If terms like LLMs, Agent, deep learning make you feel like everyone secretly attended an AI meeting without inviting you, this guide is for you. Let’s decode the jargon before the robots fully take over.
Contents
- Core AI Concepts
- Data & Training Concepts
- Learning Methods
- Modern LLM Concepts
- AI Applications
- Development & Infrastructure
- Popular AI Tools & Platforms
- AI Safety & Society
1. Core AI Concepts
| Term | Acronym | Definition | Example |
|---|---|---|---|
| Artificial Intelligence | AI | Machines performing tasks that normally require human intelligence | ChatGPT, Claude |
| Machine Learning | ML | Systems that learn patterns from data without being explicitly programmed | Netflix recommendations |
| Deep Learning | — | An advanced form of machine learning using layered neural networks | Face recognition |
| Neural Network | — | A computational model loosely inspired by the structure of the human brain | Image classification |
| Foundation Model | — | A large AI model trained on broad data that can be adapted to many tasks | GPT-4, Claude, Gemini |
| Generative AI | — | AI that creates new content — text, images, audio, code, and more | AI-generated images, ChatGPT |
| Large Language Model | LLM | AI trained on massive text datasets, capable of generating and understanding language | GPT-4, Claude, Llama |
| Agentic AI | — | AI that can autonomously plan and execute multi-step tasks, often using tools | AutoGPT, Claude with tool use |
| AI Agent | — | A single AI system that perceives its environment, makes decisions, and takes actions to achieve a goal | A coding agent that writes, tests, and fixes code on its own |
| Multimodal AI | — | AI that can process and generate multiple types of content — text, images, and audio together | GPT-4o image analysis |
| Artificial General Intelligence | AGI | A hypothetical AI that matches or exceeds human-level intelligence across all tasks — does not yet exist | Often discussed as a long-term goal in AI research |
| Chatbot | — | A software application that simulates conversation with users, often powered by an LLM | Customer support bots, ChatGPT |
| AI Alignment | — | The challenge of ensuring AI systems behave in ways that are safe and consistent with human intentions and values | Preventing an AI from pursuing goals that harm people |
2. Data & Training Concepts
| Term | Acronym | Definition | Example |
|---|---|---|---|
| Dataset | — | A structured collection of data used to train or evaluate AI models | A database of labelled customer photos |
| Training Data | — | The specific data an AI model learns from during training | Millions of labelled images |
| Model | — | The trained AI system that makes predictions or generates outputs | A spam detector |
| Parameters | — | The internal numerical values a model learns during training; more parameters generally means more capability | GPT-4 has hundreds of billions |
| Weights | — | Another word for parameters — the numerical values stored inside a model after training | "Downloading model weights" means downloading the trained model itself |
| Token | — | Small units of text that AI processes — roughly a word or part of a word | "running" = 1 token; "unbelievable" = 3 tokens |
| Embeddings | — | Numerical representations of text that capture meaning and relationships | Used in semantic search |
| Vector Database | — | A database that stores embeddings so AI can quickly retrieve relevant information | Pinecone, Weaviate |
| Pre-training | — | The initial large-scale training phase where a model learns from a huge, general dataset before any specialisation | Training an LLM on the entire internet |
| Fine-Tuning | — | Training an existing pre-trained model further on specialised data to improve it for a specific task | Training a general model on medical records to create a medical chatbot |
| Transfer Learning | — | Reusing a model trained on one task as the starting point for a different but related task | Using an image model trained on photos to kickstart a medical imaging model |
| Synthetic Data | — | Artificially generated data used to train or test models when real data is scarce or sensitive | Generating fake patient records to train a healthcare AI |
| Epoch | — | One complete pass through the entire training dataset during model training | Training for 10 epochs means the model sees all the data 10 times |
| Batch Size | — | The number of training examples processed together in one step | A batch size of 32 means the model updates its weights after every 32 examples |
| Gradient Descent | — | The core algorithm that adjusts a model's weights during training to minimise errors | How a neural network "learns" by slowly correcting its mistakes |
| Inference | — | The process of a trained model generating outputs in response to new inputs | An AI answering your question |
| Model Card | — | A short document published alongside an AI model describing what it does, how it was trained, and its limitations | Hugging Face model cards |
3. Learning Methods
| Term | Acronym | Definition | Example |
|---|---|---|---|
| Supervised Learning | — | Training a model using labelled input-output pairs | Spam detection (email → spam/not spam) |
| Unsupervised Learning | — | Finding patterns in data without predefined labels | Customer segmentation |
