Alignment
The process of ensuring that an AI system's objectives, outputs, and behaviours align with human goals and values. This is especially important in advanced systems that may develop behaviours not explicitly intended.
Example: Ensuring a chatbot for mental health support never recommends harmful actions regardless of prompts.
Application Programming Interface (API)
A set of defined rules and protocols that allow different software systems to communicate and exchange data.
Example: Using the OpenAI API to send a prompt and receive a language model-generated response in your web app.
Artificial General Intelligence (AGI)
A theoretical form of AI that can perform any intellectual task that a human being can. It generalises learning across domains.
Example: An AGI system could learn music composition, perform surgery, and pass a philosophy exam without task-specific programming.
Artificial Intelligence (AI)
The simulation of human intelligence in machines that are programmed to think, reason, and act autonomously.
Example: AI powers personal assistants like Siri and autonomous driving systems like Tesla Autopilot.
AI Ethics
A discipline concerned with the moral implications of AI development and use, including fairness, privacy, accountability, and non-discrimination.
Example: Creating guidelines to prevent hiring algorithms from discriminating based on gender or ethnicity.
Augmented Intelligence
A collaborative model where AI complements and enhances human intelligence rather than replacing it.
Example: AI-powered radiology tools that highlight anomalies for doctors, who make the final diagnosis.
Autonomous Agent
An AI system capable of making its own decisions and taking actions to achieve its goals without human intervention.
Example: A self-driving delivery robot navigating city streets and avoiding obstacles independently.
Backpropagation
A technique for training neural networks by updating weights in reverse from output to input layers, minimising prediction errors.
Example: Used in training image classifiers to reduce the error rate in recognising handwritten digits.
Bias (Algorithmic Bias)
Unintended and systematic favouritism in AI outcomes due to imbalanced or non-representative training data.
Example: A facial recognition system that misidentifies people of colour more frequently due to underrepresentation in training data.
Big Data
Extremely large datasets that require special tools to store, analyse, and extract value, often used to train AI models.
Example: Using millions of user interactions to train recommendation engines for e-commerce platforms.
Black Box Model
A type of AI or machine learning model whose internal logic is not easily interpretable by humans, making it difficult to understand how decisions are made.
Example: A deep neural network used to approve loans but offering no clear explanation for why one applicant was accepted and another rejected.
Cognitive Computing
AI systems designed to simulate human thought processes, such as reasoning and learning, using techniques like NLP and pattern recognition.
Example: A cognitive computing system that helps legal professionals analyse case law and predict outcomes.
Computer Vision
A field of artificial intelligence that enables computers to interpret and process visual data such as images and video.
Example: Facial recognition systems that identify people in security footage using computer vision.
Corpus
A large collection of written or spoken texts used to train language models.
Example: The Common Crawl dataset is a public web corpus used to train large language models like GPT.
Data Drift
The phenomenon where input data changes over time, causing model performance to degrade.
Example: A predictive maintenance model for industrial equipment becomes less accurate as new sensor technology is introduced.
Data Labelling
The process of annotating data with tags or labels to make it suitable for supervised learning.
Example: Labelling thousands of tumour images as benign or malignant to train a cancer detection model.
Data Mining
The process of discovering meaningful patterns, correlations, and anomalies in large datasets.
Example: Retailers using data mining to identify that people who buy nappies often buy beer as well.
Deep Learning
A subfield of machine learning that uses multi-layered neural networks to model complex patterns in data.
Example: Deep learning is used in language models like GPT-4 and image generation models like Stable Diffusion.
Diffusion Models
A class of generative models that learn to produce data by gradually transforming random noise into structured outputs.
Example: Stable Diffusion creates photorealistic images from text prompts using diffusion techniques.
Embedding
A numerical vector representation of data, often used to capture the semantic meaning of words, images, or sentences.
Example: In NLP, the word 'bank' may have similar embeddings to 'money' but different from 'riverbank' depending on context.
Epoch
A full iteration over the entire training dataset during the training process of a machine learning model.
Example: If a dataset has 1,000 examples and a model sees all of them once during training, that's one epoch.
Ethical AI
A design and deployment philosophy that ensures AI technologies operate transparently, equitably, and in line with societal values.
