AI Concepts

AI is easier to understand as a stack of layers than as a list of buzzwords. Each layer builds on the ones beneath it: models are trained on data, language models turn that training into fluent text, retrieval and tools connect them to real knowledge, and evaluation and governance keep the whole thing trustworthy.

This page is the map. Every concept gets a plain-language definition, and wherever I have written an in-depth guide, the concept links straight to it. New guides will be added here as they are published.

---

Deep dives

In-depth guides on individual concepts, written from basic to advanced.

---

The concept map

The seven layers of the stack, bottom to top. Where a concept is covered in depth, the last column links to the guide.

01 · Foundations

What AI is, and how machines learn from data rather than from hand-written rules.

Concept	In plain words
Artificial intelligence	Systems that perform tasks that normally need human judgement: recognising, predicting, generating and deciding.
Machine learning	Learning patterns from examples instead of following rules a programmer wrote by hand.
Deep learning	Machine learning with many-layered neural networks. It powers modern vision, speech and language models.
Neural network	Layers of simple connected units. Training adjusts the strength (weight) of each connection until the outputs are useful.
Supervised, unsupervised, reinforcement learning	Learning from labelled examples, from structure in unlabelled data, and from rewards for good actions.
Training vs inference	Training sets the model’s weights from data: expensive and occasional. Inference uses the finished model to answer: cheap and constant.

02 · Language models

The engine behind chat assistants: what a large language model is, and the limits that shape how you use it.

Concept	In plain words
Large language model (LLM)	A model trained on vast amounts of text to predict the next piece of text. Scale gives it broad language, knowledge and reasoning ability.
Token	The unit a model reads and writes, roughly a word or part of a word. Limits and pricing are counted in tokens.
Transformer and attention	The architecture behind modern LLMs (introduced in 2017). Attention lets every token weigh the relevance of every other token in the input.
Context window	The maximum amount of text the model can consider at once: instructions, documents, conversation and its own answer.
Temperature	How much randomness goes into word choice. Low for consistent, factual work; higher for variety and brainstorming.
Knowledge cutoff	The date after which the model’s training data stops. Anything later is unknown to it unless supplied.
Hallucination	Fluent, confident output that isn’t supported by facts or sources.
Multimodal model	A model that works with more than text: images, audio, documents or video, as input, output or both.

03 · Working with models

How to get reliable results from a model without changing the model itself.

Concept	In plain words
Prompt engineering	Writing clear instructions with the right context, examples and output format. The fastest, cheapest lever you have.
System prompt	Standing instructions that set the model’s role, rules and tone for a whole conversation or application.
Zero-shot and few-shot	Asking with no examples, versus including a few worked examples of the task in the prompt.
Chain of thought and reasoning models	Working step by step before answering. Reasoning models do this internally, trading speed for accuracy on hard problems.
Structured output	Asking for JSON or another fixed format so that software, not just people, can use the answer.
Context engineering	Deciding what goes into the context window, in what order: instructions, documents, history and tools.

04 · Knowledge & retrieval

Connecting models to knowledge they were never trained on: your documents, policies and data, kept current and permissioned.

Concept	In plain words	Go deeper
Embeddings	Numbers that capture the meaning of text, so similar ideas sit close together even when the words differ.	RAG guide §2
Vector database	A store built to find the embeddings nearest to a query, quickly and at scale.	RAG guide §2
Retrieval-Augmented Generation (RAG)	Find relevant evidence first, then ask the model to answer from it, with citations.	Full guide
Hybrid search and reranking	Combining meaning-based and keyword search, then re-ordering the shortlist for relevance.	RAG guide §4
Knowledge graphs and GraphRAG	Storing entities and their relationships, so questions that span documents can be answered by following the links.	RAG guide §5
Long context vs retrieval	Pasting everything in works for a few known documents; retrieval is needed for large, changing or permissioned collections.	RAG guide §1

05 · Customising models

Changing how a model behaves, rather than what it can see. A useful rule: customise for how it should answer, retrieve for what it should know.

Concept	In plain words
Fine-tuning	Further training on your own examples to change style, format or skill at a specific task.
LoRA and parameter-efficient tuning	Training a small set of add-on weights instead of the whole model: far cheaper, and easy to swap in and out.
RLHF and preference tuning	Training on human (or AI) judgements of which answer is better, to make a model more helpful and safer.
Distillation	Training a smaller, faster model to imitate a larger one on the tasks you care about.
Quantisation	Storing a model’s numbers at lower precision so it runs faster and cheaper, with a small loss in quality.
Open-weight vs hosted models	Downloading and running a model yourself, versus calling a provider’s model through an API.

06 · Agents & tools

Models that act, not just answer: calling tools, working through multi-step tasks and checking their own progress.

Concept	In plain words	Go deeper
Tool use (function calling)	The model asks, in a structured way, for a function or API to be run. The application runs it and hands back the result.
AI agent	A model working in a loop of plan, act, observe and reflect until a goal is met or a human is needed.
Model Context Protocol (MCP)	An open standard, introduced by Anthropic in 2024, for connecting AI applications to tools and data sources through one common interface.
Agentic RAG	An agent that plans its searches, routes them to the right sources and re-queries when evidence is missing.	RAG guide §5
Multi-agent systems	Several specialised agents, such as a planner, a researcher and a reviewer, coordinating on one task.
Human in the loop	Approval checkpoints before an agent takes consequential or irreversible actions.

07 · Quality & safety

What turns an impressive demo into a system people can depend on.

Concept	In plain words	Go deeper
Evaluations (evals)	Repeatable tests of a system against a set of questions with known good answers, run whenever anything changes.	RAG guide §6
LLM-as-judge	Using a model to grade outputs at scale. Useful, but calibrate it against human reviewers.	RAG guide §6
Grounding and citations	Tying every claim to a source the reader can check, and allowing “I don’t have enough evidence”.	RAG guide §2
Guardrails	Checks on what goes into and comes out of a model: content filters, policy rules, format validation.
Prompt injection	Malicious instructions hidden in content the model reads, such as a web page or document, that try to hijack its behaviour.	RAG guide §7
Observability	Tracing each request, with its inputs, sources, cost, latency and feedback, so failures can be reproduced and fixed.	RAG guide §8
Responsible AI and governance	Fairness, privacy, transparency and accountability, guided by frameworks such as the NIST AI Risk Management Framework and the EU AI Act.

---