The World Of AI
An explorable map of artificial intelligence. Start at the whole field and travel inward — machine learning, neural networks, transformers — until you reach the concept you came for.
Artificial Intelligence
The whole field: getting machines to do things that used to require a person.
Machine Learning
Finding patterns in data instead of being told the rules.
Supervised Learning
Learning from labelled examples — still the most common kind by far.
Linear Regression
Fitting a straight line through points to predict a number.
Logistic Regression
A straight boundary between two classes. The baseline everything must beat.
Explore Logistic Regression in a labDecision Trees
A flowchart the machine writes itself, one question at a time.
Explore Decision Trees in a labRandom Forests
Many trees on different samples, averaged. Still wins on tabular data.
Explore Random Forests in a labGradient Boosting
Trees built in sequence, each correcting the last one’s mistakes.
Support Vector Machines
Finding the boundary with the widest possible margin around it.
k-Nearest Neighbours
No training at all — ask the closest examples and take a vote.
Explore k-Nearest Neighbours in a labEnsemble Methods
Combining several weak models into one strong one. Still the tabular-data champion.
Class Imbalance
When 99% of your data is one class, and accuracy stops meaning anything.
Feature Importance
Which inputs the model actually leaned on. The first question anyone asks.
One-Hot Encoding
Turning categories into columns of zeros and ones, because models only eat numbers.
Naive Bayes
Probability with a deliberately wrong assumption that works anyway.
Unsupervised Learning
Finding structure when nobody has labelled anything.
Clustering
Grouping things that resemble each other, without being told the groups.
Explore Clustering in a labk-Means
The clustering everyone starts with. Round groups, chosen k, different answer each run.
Explore k-Means in a labDBSCAN
Clustering by density — finds odd shapes and calls the leftovers noise.
Principal Component Analysis
Squashing many columns into the few directions carrying most variation.
t-SNE and UMAP
Flattening high-dimensional data onto a page you can actually look at.
Hierarchical Clustering
Building a tree of nested groups instead of picking k up front.
Association Rules
People who bought this also bought that. Older than machine learning, still useful.
Silhouette Score
A number for how cleanly separated your clusters are. Imperfect, but better than eyeballing.
Anomaly Detection
Finding the rare thing without being told what rare looks like.
Explore Anomaly Detection in a labReinforcement Learning
Learning by doing, with a reward instead of an answer key.
Reward Function
The number you are asking a system to maximise. Getting it wrong is the classic failure.
Explore Reward Function in a labPolicy
The agent’s rule for what to do in any given situation.
Q-Learning
Learning the value of each action in each state, by trial and error.
Exploration vs Exploitation
Trying something that might be worse, in case it turns out better.
Markov Decision Process
The formal frame behind RL: states, actions, rewards, and what comes next.
Reward Shaping
Adding intermediate rewards so the agent learns something before the heat death of the universe.
Multi-Armed Bandit
The simplest exploration problem. Which slot machine, given limited pulls?
RLHF
Tuning a model on human comparisons rather than a written objective.
How Models Train
The machinery underneath every kind of learning above.
Gradient Descent
Roll downhill on a surface the model cannot see, one small step at a time.
Explore Gradient Descent in a labLearning Rate
How big each step is. The first thing to tune, and the one that decides everything.
Explore Learning Rate in a labLoss Function
The single number that says how wrong the model currently is.
Epochs
One full pass through the training data. Too few and it has not learned; too many and it memorises.
Batch Size
How many examples are looked at before each adjustment. Smaller means noisier steps.
Overfitting
Memorising the answers instead of learning the rule.
Explore Overfitting in a labUnderfitting
Too simple to capture the pattern. Wrong even on the data it was given.
Explore Underfitting in a labRegularisation
Deliberately handicapping a model so it cannot memorise.
Cross-Validation
Testing on several different held-out slices, so one lucky split cannot fool you.
Feature Engineering
Building better inputs. Usually worth more than a better algorithm.
Data Leakage
When a column quietly contains the answer. Scores look wonderful until launch.
Optimisers
The rules for how each step is taken. SGD, momentum, Adam and their descendants.
Explore Optimisers in a labMomentum
Carrying speed between steps, so the model rolls through shallow dips instead of stopping in them.
Explore Momentum in a labAdam
The default optimiser for a decade. Adapts the step size per parameter.
Explore Adam in a labEarly Stopping
Halting the moment held-out performance turns. The cheapest cure for overfitting.
Explore Early Stopping in a labWeight Decay
Gently pulling weights toward zero so none of them dominate.
Data Augmentation
Making more training data by flipping, cropping and distorting what you already have.
Hyperparameters
The settings you choose rather than the ones the model learns.
Measuring Models
Deciding whether the thing actually works, before your users do.
Accuracy
The share it got right. Close to useless when one class is rare.
Explore Accuracy in a labPrecision
When it raises a flag, how often is it right?
Explore Precision in a labRecall
Of the real cases out there, how many did it catch?
