature = 1.0: “Dance the tango with your espresso partner” (imaginative and much less predictable)
Can I adjust temperature in AI tools?
In most of the popular LLMs such as ChatGPT, Gemini, Claude, Copilot and Perplexity, you can’t adjust the temperature in the user interface, although your choice of model within an LLM might well change the temperature value.
An exception to the rule is Google’s AI studio which currently lets you adjust the temperature setting.
And tools that embed LLMs, such as Jasper and Writesonic, may let you adjust a “creativity” slider, which is essentially the temperature parameter.
If you’re using an API, you will be able to set the temperature parameter manually.
What’s Top-p?
You may have noticed in Google’s AI Studio that you can adjust a setting called Top-p.
Top-p AKA nucleus sampling is a dynamic method for choosing the next token in text generation. Depending on the setting, the model may not look at all possible tokens, but instead restrict itself to a smaller set of tokens.
How do you compare Temperature vs Top-p
Temperature relates to the probability distribution of the model’s output, whereas Top-p restricts the selection of tokens to a subset. In a nutshell, Temperature influences how creative or unpredictable the output is, while Top-p maintains coherence by limiting the selection to the most likely options.
7. Hallucinations
In Large Language Models (LLMs), a hallucination is when the model produces an output that is either incorrect and factually wrong, made up with fabricated or inconsistent content.
Examples of hallucinations include:
- Factual errors: “Jupiter has 85 moons” (it actually has 95 moons)
- Fabrication: Publishing a quote from a famous person that never occurred
- Inconsistent: Claiming a statistic and changing data in the same output
There are significant risks to users or systems that rely on data that could be a hallucination. This particularly relates to sectors such as law, engineering, medicine or government.
Hallucination can negatively impact your brand if LLMs or answer engines such as Google’s AI Overviews publish misinformation about you.
Why do hallucinations happen?
There are a variety of reasons why LLM’s hallucinate, and often they relate to LLM training data or LLM design.
LLMs are designed to produce answers using statistical probability based on their training data. If the training data is incomplete or not up-to-date, gaps will occur, causing the AI engine to fabricate information to fill the gaps.
When hallucinating, the AI has no idea it is wrong, it’s simply using its available resources and most plausible output based on patterns and statistical probability.
Common reasons why hallucinations occur
- Incorrect training data: the model is trained on data that is flawed or incomplete
- Overfitting: the AI engine is overtrained and isn’t able to insert new unseen data
- Lack of grounding: the LLM doesn’t have access to real-time data
- Probability-driven guesses: The AI model generates incorrect predictions due to faulty patterns
How can we mitigate against hallucinations?
One common approach is to use Retrieval Augmented Generation (RAG) where real or more up-to-date documents and data are injected into the context. For example, you may have used or created Custom GPTs in ChatGPT, which gives the AI engine access to a limited subset of data, such as PDFs, or other documents.
Better prompting and prompt engineering will produce better results. Try to use custom instructions, provide references or data in your prompts, and ask the AI tool to double-check or prove its results. Or employ chain-of-thought prompting, which will break down the steps the LLM took to providing its final answer.
8. Retrieval-Augmented Generation (RAG)
For the most part, AI’s are pretty accurate in their output. But not in all cases. Outputs can sometimes contain hallucinations, lack of specificity or out-of-date information.
Let’s use an anecdote.
When my children were younger, they would ask me challenging questions about a whole variety of topics, such as:
- “How many bones are in the human body?”
- “Who is the best paid footballer?”
- “How many people live in the UK?”
Whilst I could guess the answers, the reality is that my own knowledge was out-of-date.
The same issue can occur with LLMs. Their training data can be out-of-date, and for more specialised prompts, LLMs may not have adequate source data. This is where RAG, or Retrieval-Augmented Generation, assists.
Retrieval-Augmented Generation (RAG) is an AI framework that enhances the output of large language models (LLMs), by providing them with access to external, authoritative and sometimes specific knowledge sources
How does RAG solve LLM problems?
RAG introduces an important step in the workings on LLMs: retrieval. It works by combining a retrieval mechanism with a generative model.
When a user asks a question or inserts a prompt, the RAG system searches its designated external data e.g. product specifications, a company’s internal documentation, a private database or intranet relevant to the user’s query.
The system then takes the user’s original query and augments it by adding the retrieved information as context, enriching the original query. And the LLM then generates a response grounded in the retrieved information.
RAG offers significant benefits including:
- Up-to-date information – RAG can be used to continually update an LLM for custom use cases
- Factual accuracy – the AI is forced to uses verifiable source of data
- Domain-specific knowledge – it can be used for specific user cases e.g. an organisation’s HR policies, a legal contract, medical data, a set of product specifications
- Reduced cost – it’s more efficient than retraining a large LLM on new data
9. Context Window
In AI, context window refers to the working memory of a large language model (LLM). It defines the maximum amount of information (measured in tokens) an LLM can process at one time when generating its response.
Think of it as the model’s short term memory.
A larger context window means the AI engine can retain (and therefore understand) more information from its source – be it a conversation or document, and in turn, the output will be more contextually relevant and will have fewer hallucinations.
A smaller context window means the model won’t be able to remember earlier details, resulting in less accurate or relevant output.
What are the context window sizes of common LLMs?
Over time, the context window of well-known large language models has grown significantly since the original GPTs were released. And each successive update typically has longer context windows.
Furthermore, depending on your paid subscription, you will gain access to larger context windows for most of the common LLMs such as ChatGPT, Gemini and Claude.
For example, as of September 2025, below are the different ChatGPT plans with context window sizes highlighted.
10. Artificial General Intelligence (AGI)
Artificial General Intelligence (AGI) is a hypothetical type of AI that can perform any intellectual task that a human can, with a similar level of cognitive ability. It would have the capacity to learn, reason, and adapt across a wide range of domains, just like a person.
You may have also heard Artificial Superintelligence (ASI) spoken about in a similar context. However, ASI is an evolution beyond AGI, and refers to an AI system that would not only match human intelligence, but would far surpass it in virtually every aspect, including scientific creativity, problem-solving, and social skills.
ASI would be capable of independent self-improvement, potentially leading to a rapid, exponential increase in its capabilities.
In short, AGI is about reaching human-level intelligence, while ASI is about transcending it. AGI is seen as a potential stepping stone to ASI.