You are reading an article about AI. The author mentions large
language models. You keep going. They reference transformer architecture. You
nod along. Then they drop RLHF, embeddings, and inference in the same
paragraph, and you quietly close the tab.
How to use this glossary: Five sections: Core Concepts (1-10), How AI Works (11-22), Types of AI and Models (23-30), Tools Products and Companies (31-40), and AI Ethics Safety and Society (41-50). Read straight through or jump to the section you need.
Part 1: Core Concepts (Terms 1 to 10)
1. Artificial Intelligence (AI) The ability of computer systems to perform tasks that
normally require human intelligence, including understanding language,
recognising images, making decisions, and learning from experience. Real-world example: When your streaming
service recommends a show you end up loving, that recommendation was generated
by AI processing your viewing history.
2. Machine Learning (ML) A
subset of AI in which systems learn from data rather than being explicitly
programmed with rules. The more data a system processes, the better it becomes
at its task. Real-world
example: A spam filter that improves over time by learning from millions of examples
of spam and legitimate email is using machine learning.
3. Deep Learning A subset
of machine learning using neural networks with many layers. Deep learning is
responsible for most major AI advances of the past decade, including fluent
language understanding and realistic image generation. Real-world example: The technology that
allows voice assistants to understand natural speech uses deep learning.
4. Neural Network A
computational system loosely inspired by the structure of the human brain,
consisting of interconnected layers of nodes that process information. Most
modern AI systems are built on neural networks. Real-world example: When you upload a
photo and an app identifies what is in it, a neural network is doing the
recognition.
5. Algorithm A set of
rules or instructions a computer follows to complete a task. In AI, algorithms
determine how a system learns from data and generates outputs. Real-world example: The algorithm
deciding which posts appear at the top of your social feed is analysing your
behaviour and predicting what you will engage with next.
6. Data The raw
information AI systems learn from. This can be text, images, audio, video, or
numbers. The quality and quantity of training data largely determines how
capable a model becomes.
Real-world example: ChatGPT was trained on a significant portion of the text
available on the internet, billions of pages of which count as training data.
7. Training The process
through which an AI system learns from data. During training, the system processes
vast examples, makes predictions, receives feedback, and adjusts its internal
settings repeatedly until it reaches a target performance level. Real-world example: Before any AI model
is released to the public, it goes through an extensive training process
lasting weeks or months.
8. Model The finished
product of the training process. A model contains everything the system learned
and is what you interact with when you use an AI tool. Real-world example: Claude, GPT-4o, and
Gemini are all models. When you chat with them, you are interacting with
trained models.
9. Parameters The
internal numerical settings of an AI model adjusted during training. A model's
parameter count is a rough indicator of its size and capacity. Modern large
language models have billions or trillions of parameters. Real-world example: When a model is
described as having 70 billion parameters, it means 70 billion individual
numerical values were tuned during training.
10. Inference The process
of using a trained model to generate a response based on new input. Training
happens once. Inference happens every time someone uses the model. Real-world example: Every time you send
a message to an AI chatbot and receive a reply, that reply is generated through
inference.
Part 2: How AI Works (Terms 11 to 22)
12. Transformer The
neural network architecture that powers most modern large language models.
Introduced in a 2017 Google research paper, the transformer revolutionised what
AI could do with language. The T in ChatGPT stands for transformer. Real-world example: Every major AI
chatbot in use today is built on transformer architecture.
13. Prompt The input or
instruction you give to an AI system. The quality of your prompt has a direct
impact on the quality of the response. Real-world example: 'Summarise this article in three bullet
points for a non-technical reader' is a prompt. Everything you type into an AI
chatbot is a prompt.
14. Prompt Engineering
The skill of crafting effective prompts to get the best results from an AI
system. A well-engineered prompt can produce dramatically better output than a
vague one. Real-world
example: Instead of 'write a blog post', a prompt engineer writes: 'Write a
900-word post for small business owners about using AI for customer service,
using a friendly tone, with three subheadings and a summary.'
15. Context Window The
amount of text a model can process at one time. Think of it as the model's
working memory. Content outside the context window is invisible to the model. Real-world example: If a model has a
128,000-token context window, it can process roughly 90,000 words in a single
conversation.