| Reinforcement Learning | — | Training a model through a system of rewards and penalties for its actions | Game-playing AI like AlphaGo |
| Reinforcement Learning from Human Feedback | RLHF | A training technique where human raters score AI outputs, and the model learns to produce responses humans prefer | How ChatGPT and Claude were fine-tuned to be helpful and safe |
| Classification | — | A model predicting which category an input belongs to | Fraud detection (fraudulent vs. legitimate) |
| Regression | — | A model predicting a continuous numeric value | Predicting house prices |
| Clustering | — | Grouping similar data points together without predefined labels | Market segmentation |
| Data Augmentation | — | Artificially expanding a training dataset by creating modified versions of existing data | Flipping, rotating, or cropping images to give a model more variety to learn from |
| Overfitting | — | When a model memorises training data too closely and performs poorly on new data | A model that aces training tests but fails in the real world |
| Underfitting | — | When a model is too simple to learn the underlying patterns in the data | A model that makes weak or random predictions |
| Cross-Validation | — | A technique for testing how well a model generalises by training and evaluating it on different subsets of data | Splitting data into 5 "folds" and rotating which one is used for testing |
4. Modern LLM Concepts
| Term | Acronym | Definition | Example |
|---|---|---|---|
| Natural Language Processing | NLP | The field of AI focused on enabling machines to understand and generate human language | Language translation, sentiment analysis |
| Prompt | — | The instruction or question you give to an AI model | "Write a blog post about climate change" |
| System Prompt | — | A hidden set of instructions given to an AI model before the conversation starts, used to set its behaviour, tone, or rules | A company using a system prompt to make Claude respond only about their product |
| Prompt Engineering | — | The practice of crafting and refining prompts to get better, more reliable AI outputs | Using structured formatting or examples in your prompt |
| Zero-Shot Prompting | — | Asking an AI to complete a task with no examples provided | "Translate this sentence to French." |
| Few-Shot Prompting | — | Giving an AI a small number of examples before asking it to complete a task | Showing 2–3 example summaries before asking it to summarise a new article |
| Chain-of-Thought Prompting | — | Encouraging an AI to reason step by step before giving a final answer, which improves accuracy on complex tasks | Adding "Think step by step" to a maths or logic prompt |
| Transformer | — | An attention-based neural network architecture that is the foundation of most modern LLMs | GPT, Claude, and Gemini are all transformer-based models |
| Attention Mechanism | — | The part of a transformer that lets the model focus on the most relevant parts of the input when generating each word | How a model knows "it" in "The cat sat because it was tired" refers to the cat |
| Retrieval-Augmented Generation | RAG | A technique that combines AI generation with real-time retrieval of relevant documents or data | A chatbot that searches your company's PDF documents before answering |
| Function Calling | — | A feature that lets an LLM trigger external tools or APIs — such as searching the web or running code — as part of its response | An AI assistant that calls a weather API to answer "Will it rain tomorrow?" |
| Context Window | — | The maximum amount of text an AI can read and "remember" in a single interaction | A model with a 200,000-token context window can read roughly 150,000 words at once |
| Temperature | — | A setting that controls how predictable or creative an AI's output is. Low = more focused; high = more varied and creative | Set low for factual Q&A; set high for creative writing |
| Top-p Sampling | — | A setting that controls AI output variety by limiting the pool of possible next words to a cumulative probability threshold | Often used alongside temperature to tune output quality |
| Hallucination | — | When an AI confidently states something that is factually incorrect or entirely made up | An AI inventing a citation to a research paper that doesn't exist |
| Guardrails | — | Rules or filters applied to an AI to prevent it from producing harmful, off-topic, or inappropriate outputs | A customer service bot that refuses to discuss competitors |
| Jailbreak | — | A technique used to trick an AI into bypassing its safety guidelines or guardrails | Roleplaying prompts designed to make an AI ignore its rules |
| Prompt Injection | — | An attack where malicious instructions are hidden in content the AI reads, trying to hijack its behaviour | A webpage that contains hidden text telling a browsing AI to send your data elsewhere |
5. AI Applications
| Term | Acronym | Definition | Example |
|---|---|---|---|
| Computer Vision | — | AI that can interpret and understand visual information from images and video | CCTV object recognition, medical imaging |
| Speech Recognition | — | AI that converts spoken audio into text | Siri, Google Voice |
| Text-to-Speech | TTS | AI that converts written text into natural-sounding spoken audio | ElevenLabs, Google Text-to-Speech |