Explore Recall in a labF1 Score
One number balancing precision against recall, when you must have just one.
ROC and AUC
How well a model separates two classes across every possible threshold.
Confusion Matrix
The four outcomes: caught, missed, false alarm, correctly cleared.
Explore Confusion Matrix in a labThresholds
A model outputs a number. A person decides where the line falls — and who it hits.
Explore Thresholds in a labCalibration
Whether a model that says 70% is right about 70% of the time. Rarely checked, often wrong.
Baseline
The dumbest thing that could work. If you have not beaten it, you have nothing.
Holdout Set
Data locked away and touched exactly once, at the very end.
Train/Test Split
Holding data back and not touching it. The discipline everything else rests on.
Explore Train/Test Split in a labNeural Networks
Layers of very simple units that, stacked deep enough, learn almost anything.
Perceptron
The 1958 ancestor of every neural network. One neuron, one line.
Weights and Biases
The numbers that get adjusted during training. Everything a model knows lives here.
Layers
Stacks of units. Each one works on the previous layer’s output, not on the raw input.
Activation Functions
The small nonlinear step that lets a network learn curves instead of straight lines.
Backpropagation
Sending the error backwards through the network to work out who to blame.
Dropout
Switching off random units during training so the network cannot rely on any one of them.
Batch Normalisation
Keeping the numbers flowing through a deep network in a sane range.
Softmax
Turning a row of scores into probabilities that add to one. The last step of most classifiers.
Residual Connections
Letting a layer be skipped, so the signal survives a hundred layers of depth.
Embedding Layer
The lookup table that turns a token id into a vector of meaning.
Explore Embedding Layer in a labVanishing Gradients
Why deep networks would not train before the 2010s — the signal faded before it arrived.
Deep Learning
Neural networks with many layers — the reason the last fifteen years happened.
Computer Vision
Getting machines to work with images and video.
Convolution
Sliding nine numbers across an image, which is how machines learned to see.
Explore Convolution in a labCNNs
Networks built from convolutions. Vision, and anything with local structure.
Explore CNNs in a labPooling
Shrinking the picture between layers, keeping what matters.
ResNet
Letting layers be skipped, which is how networks got genuinely deep.
Object Detection
Not just what is in the picture, but where — and how many.
Vision Transformers
Treating an image as a sequence of patches, and using attention instead of convolution.
Optical Character Recognition
Reading text out of pictures. One of the oldest jobs in the field.
Transfer Learning
Starting from a model trained on millions of images and adapting it to your few thousand.
Image Segmentation
Labelling every pixel rather than the picture as a whole.
Sequence Models
Architectures for data with an order: text, audio, time series.
RNNs
Reading one step at a time, carrying a memory forward. Slow, and forgetful.
LSTMs
RNNs with gates that decide what to keep and what to drop.
GRUs
A lighter LSTM with fewer gates. Often just as good, and faster.
Speech Recognition
Turning sound into text. Solved well enough that it disappeared into products.
Time Series
Forecasting what comes next when the order of the data is the point.
Generative Models
Models that produce something new rather than labelling something existing.
Diffusion
Image generation by removing noise, forty times over.
Explore Diffusion in a labGANs
Two networks competing: one forging, one detecting. The previous generation of image models.
Autoencoders
Squeezing data through a narrow middle and rebuilding it, learning what matters on the way.
Text-to-Image
Words in, picture out. The application that made generative AI visible to everyone.
Explore Text-to-Image in a labGuidance Scale
How hard the model is pushed toward your prompt. Too high and colours burn out.
Explore Guidance Scale in a labSeed
The exact static a generation starts from. Same seed, same prompt, same picture, forever.
Explore Seed in a labLatent Space
The compressed space a generative model works in. Meaning as coordinates.
Explore Latent Space in a labTransformers
The architecture behind every modern language model. Attention, stacked deep.
Tokens
Text chopped into pieces before a model sees it. Not words — something stranger.
Explore Tokens in a labEmbeddings
Meaning represented as coordinates. Similar words end up as neighbours.
Explore Embeddings in a labAttention
The mechanism that decides which earlier words matter right now.
Explore Attention in a labMulti-Head Attention
Running attention many times in parallel, each head tracking something different.
Explore Multi-Head Attention in a labPositional Encoding
Telling a model the order of the words, since attention sees them all at once.
Context Window
How much a model can hold at once. Outside it, nothing exists.
Explore Context Window in a labPre-training
The enormous first pass: predict the next token, across most of the written internet.
Fine-tuning
Nudging a trained model toward your data, your tone, your task.
LoRA
Fine-tuning by training a small extra piece instead of the whole model.
Quantisation
Storing weights at lower precision so a model fits on smaller hardware.
Distillation
Training a small model to imitate a large one. How fast models get made.
Temperature
How adventurous the next-token choice is. Low is predictable, high is loose.
Explore Temperature in a labSampling
Choosing the next token from a distribution rather than always taking the top one.
Explore Sampling in a labPerplexity
How surprised a model is by text. The oldest measure of language modelling.
