The map

The World of AI

Every field nests inside the one before it. 95 terms across five rings, each with its own explainer.

Artificial Intelligence

The outermost ring: any system that performs tasks we would call intelligent, however it is built.

Reinforcement Learning

An agent learns by acting and receiving rewards, rather than from labelled examples.

Read →

Speech Recognition

Turning spoken audio into text, now usually with a single end-to-end neural network.

Read →

Emergent Behavior

Capabilities that appear at scale without being trained for directly.

Read →

Augmented Programming

Writing software with a model as a collaborator — completion, review and refactoring.

Read →

AI Ethics

Fairness, accountability, privacy and harm — who is affected when the system is wrong.

Read →

Expert Systems

Rule-based reasoning from hand-written knowledge — the dominant approach before statistical learning.

Read →

Computer Vision

Extracting meaning from images and video — detection, segmentation, tracking, recognition.

Read →

Symbolic AI

AI that represents knowledge using symbols, logic and explicit rules.

Read →

Automated Reasoning

The use of logic to automatically derive conclusions from known information.

Read →

Constraint Satisfaction

Solving problems by finding values that satisfy a set of restrictions.

Read →

Natural Language Processing

AI techniques that enable machines to understand and generate human language.

Read →

Robotics

The combination of perception, reasoning and control in physical machines.

Read →

Knowledge Graphs

Networks that represent entities and the relationships between them.

Read →

Multi-Agent Systems

Systems in which multiple intelligent agents communicate and coordinate.

Read →

Evolutionary Computation

Optimisation techniques inspired by biological evolution and natural selection.

Read →

Swarm Intelligence

Collective problem-solving inspired by ants, birds and other social organisms.

Read →

Machine Learning

Systems that improve from data rather than from rules written by hand.

Supervised Learning

Learning a mapping from inputs to known answers. Most production ML is still this.

Read →

Unsupervised Learning

Finding structure with no labels — grouping, compressing and spotting the unusual.

Read →

K-Means

Splits data into k groups by repeatedly moving cluster centres to the mean of their members.

Read →

K-Nearest Neighbours

Classifies a point by the labels of the points closest to it. No training step at all.

Read →

Logistic Regression

A linear model that outputs a probability. The baseline every classification project should start from.

Read →

Linear Regression

Fits a straight-line relationship between inputs and a continuous outcome.

Read →

PCA

Rotates data onto the axes that carry the most variance, so you can drop the rest.

Read →

Hypothesis Testing

Deciding whether an observed difference is real or noise.

Read →

Decision Trees

Splits data by asking one question at a time. The building block of forests and boosting.

Read →

Support Vector Machines

Finds the boundary with the widest margin between classes, using kernels for non-linear data.

Read →

Semi-Supervised Learning

Learning from a small labelled dataset combined with a larger unlabelled dataset.

Read →

Self-Supervised Learning

Learning useful representations from labels automatically created from the data.

Read →

Active Learning

A training process where the model selects the examples it wants labelled.

Read →

Federated Learning

Training models across decentralised devices without centralising their raw data.

Read →

Anomaly Detection

Identifying unusual observations that differ significantly from normal patterns.

Read →

Naive Bayes

A probabilistic classifier based on Bayes’ theorem and feature independence.

Read →

Random Forests

An ensemble that combines predictions from many decision trees.

Read →

Gradient Boosting

An ensemble method that builds models sequentially to correct earlier errors.

Read →

XGBoost

A highly optimised implementation of gradient-boosted decision trees.

Read →

LightGBM

A fast gradient-boosting framework designed for large structured datasets.

Read →

CatBoost

Gradient boosting designed to handle categorical features effectively.

Read →

DBSCAN

A density-based clustering algorithm that can identify noise and irregular clusters.

Read →

Hierarchical Clustering

Clustering that organises observations into a tree of nested groups.

Read →

Gaussian Mixture Models

Probabilistic models that represent data as a mixture of Gaussian distributions.

Read →

t-SNE

A dimensionality-reduction technique commonly used to visualise complex data.

Read →

UMAP

A fast dimensionality-reduction method that preserves local and global structure.

