Eleven short chapters, no jargon left unexplained. Read top to bottom for a full mental model, or jump to the one you need — and draw the illustration whenever a picture helps.
AI Concept Flashcards
Flip through all 21 flashcards — a rapid revision system covering 300+ AI concepts across the full stack.
1 / 21AI Foundations
Artificial intelligence is software that learns patterns from examples instead of being told every rule. Classic software follows instructions a person wrote; AI writes its own rules from data and then applies them to new situations.
Traditional software is a recipe. AI is a cook who has tasted ten thousand dishes and can now improvise.
- Narrow AI
- Good at one job — translation, fraud detection, image tagging. Everything in use today is narrow AI.
- AGI
- A hypothetical system as broadly capable as a person across any task. Not here yet.
- Algorithm
- The step-by-step method the system uses to learn or decide.
- Model
- The trained result — the learned patterns saved as numbers.
- Data
- The examples the model learns from. Quality of data caps quality of model.
Machine Learning
Machine learning is the practical craft of getting a model to learn from data. You show it examples, it guesses, you measure how wrong it was, and it nudges itself closer. Repeat millions of times.
Learning to shoot hoops: shoot, see how far off you were, adjust the wrist, shoot again.
- Supervised learning
- Learning from labelled examples — this email is spam, this one isn't.
- Unsupervised learning
- Finding structure in unlabelled data, like grouping customers who behave alike.
- Reinforcement learning
- Learning by trial, reward and penalty — how game-playing and robot systems improve.
- Features
- The input signals the model looks at: price, age, word counts, pixels.
- Overfitting
- Memorising the training set instead of learning the pattern — great in practice runs, poor in the real world.
Deep Learning
Deep learning stacks many layers of simple maths units — neurons — so the system can build understanding in steps. Early layers spot edges and fragments; later layers assemble them into faces, sentences and meaning.
A relay team: each runner refines the baton a little before passing it on, and the last one crosses the line with the answer.
- Neuron
- A tiny unit that takes numbers in, weighs them, and passes a number out.
- Weights
- The learned strengths of each connection — this is what training actually changes.
- Backpropagation
- The method that traces the error backwards and tells every weight how to adjust.
- GPU
- The chip that makes the huge parallel maths of deep learning practical.
- Neural architecture
- The shape of the network — how many layers, how they connect.
Generative AI
Most earlier AI classified or predicted. Generative AI creates: text, images, audio, video, code. It learns the statistical shape of a domain and then samples brand-new things that fit that shape.
A musician who has absorbed thousands of songs and can now compose a new one in the same style.
- Prompt
- The instruction you give. Clear context and a stated goal beat clever wording.
- Diffusion model
- Image generators that start from noise and repeatedly clean it into a picture.
- Multimodal
- A model that handles more than one kind of input or output — text plus images plus audio.
- Hallucination
- A confident, fluent answer that is simply wrong. Always verify facts that matter.
- Temperature
- A dial for randomness: low is predictable and safe, high is creative and loose.
Large Language Models
An LLM is a very large network trained to predict the next chunk of text. That single skill, at scale, produces summarising, translating, reasoning and coding as side effects.
An extraordinarily well-read colleague who answers instantly, sounds sure of themselves, and occasionally invents a source.
- Token
- A word fragment. Models read and bill in tokens, not words.
- Context window
- How much text the model can hold in view at once. Older text falls off the desk.
- Attention
- The mechanism that lets each token look at the other tokens that matter to it.
- Fine-tuning
- Extra training on your own examples so the model adopts your style or domain.
- System prompt
- Standing instructions that set the model's role and rules before the chat starts.
RAG — Retrieval-Augmented Generation
A model only knows what it was trained on. RAG fixes that by searching your documents first and handing the relevant passages to the model along with the question, so the answer is grounded in your material.
An open-book exam: same student, but now allowed to look up the right page before answering.
- Embedding
- Text turned into a list of numbers so similar meanings sit close together.
- Vector database
- The store that finds the closest passages by meaning rather than keyword.
- Chunking
- Splitting documents into passages small enough to retrieve precisely.
- Grounding
- Forcing the answer to be based on retrieved sources, usually with citations.
- Re-ranking
- A second pass that reorders retrieved passages by true relevance.
AI Agents
An agent is a model given a goal, a set of tools and permission to loop. It plans, calls a tool, reads the result, and decides the next move until the task is done — rather than answering once and stopping.
A chatbot answers your question. An agent is an intern who goes away, does the work, and comes back with it finished.
- Tool calling
- The model invoking real functions — search, database, email, code execution.
- Planning
- Breaking a goal into ordered steps before acting.
- Memory
- Carrying state across steps and sessions so the agent doesn't restart from zero.
- Human in the loop
- An approval gate before consequential actions like sending or paying.
- Multi-agent
- Several specialised agents cooperating, often with one coordinating.
AI Protocols
Protocols are the shared plugs and sockets of the AI world. They let models, tools and other agents connect without a custom integration being built every single time.
USB-C for AI: one standard connector instead of a drawer full of proprietary cables.
- API
- The basic contract for calling a service: send a request, get a structured response.
- MCP
- Model Context Protocol — a standard way to expose tools and data sources to a model.
- Function schema
- A machine-readable description of a tool so the model knows how to call it.
- Agent-to-agent
- Emerging standards for agents from different vendors to negotiate and hand off work.
- Webhooks
- The reverse direction — a system pushing events to your app as they happen.
AI Infrastructure
Everything under the model: the chips that do the maths, the pipelines that move data, and the serving layer that answers users quickly and affordably at scale.
The model is the engine. Infrastructure is the fuel, the road, the pit crew and the electricity bill.
- Training vs inference
- Training builds the model once at great cost; inference runs it millions of times.
- GPU / accelerator
- Specialised chips that do huge parallel maths — the scarce resource in AI.
- Latency
- How long a user waits. Often matters more to adoption than raw accuracy.
- Caching
- Reusing previous work to cut cost and speed up repeated requests.
- Observability
- Logging prompts, outputs, cost and failures so problems are visible early.
AI Strategy & Business
The hard part is rarely the model. It is choosing use cases with real economics, redesigning the workflow around them, and building the habits and data that compound over time.
Buying a race car doesn't make you fast. The track, the crew and the driver's practice do.
- Use-case selection
- High volume, tolerant of small errors, measurable value. Start there.
- Build vs buy
- Buy the commodity, build only what genuinely differentiates you.
- Workflow redesign
- Value comes from changing the process, not from bolting AI onto the old one.
- Data moat
- Proprietary data and feedback loops competitors cannot simply copy.
- Adoption
- Training, incentives and trust. Unused capability produces zero return.
AI Safety & Governance
Governance is how you keep AI useful without causing harm: knowing where it is used, what data it touches, who approved it, and what happens when it gets something wrong.
Brakes and seatbelts. They are not the opposite of speed — they are what makes speed survivable.
- Bias
- Skewed training data producing unfair outcomes for some groups.
- Privacy
- Controlling what personal or confidential data ever reaches a model.
- Explainability
- Being able to show why a decision was made, especially in regulated settings.
- Prompt injection
- Hidden instructions in content that hijack an agent's behaviour.
- Accountability
- A named human owner for every automated decision, with audit trails.