| Text-to-Image | — | AI that generates images from a text description | DALL·E, Midjourney, Stable Diffusion |
| Sentiment Analysis | — | AI that identifies the emotional tone of a piece of text — positive, negative, or neutral | Analysing customer reviews to gauge satisfaction |
| Recommendation System | — | AI that predicts what a user might want to see or do next, based on past behaviour | YouTube's "Up Next" queue, Spotify's Discover Weekly |
| Automation | — | Using AI to reduce or eliminate manual, repetitive tasks | Auto-generating reports, routing support tickets |
| Explainable AI | XAI | AI systems designed so that their reasoning and decisions can be understood by humans | A loan-rejection system that shows which factors (income, credit score) influenced the decision |
| AI Ethics | — | The principles and practices for developing and deploying AI responsibly and fairly | Preventing bias, ensuring transparency, protecting privacy |
| Bias | — | When an AI system produces unfair or skewed outcomes, often because of imbalanced training data | A hiring tool that systematically ranks male applicants higher than equally qualified female applicants |
| Red Teaming | — | Deliberately trying to break or misuse an AI system to find safety vulnerabilities before release | Researchers probing a model with harmful prompts to see how it responds |
6. Development & Infrastructure
| Term | Acronym | Definition | Example |
|---|---|---|---|
| Application Programming Interface | API | A defined way for software systems to communicate with each other | The OpenAI API lets developers build apps powered by GPT |
| API Key | — | A private authentication token that identifies you when making API calls | You paste your API key into code to give it permission to use a service |
| Graphics Processing Unit | GPU | Specialised hardware that dramatically accelerates AI training and inference workloads | NVIDIA A100 GPUs used in data centres |
| Tensor Processing Unit | TPU | Hardware designed specifically for AI workloads, developed by Google | Used to train Google's AI models |
| Cloud Computing | — | Running applications and storing data on remote internet-connected servers rather than locally | AWS, Azure, Google Cloud |
| Edge AI | — | Running AI models directly on a local device rather than in the cloud | AI on a smart camera that processes footage without sending it to a server |
| Latency | — | The delay between sending a request to an AI and receiving a response | A model with low latency feels instant; high latency feels slow |
| Model Quantisation | — | A technique that reduces a model's size and memory usage by representing its weights with less precision, making it faster and cheaper to run | Running a compressed version of Llama on a laptop instead of a server |
| Open Source Model | — | An AI model whose weights and/or code are publicly available for anyone to use and modify | Meta's Llama models |
| Hugging Face | — | A popular platform for sharing, discovering, and running open-source AI models and datasets | Often called "the GitHub of AI" |
| Benchmark | — | A standardised test used to evaluate and compare AI model performance | MMLU, HumanEval |
7. Popular AI Tools & Platforms
| Category | Tools |
|---|---|
| AI Assistants | ChatGPT, Claude, Gemini |
| AI Model Hub | Hugging Face |
| Image Generation | DALL·E, Midjourney, Stable Diffusion |
| Coding & Development | Python, Jupyter Notebook |
| Data Analysis | Pandas, NumPy |
| Data Visualisation | Power BI, Tableau |
| Machine Learning Frameworks | Scikit-learn, TensorFlow, PyTorch |
| Cloud Platforms | AWS, Azure, GCP |
| Workflow & AI Orchestration | LangChain, n8n |
| API Testing | Postman, SoapUI |
8. AI Safety & Society
These terms come up constantly in news, policy, and real-world AI discussions. Every beginner should know them.
| Term | Acronym | Definition | Example |
|---|---|---|---|
| AI Safety | — | The field focused on ensuring AI systems behave reliably and don't cause unintended harm as they become more capable | Research into preventing models from pursuing dangerous goals |
| AI Alignment | — | The challenge of ensuring AI systems pursue goals that are actually consistent with human intentions and values | Ensuring a powerful AI optimises for human wellbeing, not just task completion |
| Deepfake | — | AI-generated video, audio, or images that realistically depict someone saying or doing something they never did | Synthetic video of a public figure making a fake speech |
| Copyright & IP | — | Legal questions about who owns AI-generated content and whether training data was used lawfully | Ongoing lawsuits between AI companies and artists or publishers |
| Data Privacy | — | The concern about how personal data is collected, stored, and used to train AI models | Whether your chat history is used to improve a model |
| AI Regulation | — | Government laws and policies designed to govern how AI is developed and deployed | The EU AI Act, US executive orders on AI |
| Carbon Footprint of AI | — | The energy and environmental cost of training and running large AI models | Training GPT-4 is estimated to have used millions of kilowatt-hours of electricity |
| Human-in-the-Loop | HITL | A system design where a human reviews or approves AI decisions before they take effect | A doctor reviewing an AI's diagnosis before acting on it |
Top comments (0)