Read →

Hidden Markov Models

Probabilistic models for sequences with unobserved underlying states.

Read →

Gaussian Processes

Probabilistic models that make predictions together with uncertainty estimates.

Read →

Recommendation Systems

Models that predict which products or content a user may prefer.

Read →

Neural Networks

Models built from layers of simple connected units, loosely inspired by biological neurons.

Perceptron

A single artificial neuron: weighted inputs, a sum, a threshold.

Read →

Backpropagation

Sends the error backwards through the network to find how much each weight is to blame.

Read →

Feed Forward

Data flows in one direction, input to output, with no loops.

Read →

Hopfield Network

A recurrent network that settles into stored patterns — an early model of associative memory.

Read →

Boltzmann Machine

A stochastic network that learns a probability distribution over its inputs.

Read →

Self Organising Maps

Projects high-dimensional data onto a grid while preserving which things are near which.

Read →

Deep Belief Network

Stacked probabilistic layers trained greedily — a stepping stone to modern deep learning.

Read →

Liquid State Machines

A pool of spiking neurons whose rich dynamics make a signal easy to read off.

Read →

Weights and Biases

Trainable parameters that determine how signals move through a neural network.

Read →

Activation Functions

Functions that introduce non-linearity into neural-network computations.

Read →

Gradient Descent

An optimisation method that updates parameters to reduce the model’s loss.

Read →

Adam

An adaptive optimiser commonly used to train neural networks.

Read →

Residual Connections

Shortcut paths that allow information to skip one or more network layers.

Read →

Deep Learning

Neural networks deep enough to learn their own features straight from raw data.

CNN

Convolutional networks slide small filters over an image to learn edges, then shapes, then objects.

Read →

RNN

Processes sequences one step at a time, carrying a hidden state forward as memory.

Read →

LSTM

A recurrent cell with gates that decide what to keep and what to forget.

Read →

Transformers

Attention lets every position look at every other in parallel.

Read →

Auto Encoders

Compresses input to a small code and reconstructs it.

Read →

Deep RL

Reinforcement learning where the policy is a deep network.

Read →

Epochs

One full pass over the training data.

Read →

Attention

A mechanism that lets models focus on the most relevant parts of an input.

Read →

Self-Attention

Attention that models relationships between elements within the same sequence.

Read →

Gated Recurrent Unit

A recurrent network designed to retain information across long sequences.

Read →

ResNet

A deep vision architecture built around residual connections.

Read →

U-Net

An encoder–decoder architecture designed for precise image segmentation.

Read →

Vision Transformers

Transformer models that process images as sequences of visual patches.

Read →

Model Pruning

Removing unimportant parameters or structures from a neural network.

Read →

Quantisation

Representing model parameters with lower-precision numerical formats.

Read →

Mixture of Experts

An architecture that activates selected specialist networks for each input.

Read →

Generative AI

Deep models that produce new text, images, audio or code rather than only classifying.

LLM

A large language model trained to predict the next token — which turns out to be enough.

Read →

GPT

A decoder-only transformer trained on broad text, then adapted to follow instructions.

Read →

BERT

An encoder trained by masking words — still the workhorse for classification and search.

Read →

GANs

Two networks compete: one generates, one judges, until the fakes are convincing.

Read →

RLHF

Reinforcement learning from human feedback — how a raw model becomes a usable assistant.

Read →

QLoRA

Fine-tuning a quantised model through small adapter matrices, so it fits on one GPU.

Read →

Agents

A model in a loop with tools, memory and a goal, deciding its own next action.

Read →

Multimodal AI

One model handling text, images and audio in a shared representation.

Read →

Transfer Learning

Starting from a model trained on something else and adapting it.

Read →

Foundational Model

A large general model trained once and reused across many downstream tasks.

Read →

N-Shot Learning

Teaching a task inside the prompt with a handful of examples — zero, one or few.

Read →

JOIN NOW

Learn these properly, in order

The glossary is the map. The program is the six months of building that makes it stick.

See the curriculum

Join any time · Build AI skills at your pace