LATENT
HOW AI WORKS - FRAME BOOK
VOL. 01·SEP 2026

TERMSACROSSTHE BOOK.

  1. 01ATTENTION11
  2. 02REPRESENTATION19
  3. 03TRANSFORMATION20
  4. 04GENERATION27

A BACK COVER THAT DOUBLES AS AN INDEX: 77 TERMS, FOUR THEMES, EVERY PAGE ONE CLICK AWAY.

  1. 001
  2. 002
  3. 003
  4. 004
  5. 005
  6. 006
  7. 007
  8. 008
  9. 009
  10. 010
  11. 011
  12. 012
  13. 013
  14. 014
  15. 015
  16. 016
  17. 017
  18. 018
  19. 019
  20. 020
  21. 021
  22. 022
  23. 023
  24. 024

SYNTHESIS GLYPH

24 PAGES ON ONE RING. EACH CHORD JOINS TWO CHAPTERS THAT SHARE AN IDEA; PULSES RUN IN READING ORDER.

THE LANGUAGE-MODEL LOOP IN ONE LINE:

pθ(xt+1∣x≤t)p_\theta(x_{t+1} \mid x_{\le t})
=softmax⁡(WU ht(L))= \operatorname{softmax}\big(W_U\, h^{(L)}_t\big)

pages = 24terms = 77chords = 31loop = 8 s

HOVER OR FOCUS A PAGE NUMBERTO SEE ITS LINKS.

TAP A PAGE NUMBER TO OPEN IT.

softmax⁡ ⁣(QK⊤dk)V\operatorname{softmax}\!\left(\dfrac{QK^{\top}}{\sqrt{d_k}}\right)V

TOKENS SHARE CONTEXT BY WEIGHTED LOOKUP.

cos⁡(xi,xj)=xi⋅xj∥xi∥ ∥xj∥\cos(x_i, x_j) = \dfrac{x_i \cdot x_j}{\lVert x_i \rVert\, \lVert x_j \rVert}

MEANING BECOMES DISTANCE AND DIRECTION.

θ←θ−η ∇θL(θ)\theta \leftarrow \theta - \eta\, \nabla_{\theta} L(\theta)

LAYERS COMPUTE; GRADIENTS TEACH THEM.

xt+1∼softmax⁡(z/T)x_{t+1} \sim \operatorname{softmax}(z / T)

SAMPLE A STEP, APPEND, REPEAT.

Glossary

Terms Across the Book

77 KEY TERMS FROM PAGES 002 – 023, GROUPED BY THE BOOK'S FOUR THEMES. EACH ONE HAS A SHORT DEFINITION AND LINKS BACK TO THE PAGES WHERE THE IDEA IS DRAWN.

The book's four themes as a stack of layers — 04 generation: 27 terms, 03 transformation: 20 terms, 01 attention: 11 terms, 02 representation: 19 terms.

HOW TO READ

Embeddingx∈Rdx \in \mathbb{R}^{d}
A learned vector that stands for a token or an image patch.
Pages: 004017019010

THEMES

01 ATTENTION11
02 REPRESENTATION19
03 TRANSFORMATION20
04 GENERATION27
TOTAL77

GLOSSARY77 TERMS · 002 — 023

01ATTENTIONHOW TOKENS SHARE INFORMATION AND CONTEXT.11 TERMS

Attention
Each token builds its new vector as a weighted mix of other tokens' values.
Pages: 005001007
Query, key, valueQ, K, VQ,\ K,\ V
Three vectors per token: what it looks for, what it offers, and what it passes on.
Pages: 005014020
Attention head
One attention computation; several run in parallel, each learning its own pattern.
Pages: 005001
Causal mask
Hides later tokens, so each position can only attend to itself and earlier ones.
Pages: 014011
Positional encoding
A signal that tells the model where each token sits in the sequence.
Pages: 005019
Context window
The maximum number of tokens a model can take in at once.
Pages: 014
KV cache
Keys and values of earlier tokens, saved so each new token needs less computation.
Pages: 014
Grouped-query attention (GQA)
Several query heads share one set of keys and values, so the KV cache needs less memory.
Pages: 014
Prefill and decode
The two phases of inference: read the whole prompt in parallel, then write one token per step.
Pages: 014
Cross-attention
Attention where one sequence queries another, e.g. image features reading a text prompt.
Pages: 020
System prompt
Instructions placed before the conversation that set the model's role and rules.
Pages: 023014017