Example: An AI hiring tool that includes bias checks to prevent discrimination against minority candidates.
Expert System
An AI system that mimics the decision-making abilities of a human expert in a specific domain using rules and logic.
Example: An expert system used in agriculture to recommend crop treatments based on soil data and pest history.
Explainable AI (XAI)
AI systems designed to make their internal processes and decisions understandable to humans, increasing trust and accountability.
Example: A medical diagnostic AI that not only provides a recommendation but also explains which symptoms led to that conclusion.
Few-shot Learning
A machine learning method where a model is trained or fine-tuned using only a small number of labelled examples.
Example: Customising an LLM to write legal emails after showing it just 10 examples.
Fine-tuning
The process of taking a pre-trained model and further training it on a new, smaller dataset to specialise it for a specific task.
Example: Fine-tuning a general LLM like GPT on internal legal documents to create a legal drafting assistant.
Foundation Model
A large-scale model trained on diverse and broad data that can be adapted to many downstream tasks.
Example: GPT-4 and PaLM 2 are foundation models capable of summarisation, Q&A, translation, and more.
Fuzzy Logic
A form of logic that deals with approximate values instead of fixed true/false (binary) logic, useful for reasoning under uncertainty.
Example: Used in climate control systems to adjust temperature based on fuzzy inputs like 'a bit hot' or 'very cold'.
Generative Adversarial Network (GAN)
A generative model architecture where two networks — a generator and a discriminator — compete to improve output quality.
Example: GANs are used to create deepfake videos or generate realistic product photos from sketches.
Generative AI
A category of artificial intelligence that can create new content — such as text, images, music, or video — from training data.
Example: ChatGPT generating blog posts or Midjourney creating digital artwork from textual prompts.
Generative Pre-trained Transformer (GPT)
A class of large language models developed by OpenAI that uses transformer architecture and is pre-trained on vast amounts of text data to perform a variety of language tasks.
Example: GPT-4 is capable of writing essays, translating languages, and summarising documents with minimal prompting.
Genetic Algorithm
An optimisation technique inspired by natural selection where solutions evolve over time through mutation, crossover, and selection.
Example: Used to design efficient neural network architectures by simulating survival of the fittest.
Hallucination
The generation of plausible-sounding but factually incorrect or nonsensical content by an AI model.
Example: A language model invents a non-existent citation or provides false historical facts.
Heuristic
A practical approach to problem-solving that doesn't guarantee a perfect solution but is sufficient for immediate goals.
Example: Using a rule of thumb to estimate delivery time in a logistics AI system.
Hyperparameter
A configuration value set before training a machine learning model, such as learning rate or number of layers.
Example: Adjusting the batch size from 32 to 128 to improve training speed and model performance.
Inference
The process of using a trained machine learning model to make predictions or generate outputs from new input data.
Example: Using a fine-tuned GPT model to draft emails for a customer support team.
Intent Detection
A task in natural language understanding where the system identifies the user's goal or purpose in a message.
Example: In a chatbot, recognising 'I want to book a flight' as a travel booking intent.
Internet of Things (IoT)
A network of interconnected physical devices embedded with sensors, software, and other technologies to collect and exchange data.
Example: Smart thermostats and fridges that report usage data and adjust settings using AI analytics.
Interpretability
The extent to which a human can understand the internal mechanics of a machine learning model and its decision-making process.
Example: A decision tree is more interpretable than a deep neural network because its decisions are traceable.
Jupyter Notebook
An open-source interactive computing environment that allows users to write code, visualise outputs, and document analysis in a single interface.
Example: Data scientists use Jupyter Notebooks to prototype machine learning models and share results.
K-Nearest Neighbours (KNN)
A simple, non-parametric machine learning algorithm used for classification and regression. It makes decisions based on the closest training examples in the feature space.
Example: To classify a new fruit as an apple or pear, KNN checks which labelled fruits are closest in shape and colour.
Knowledge Graph
A data structure that uses nodes and edges to represent and store interlinked descriptions of entities and their relationships.
Example: Google's knowledge panel is powered by a knowledge graph that connects entities like people, places, and events.
Language Learning Model Optimisation (LLMO)
Techniques used to improve the performance, efficiency, or adaptability of large language models for specific tasks or domains.