Scaling Laws
The observation that more data and compute buy predictable improvement.
Mixture of Experts
Only waking part of the model for each token, so a huge model stays affordable.
Decoder-Only
The architecture behind chat models: read left to right, predict the next token.
Encoder-Decoder
Read the whole input, then write the output. Translation and summarisation live here.
KV Cache
Remembering earlier attention work so each new token does not redo everything.
Top-p Sampling
Choosing from the smallest set of tokens that covers most of the probability.
Explore Top-p Sampling in a labInstruction Tuning
The step that turned a text predictor into something that follows instructions.
Emergent Abilities
Skills that appear abruptly at scale rather than improving gradually. Still debated.
Small Language Models
Compact models you can run yourself. Cheaper, faster, private, often good enough.
Applied AI
What happens when these systems meet real users, real documents and real money.
Building With Models
The patterns you reach for when putting a model into a product.
Retrieval
Giving a model your documents to answer from, instead of what it half-remembers.
Explore Retrieval in a labRAG
Retrieval-augmented generation: search, then answer from what was found.
Explore RAG in a labChunking
How you cut documents up. Split on structure, not on character count.
Explore Chunking in a labVector Database
Storage that finds things by meaning rather than by exact words.
Explore Vector Database in a labReranking
A second pass that re-reads the shortlist properly and reorders it.
Explore Reranking in a labHybrid Search
Keyword and meaning-based search together, because each misses what the other catches.
Prompt Engineering
Writing the instruction well. Real, useful, and much smaller than the hype suggested.
System Prompt
The standing instruction a model carries into every conversation.
Explore System Prompt in a labFew-Shot Prompting
Showing two or three worked examples instead of explaining the rule.
Chain of Thought
Asking for the working, which measurably improves the answer.
Structured Output
Forcing the reply into JSON or a schema so software can rely on it.
Semantic Search
Finding documents by meaning rather than by matching words.
Explore Semantic Search in a labPrompt Templates
Reusable, versioned prompts. The moment prompting becomes engineering.
Context Stuffing
Throwing everything into the window and hoping. Usually worse than retrieving well.
Output Validation
Checking the reply is the shape you asked for, before anything downstream trusts it.
Function Calling
Letting a model trigger real code — the doorway between chat and action.
Agents
Systems that plan and act rather than answer.
Tool Use
Giving a model buttons it can press. Where capability and risk both jump.
Planning
Breaking a goal into steps before starting. Still the weakest part of most agents.
Agent Memory
What an agent carries between steps and between sessions.
Multi-Agent Systems
Several agents with different jobs, talking to each other. Often more fragile than one.
ReAct
Reason, then act, then observe, then repeat. The pattern most agents are built on.
Task Decomposition
Splitting a goal into steps small enough to actually succeed.
Sandboxing
Giving an agent a room it cannot break out of before you give it tools.
Explore Sandboxing in a labHuman in the Loop
A confirmation step before anything that sends, spends or deletes.
Explore Human in the Loop in a labRisk and Safety
The ways these systems fail, and what actually prevents it.
Hallucination
When a model confidently invents a citation, a case, a clause.
Explore Hallucination in a labPrompt Injection
Instructions hidden in a document the model was asked to read.
Explore Prompt Injection in a labJailbreaking
Talking a model out of its own guardrails. A permanent arms race.
Model Bias
Patterns absorbed from training text, including ones nobody would defend.
Explore Model Bias in a labReward Hacking
An agent satisfying your rule exactly while missing your intention entirely.
Explore Reward Hacking in a labAlignment
Getting a system to pursue what you meant rather than what you wrote.
Explore Alignment in a labGuardrails
Checks around the model rather than inside it. Mitigation, never protection.
Red Teaming
Attacking your own system on purpose, before somebody else does.
Explore Red Teaming in a labExplainability
Being able to say why a decision was made. Increasingly a legal requirement, not a nicety.
Model Cards
A short honest document: what it was trained on, what it is for, where it fails.
Differential Privacy
Adding noise so individual records cannot be recovered from a trained model.
Data Privacy
What may be sent to a third-party model, and what must never leave your building.
Running It
The unglamorous part that decides whether any of it survives contact with users.
Evaluation
Deciding whether the thing works, before your users do it for you.
Explore Evaluation in a labEval Sets
A fixed set of cases you re-run on every change. The regression test for models.
Observability
Being able to see what your system actually did, after it did it.
Cost per Token
The unit economics. Easy to ignore in a demo and impossible to ignore at scale.
Explore Cost per Token in a labLatency
How long the user waits. Often matters more to them than a small quality gain.
A/B Testing
Shipping to half your users and measuring. The only opinion that settles arguments.
Fallback Models
What runs when the good model is down, slow, or too expensive for this request.
Token Budgeting
Deciding what earns a place in the context window, because not everything can.
Explore Token Budgeting in a labRate Limits
The ceiling on how fast you may call. Discovered, always, in production.
Caching
Not paying twice for the same answer.