02REPRESENTATIONHOW TEXT, IMAGES AND MEANING BECOME VECTORS.19 TERMS

Token
A unit of text the model reads: a word, part of a word, or a symbol.
Pages: 004010
Tokenizer
Fixed rules that split text into tokens and give each one an ID.
Pages: 004010
Vocabulary∣V∣|V|
Every token a model knows, each with an integer ID (often 50,000 or more).
Pages: 004006011
Vector
An ordered list of numbers; also a point or a direction in space.
Pages: 002004
Embeddingx∈Rdx \in \mathbb{R}^{d}
A learned vector that stands for a token or an image patch.
Pages: 004017019010
Latent space
The space of a model's internal vectors, where related things lie close together.
Pages: 022001009020
Cosine similaritycos⁡(xi,xj)\cos(x_i, x_j)
How closely two vectors point the same way, from −1 to 1.
Pages: 017004
Retrieval-augmented generation (RAG)
Search a document store, often by embedding similarity, and add the best passages to the prompt.
Pages: 017021
Vector index
A store of embeddings built to find the ones nearest to a query, fast.
Pages: 017
Patch
A small square of an image, flattened and embedded like a token.
Pages: 019
Vision Transformer (ViT)
A transformer that reads an image as a sequence of patch embeddings, just as it reads text.
Pages: 019
CLIP
An image encoder and a text encoder trained so each picture lands near its caption in one shared space.
Pages: 019020
Residual stream
Each token's running vector; every layer reads it and adds its output back.
Pages: 007022
Featuredid_i
A direction in activation space that tracks a concept people can name.
Pages: 022
Polysemantic neuron
A single neuron that responds to several unrelated concepts.
Pages: 022
Superposition
Hypothesis that models store more features than dimensions as nearly orthogonal directions.
Pages: 022
Sparse autoencoder (SAE)
Rewrites an activation as a sum of a few features from a large learned dictionary.
Pages: 022
Probe
A small classifier trained on activations to test what information they contain.
Pages: 022
Steeringx+α dix + \alpha\, d_i
Adding a feature direction to activations to change what the model does.
Pages: 022

03TRANSFORMATIONHOW LAYERS COMPUTE AND HOW MODELS LEARN.20 TERMS

Neuron
A weighted sum of inputs plus a bias, passed through an activation function.
Pages: 006
Parametersθ\theta
The weights and biases a model learns; billions in a large model.
Pages: 006008012
Activation functionσ\sigma
A nonlinear function, such as ReLU, applied after a weighted sum.
Pages: 006
Transformer layer
Attention plus a feed-forward network, joined by residual connections and normalization.
Pages: 007010
Feed-forward network (MLP)
A small two-layer network applied to each token separately inside every layer.
Pages: 007
LossL(θ)L(\theta)
A number that measures how wrong the model is; training lowers it.
Pages: 008011012
Cross-entropy−log⁡p-\log p
The next-token loss: minus the log of the probability given to the true token.
Pages: 011021
Gradient∇L\nabla L
The direction in which the loss rises fastest: one slope per parameter.
Pages: 008007
Backpropagation
Computing every gradient layer by layer, backward, with the chain rule.
Pages: 007
Gradient descent
Repeatedly moving the parameters a small step against the gradient.
Pages: 008011
Learning rateη\eta
The size of each gradient-descent step.
Pages: 008
Pre-training
Training on a huge text corpus by predicting the next token.
Pages: 011012015
Scaling laws
Loss falls predictably, roughly as a power law, as parameters, data and compute grow.
Pages: 012
Compute-optimal training
For a fixed compute budget, grow parameters and data together: about 20 training tokens per parameter.
Pages: 012
Emergent abilities
Skills that seem to switch on suddenly at scale; partly an effect of all-or-nothing scoring.
Pages: 012
Fine-tuning
Further training of a pre-trained model on a smaller, targeted dataset.
Pages: 015
RLHF
Reinforcement learning from human feedback: training toward answers people prefer.
Pages: 015
Reward model
A model trained on people's choices between two answers, used to give any answer a score.
Pages: 015
KL penalty
A cost for drifting from the reference model, so RLHF gains reward without forgetting how to write.
Pages: 015
Alignment
Shaping a model's behavior to match human intentions and values.
Pages: 015023