Example: Using quantisation and instruction tuning to optimise an LLM for enterprise use.
Large Language Model (LLM)
A type of deep learning model trained on vast amounts of textual data capable of generating, understanding, and reasoning with human language.
Example: ChatGPT and Claude are LLMs trained to assist in writing, coding, and answering questions.
Latent Space
A high-dimensional abstract representation where similar inputs are grouped close together, used in generative models and embeddings.
Example: In image generation, manipulating the latent space can change features like brightness or emotion.
Learning Rate
A key hyperparameter in training that controls how much model weights are adjusted with respect to the loss gradient.
Example: A high learning rate may lead to overshooting minima, while too low a rate slows training progress.
Machine Learning (ML)
A branch of AI that enables systems to learn from data and improve performance without being explicitly programmed.
Example: Spam filters use machine learning to classify emails as spam or not based on past examples.
Model Drift
A phenomenon where the accuracy of a model declines over time due to changes in data or environment.
Example: A fraud detection model becomes less accurate as fraud tactics evolve.
Model Training
The process of feeding data to a machine learning model and adjusting its parameters to minimise error.
Example: Training a recommendation engine on customer purchase history to suggest new products.
Multimodal AI
AI systems capable of processing and integrating multiple types of data such as text, images, audio, and video.
Example: A model like GPT-4 Vision that can read text and interpret images at the same time.
Natural Language Processing (NLP)
A subfield of AI focused on the interaction between computers and human (natural) languages. It enables machines to read, understand, and respond in human language.
Example: NLP is used in voice assistants, language translation apps, and chatbots.
Neural Network
A machine learning model inspired by the structure of the human brain, composed of layers of interconnected nodes (neurons).
Example: Neural networks are behind deep learning models used in image and speech recognition.
Noise
Random or irrelevant information in data that can obscure meaningful patterns and negatively affect model performance.
Example: Sensor errors or typo-filled data entries can be considered noise.
Ontology
A structured framework that categorises and defines relationships among concepts within a domain, often used in semantic AI systems.
Example: An ontology in healthcare might define how symptoms relate to diseases and treatments.
Overfitting
A modelling error where a machine learning model captures noise in the training data and performs poorly on new data.
Example: A model that memorises training answers but can't handle unseen test data is overfitted.
Predictive Analytics
The use of data, algorithms, and AI to identify the likelihood of future outcomes based on historical data.
Example: Retailers use predictive analytics to forecast demand for certain products.
Pre-training
The process of initially training a model on a large, general dataset before fine-tuning it for specific tasks.
Example: GPT models are pre-trained on large corpora before being customised for customer service chatbots.
Prompt Engineering
The art and science of crafting effective prompts to steer the output of large language models.
Example: Adding system instructions like 'Reply as a polite tutor' is an example of prompt engineering.
Quantisation
A model compression technique that reduces the number of bits used to represent weights and activations, enhancing efficiency.
Example: Quantising a model from 32-bit to 8-bit improves performance on mobile devices.
Quantum Computing
A new paradigm of computing based on quantum mechanics, which holds potential for exponential processing capabilities.
Example: Quantum computing may one day accelerate AI training beyond classical limits.
Reasoning Engine
A system in AI that derives logical conclusions from a set of facts or data using rules or inference algorithms.
Example: An AI diagnosis tool uses a reasoning engine to deduce possible medical conditions based on symptoms.
Reinforcement Learning (RL)
An area of machine learning where agents learn by interacting with their environment to maximise cumulative rewards.
Example: A robot learning to walk by trial and error using RL techniques.
Reinforcement Learning with Human Feedback (RLHF)
A learning method where human preferences guide the AI's reward signal, often used in fine-tuning language models.
Example: ChatGPT was trained with RLHF to produce more helpful and safe responses.
Retrieval-Augmented Generation (RAG)
A method combining information retrieval with generation, where an LLM fetches relevant documents to improve its response.
Example: An AI assistant retrieves and cites product specs while generating an answer to a technical question.
Self-Supervised Learning
A training approach where the model learns patterns by generating its own labels from raw data, reducing reliance on human-annotated data.
Example: BERT is trained with self-supervised learning by predicting missing words in text.
