Learn how AI really works

Paste any text and watch how a language model chops it into tokens — the subword chunks LLMs actually read. Words, spaces and punctuation, split live.

• Live tokenisation
• Token & character counts
• Try your own text

See how AI turns words into vectors — explore a semantic map, find nearest neighbours, and watch king − man + woman ≈ queen as real vector maths.

• 2-D semantic map
• Nearest neighbours
• Analogy arithmetic

Explore how temperature, top-k and top-p control the randomness of language models. Adjust sliders and watch the probability distribution change live.

• Real softmax maths
• Interactive sliders
• Sample from the distribution

See how transformer attention works — type a sentence and watch each token decide what to attend to via scaled dot-product attention.

• Interactive heatmap
• Click to highlight connections
• Real attention maths

How do transformers know word order? Explore the sinusoidal positional-encoding heatmap and watch each dimension trace a sine wave across positions.

• Sinusoidal PE heatmap
• Dimension wave plot
• Real transformer maths

See how a model decodes text — compare greedy decoding against beam search and watch the search tree keep and prune candidate sequences by log-probability.

• Adjustable beam width
• Live search tree
• Real log-prob scoring

Why long documents get sliced up for RAG — set a chunk size and overlap, and see how a document splits to fit inside a model's context budget.

• Adjustable size & overlap
• Visual chunks
• Context budget bar

Watch Retrieval-Augmented Generation in action — type a question, see documents retrieved from a knowledge base, and read the grounded answer with citations.

• Retrieve → Augment → Generate
• Live step-by-step
• Illustrative bigspark corpus

Generate text the pre-LLM way — build a word-level Markov model from a corpus and sample next words by frequency. Predicting words, with no understanding.

• Order 1 or 2
• Paste your own text
• Transition tables

The decoder's defining trick — toggle the causal mask and watch each token lose sight of the future, so it can only attend to itself and earlier tokens.

• Causal vs full attention
• Masked softmax
• Interactive heatmap

How an LLM actually writes — generate one token at a time, feed it back in, and repeat. Watch the sequence grow and the next-token odds shift each step.

• One token at a time
• Live next-token odds
• Temperature control

How a decoder looks back at the encoder — the mechanism behind translation and summarisation. Click a target word to see which source words it attends to.

• Encoder → decoder
• Word alignment
• Seq2seq intuition

Why generation speeds up after the first token — cache keys and values instead of recomputing the whole prefix, turning quadratic work into linear.

• Cached vs recomputed
• O(n²) vs O(n)
• Step-by-step cost

Place two classes of points, hit train, and watch a from-scratch neural network learn the boundary between them — live, in your browser.

• Real backprop, no libraries
• Live decision boundary
• Preset datasets

Watch unsupervised learning find structure — drop points, pick k, and step through Lloyd's algorithm as centroids move and clusters settle.

• Click to add points
• Step or run
• Live inertia readout

Drag points and watch the least-squares best-fit line update live, with residuals, the equation and R² — the simplest model there is, made tangible.

• Drag & add points
• Live best-fit line
• Equation & R²

Classify by proximity — place labelled points, move your cursor, and see the decision boundary and the k nearest neighbours that decide each vote.

• Live decision regions
• Adjustable k
• Real Euclidean distance

How trees split data — slide a threshold across a scatter plot and watch Gini impurity and information gain change, then let it find the best split.

• Interactive splits
• Gini & information gain
• Find best split

The building block of CNNs — apply edge-detect, blur, sharpen and emboss kernels to an image, or type your own 3×3 kernel and see the result live.

• Preset & custom kernels
• Live filtered output
• Real convolution maths

Watch an agent learn by trial and error — train tabular Q-learning in a grid world and see state values and the greedy policy emerge, then follow the path.

• Train & watch it learn
• Value & policy map
• Tune α, γ, ε

See how a system prompt, untrusted content and a user question stack into one context — then watch an illustrative guardrail catch an injection attempt.

• Colour-coded context
• Try an injection
• Guardrail in action

One model, two groups — adjust thresholds and see selection rates, the four-fifths rule and the tension between competing fairness definitions.

• Per-group thresholds
• Disparate impact
• Equalised-odds view

Move a fraud model's decision threshold across overlapping score distributions and watch precision, recall, F1 and false alarms shift in real time.

• Precision vs recall
• Live confusion matrix
• The fincrime trade-off

Why a 90%-accurate test can still be wrong — set prevalence, sensitivity and specificity and see why false positives can dwarf true ones.

• Prior → posterior
• 1,000-person view
• Base-rate fallacy

Estimate π by throwing random darts at a circle — a hands-on look at Monte Carlo sampling and how error shrinks as the number of samples grows.

• Throw random darts
• Live π estimate
• Convergence plot

Watch optimisation roll downhill across a 2-D loss surface — tune the learning rate to see smooth convergence, oscillation, or outright divergence.

• 2-D loss surface
• Tune learning rate
• Real analytic gradient

A quick, interactive self-assessment across five dimensions — get an instant radar-chart snapshot of your organisation's AI maturity.

• 5 dimensions
• Instant radar chart
• Scored in your browser

Estimate what AI adoption could save your team — set your assumptions and see hours, full-time equivalents and cost saved per year, live.

• Live estimate
• Hours & £ saved
• Illustrative

Fancy a break? A few small browser games our engineers built for fun live over in the Playground — pure vanilla JavaScript, no downloads.

• Snake
• 2048
• Memory match

A curated shelf of talks and tutorials on AI, data engineering and building software well. Filter by topic or by learning pathway.

• Curated video picks
• Filter by pathway
• Updated regularly