04GENERATIONHOW OUTPUTS ARE PRODUCED, USED AND CHECKED.27 TERMS

Logitziz_i
The raw score a model gives each vocabulary token before softmax.
Pages: 006013
Softmaxezi/∑jezje^{z_i} / \textstyle\sum_j e^{z_j}
Turns a list of scores into positive probabilities that sum to 1.
Pages: 006013005
Next-token prediction
The core task: assign a probability to every possible next token.
Pages: 011006010
Autoregressive generation
Generate one token, append it to the input, repeat.
Pages: 010014013
Sampling
Picking the next token at random, weighted by its probability.
Pages: 013021016
Greedy decoding
Always picking the single most likely next token, so the same prompt gives the same text.
Pages: 013006
TemperatureTT
Divides the logits before softmax: low is focused, high is more varied.
Pages: 013021
Top-k
Keep only the k best candidates: the k likeliest tokens when sampling, or the k closest chunks in a search.
Pages: 013017
Top-p (nucleus) sampling
Sampling only from the smallest set of tokens whose probabilities add up to p.
Pages: 013
Chain of thought
Writing out intermediate reasoning steps as tokens before the answer.
Pages: 016
Test-time compute
Spending more computation while answering, by writing longer reasoning or sampling more attempts.
Pages: 016
Self-consistency
Sampling several chains of reasoning and keeping the answer most of them reach.
Pages: 016
Tool use
The model writes a structured call; software runs it and returns the result.
Pages: 018
Model Context Protocol (MCP)
An open standard for plugging tools and data sources into AI applications.
Pages: 018
Agent
A loop in which a model plans, acts with tools, observes the results, and repeats.
Pages: 018
Diffusion model
Generates an image by removing noise step by step, starting from pure noise.
Pages: 009020
Latent diffusion
Diffusion run in a compressed latent space instead of on pixels.
Pages: 020009
Classifier-free guidance
Pushes each denoising step toward the prompt by comparing predictions with and without it.
Pages: 020
Hallucination
Fluent, confident output that is false or not supported by any source.
Pages: 021017
Knowledge cutoff
The date the training data ends; the model has seen nothing that happened after it.
Pages: 021
Calibration
How well a model's confidence matches how often it is actually right.
Pages: 021
Benchmark
A fixed, held-out set of test cases with a scoring rule.
Pages: 023
pass@k
The chance that at least one of k sampled solutions passes the tests.
Pages: 023
Contamination
Test items leaking into training data, which inflates scores.
Pages: 023
LLM-as-a-judge
Using a model to compare or grade answers when there is no single correct one.
Pages: 023
Red-teaming
Deliberately attacking a model to find harmful or failing behavior.
Pages: 023
Safety classifier
A separate model that screens inputs or outputs and blocks unsafe ones.
Pages: 023

LATENT

How AI Works

A MOTION DESIGN STYLE-FRAME BOOK

COLOPHON

EDITION
VOL. 01 · SEP 2026
PAGES
001 — 024
CREDIT
Compiled & developed by Vu Dinh
TYPE
IBM Plex Mono · Inter Tight
BACK TO THE COVER · 001