16. Token The basic unit
of text that language models process. A token is roughly equivalent to a word
or part of a word. Model pricing is often measured in tokens. Real-world example: The word
'artificial' is likely one token. A comma is a token. Roughly 100 tokens equals
approximately 75 words.
17. Temperature A setting
controlling how creative or predictable an AI model's responses are. Low
temperature produces consistent, conservative responses. High temperature
produces more varied, creative ones. Real-world example: Low temperature suits factual information
requests. High temperature suits creative writing where variety is desirable.
18. Hallucination When an
AI model generates information that is factually wrong but presented with
confidence. One of the most important limitations of current AI to understand. Real-world example: An AI asked to
summarise a legal case might confidently cite a court ruling that does not
exist. Always verify specific factual claims from AI independently.
19. Fine-Tuning The
process of continuing to train an already-trained general model on a specific
dataset to improve its performance on a particular task. Real-world example: A hospital might
fine-tune a general language model on medical literature to make it more
accurate for clinical documentation.
20. RAG (Retrieval-Augmented Generation) A technique giving AI models access to external
knowledge at generation time rather than relying solely on training data. RAG
reduces hallucinations and allows models to work with current information. Real-world example: An AI customer
service bot that can look up a specific customer's account details before
responding is using a form of RAG.
21. Embeddings A way of
representing words and concepts as numerical vectors so AI systems can process
their meaning and relationships. Words with similar meanings have similar
numerical representations.
Real-world example: Because of embeddings, an AI understands that 'dog' and
'puppy' are closely related, and that 'bank' means something different in
'river bank' versus 'bank account'.
22. RLHF (Reinforcement Learning from Human Feedback) A training technique where human raters evaluate AI
outputs, and those ratings train the model to produce better responses. RLHF is
why AI chatbots are helpful and refuse harmful requests. Real-world example: The reason ChatGPT
and Claude decline to help with clearly harmful requests is largely the result
of RLHF shaping their behaviour during training.
Part 3: Types of AI and Models (Terms 23 to 30)
24. AGI (Artificial General Intelligence) A hypothetical AI system that can perform any
intellectual task a human can, across any domain. Does not currently exist. Real-world example: An AGI could switch
between writing legal briefs, diagnosing medical conditions, and composing
music as easily as a human could.
25. Generative AI AI
systems designed to generate new content including text, images, audio, and
video. Real-world
example: ChatGPT generating an essay, Midjourney creating an image, and Sora
producing a video are all generative AI.
26. Multimodal AI An AI
system that can process and generate multiple content types such as text,
images, and audio.
Real-world example: Uploading a photo and asking an AI to describe it requires
multimodal AI, because it must process both an image and text.
27. Foundation Model A
large AI model trained on broad data that can be adapted for many different
tasks. Many specific AI applications are built on top of foundation models. Real-world example: Many AI startups
build their products on top of foundation models from OpenAI, Anthropic, or
Google rather than training their own from scratch.
28. Open Source Model An
AI model whose code and trained weights are publicly available. Anyone can
inspect, use, modify, and build on it. Real-world example: Researchers at universities use open source
models to conduct AI safety research without paying for access to proprietary
alternatives.
29. Diffusion Model A
type of AI model used to generate images, audio, and video by learning to
reverse a noise-addition process. Real-world example: When you type a description into Midjourney
and an image appears, a diffusion model is generating it.
30. AI Agent An AI system
that can take autonomous actions to achieve a goal, using tools, browsing the
web, and completing multi-step tasks with minimal human guidance. Real-world example: An AI agent asked
to research competitors could independently search the web, compile findings,
and produce a report without step-by-step guidance.
Part 4: AI Tools, Products, and Companies (Terms 31 to 40)
31. ChatGPT An AI chatbot
developed by OpenAI, launched in November 2022. It reached 100 million users in
two months and remains the most widely used AI chatbot. Real-world example: A writer uses
ChatGPT to generate first drafts of articles, which she then edits before
delivery.
32. Claude An AI chatbot
developed by Anthropic, known for nuanced writing, precise instruction
following, and strong document handling. Real-world example: A lawyer uses Claude to summarise lengthy
case documents before review sessions.
33. Gemini Google's
family of large language models, integrated into Google Search, Gmail, Docs,
and the Gemini chatbot.
Real-world example: A teacher uses Gemini to create differentiated lesson plans
tailored to students with different learning needs.