Semantic Search
A search technique that understands user intent and contextual meaning, not just keyword matching.
Example: Searching 'how to fix a leaking tap' returns guides even if the term 'leaking tap' isn't present in the document.
Sentiment Analysis
The process of identifying emotions, opinions, or attitudes in text, often classifying as positive, negative, or neutral.
Example: Analysing tweets to gauge public reaction to a new product.
Stochastic
Involving randomness or probabilistic behaviour, often used in generative AI and optimisation algorithms.
Example: The output of GPT-4 varies for the same input due to its stochastic decoding process.
Strong AI
Also known as Artificial General Intelligence (AGI), refers to machines with human-level cognitive capabilities across all domains.
Example: A future AI that can autonomously write novels, plan cities, and solve ethical dilemmas equally well.
Super Artificial Intelligence (SAI)
A theoretical AI far surpassing human intelligence in all aspects—reasoning, creativity, emotional intelligence, etc.
Example: An SAI could theoretically develop new sciences and philosophies independently.
Supervised Learning
A machine learning technique where models are trained on labelled data to learn input-output mappings.
Example: Teaching a model to classify emails as spam or not using historical examples.
Synthetic Data
Artificially generated data that simulates real-world data, often used for training when real data is scarce or sensitive.
Example: Creating synthetic medical images to train diagnostic models without violating patient privacy.
Token
A unit of text processed by LLMs—typically a word or word piece.
Example: The sentence 'Hello world!' is split into 3 tokens: 'Hello', 'world', and '!'.
Tokenisation
The process of breaking down text into tokens for processing by a model.
Example: In NLP, 'ChatGPT is great' becomes ['Chat', 'G', 'PT', 'is', 'great'].
Transfer Learning
Using knowledge from one task to enhance learning on another related task, reducing training time and data needs.
Example: Fine-tuning a model trained on English text to perform sentiment analysis in another language.
Underfitting
When a model is too simplistic to capture the patterns in the training data, resulting in poor performance.
Example: A linear model trying to predict complex image classifications may underfit.
Unsupervised Learning
A learning approach where models identify patterns or clusters in unlabelled data.
Example: Grouping customers based on purchasing behaviour without predefined labels.
User Intent
The goal or purpose behind a user's query or interaction.
Example: A user typing 'how to bake a cake' likely intends to find a recipe.
Validation Set
A subset of data used to evaluate model performance during training and tune hyperparameters.
Example: Used to detect overfitting before final testing.
Vector Database
A database designed to store and search vector embeddings used in AI tasks like similarity search and RAG.
Example: Pinecone and Weaviate are vector databases for storing text or image embeddings.
Vector Embedding
A numeric representation of data that preserves semantic meaning and relationships in a vector space.
Example: The words 'king' and 'queen' have similar embeddings with subtle gender differences.
Virtual Assistant
An AI-powered software agent that helps users complete tasks via conversation or voice commands.
Example: Siri, Alexa, and Google Assistant are popular virtual assistants.
Voice Recognition
Technology that interprets and converts spoken language into text or action.
Example: Voice typing and voice commands rely on voice recognition systems.
Weak AI
AI systems designed to perform a narrow, specific task without general intelligence.
Example: A chess-playing AI that cannot understand language or drive a car is an example of weak AI.
Web Scraping
Automated extraction of information from websites, often used to collect training data or monitor content.
Example: Scraping real estate listings to train a property valuation model.
Weight
A parameter in neural networks that determines the strength of influence one node has on another.
Example: Weights adjust during training to minimise the model's error.
Whisper
A speech-to-text model developed by OpenAI capable of transcribing audio in multiple languages.
Example: Whisper can transcribe lectures and podcasts with high accuracy.
YAML
A human-readable format for data serialisation, commonly used for configuration files in machine learning workflows.
Example: Defining model parameters in a YAML file for training in PyTorch.
Zero-shot Learning
The ability of a model to perform tasks it has never been explicitly trained on by leveraging general knowledge.
Example: A model answering legal questions despite not being trained specifically on legal data.
Zettabyte
A unit of digital data equal to one sextillion (10^21) bytes, often used to describe the scale of internet data.
Example: The global internet traffic surpassed 1 zettabyte per year by 2016.