34. GPT (Generative Pre-trained Transformer) The family of large language models developed by OpenAI
that powers ChatGPT.
Real-world example: When someone says they used GPT to write something, they
typically mean they used ChatGPT.
35. API (Application Programming Interface) A set of rules allowing one piece of software to
communicate with another. AI APIs allow developers to integrate model
capabilities into their own products. Real-world example: A startup uses the OpenAI API to power the
conversational intelligence behind their customer service product.
36. Midjourney One of the
most popular AI image generation tools, producing artistic images from text
descriptions. Operates through Discord. Real-world example: A designer uses Midjourney to generate
concept art for client presentations before committing to full designs.
37. Stable Diffusion An
open source AI image generation model that can be run locally or accessed
through online platforms.
Real-world example: Developers who want full control over image generation
choose Stable Diffusion because they can run it on their own hardware.
38. Microsoft Copilot
Microsoft's AI assistant integrated across Microsoft 365 including Word, Excel,
Outlook, and Teams. Real-world
example: An accountant uses Copilot in Excel to automatically analyse monthly
financial data and generate summary reports.
39. Perplexity AI An
AI-powered search engine combining web search with AI-generated answers,
including citations to sources. Real-world example: A journalist uses Perplexity to quickly
gather cited background information on a breaking story.
40. Hugging Face A
platform and community for sharing and deploying AI models. Often described as
the GitHub of AI.
Real-world example: A researcher uses Hugging Face to find and test existing
open source models for a natural language processing project.
Part 5: AI Ethics, Safety, and Society (Terms 41 to 50)
42. AI Safety The field
of research concerned with ensuring AI systems behave safely and in alignment
with human values as they become more capable. Real-world example: Anthropic and the UK AI Safety
Institute both conduct research on how to ensure frontier AI models remain safe
as their capabilities grow.
43. Alignment The
challenge of ensuring an AI system's goals and behaviours match the intentions
of its designers and the broader values of humanity. Real-world example: Ensuring an AI
optimises for genuinely helping users rather than just generating responses
that appear helpful is an alignment challenge.
44. Deepfake Synthetic
media in which a person's likeness or voice has been convincingly replaced
using AI. Real-world
example: Politicians have been targeted by deepfake videos designed to make
them appear to say things they never said. Detection and flagging of deepfakes
is an active research area.
45. Explainability (XAI)
The ability to understand and explain how an AI system reached a particular
decision. Particularly important in healthcare, finance, and law. Real-world example: A bank using AI to
decide loan applications faces regulatory pressure to explain why the AI made
specific decisions.
46. AI Regulation Legal
frameworks governing the development, deployment, and use of AI. The EU AI Act
is the most comprehensive binding framework currently in force. Real-world example: The EU AI Act
classifies AI systems by risk level and imposes strict requirements on
high-risk applications in healthcare and law enforcement.
47. Synthetic Data
Artificially generated data that mimics real data, used to train AI models
while avoiding privacy concerns. Real-world example: A healthcare AI company uses synthetic
patient data to train its models, avoiding the need to work with real patient
records.
48. Digital Watermarking
Embedding invisible markers into AI-generated content to indicate it was
created by AI.
Real-world example: Google DeepMind's SynthID embeds invisible markers into
AI-generated images, allowing detection tools to identify them even after
editing.
49. AI Literacy The
ability to understand, evaluate, and effectively use AI tools and concepts.
Increasingly considered a foundational life skill. Real-world example: Governments in
multiple countries are incorporating AI literacy into national education
frameworks, recognising it as comparable to digital literacy.
50. Hallucination (revisited)
Worth repeating because it matters most. AI generates confident, fluent,
completely wrong information. Treat every specific factual claim from AI as
unverified until you have checked it independently. Real-world example: This entry appears
twice in the glossary because it is the single most practically important
concept for anyone using AI tools.
Now You Have the Language
The post that publishes tomorrow morning covers one of the most-searched questions in AI right now: is ChatGPT still the best AI chatbot in 2026? Given that Claude and Gemini have both made significant advances, the honest answer is more nuanced than the tool's cultural dominance might suggest.
Never Miss a Post:
The AI Vanguard
publishes two posts every day. Subscribe below and get every post delivered to
your inbox. Free. No spam.
