25  Chapter 25: The Grammar of AI

How vectors, matrices, embeddings, attention, and optimization become intelligent systems

25.1 Opening story: from numbers to language, vision, and decisions

At the beginning of this book, we learned a simple but powerful idea:

To compute with the world, we first turn the world into numbers.

A house became a list of features.
A picture became a matrix of pixel values.
A sentence became a vector of word counts.
A dataset became a cloud of points.
A matrix became a machine that moves, mixes, compresses, and transforms information.

Modern artificial intelligence is built from the same story, only at a much larger scale.

An AI system may look mysterious from the outside. It can recognize an image, recommend a movie, translate a sentence, summarize a paper, generate code, answer a question, or write a paragraph. But inside the machine, the basic grammar is familiar:

  • objects become vectors;
  • collections of objects become matrices;
  • similarity becomes dot product or cosine similarity;
  • memory becomes stored vectors;
  • attention becomes a matrix of comparisons;
  • a neural network layer becomes a matrix machine followed by a nonlinear gate;
  • learning becomes optimization over many matrix entries;
  • compression becomes low-rank structure;
  • generalization depends on geometry in high-dimensional space.

This chapter is the closing chapter of the book. It is not a full textbook on deep learning. Instead, it is a map showing how the central ideas of linear algebra appear again and again inside modern AI.

ImportantCentral message

Linear algebra is the grammar of AI.

It gives AI systems a language for representing data, comparing meaning, transforming information, compressing structure, and learning from examples.

25.2 Learning goals

By the end of this chapter, you should be able to:

  1. Explain why AI systems turn text, images, users, items, and actions into vectors.
  2. Interpret embeddings as learned coordinates for meaning.
  3. Use dot products and cosine similarity to compare embeddings.
  4. Explain vector search as nearest-neighbor search in embedding space.
  5. Interpret a neural network layer as \(h = \sigma(Wx+b)\).
  6. Explain attention as a matrix of dot-product comparisons.
  7. Connect SVD, PCA, low-rank structure, and embeddings.
  8. Explain training as optimization over matrix parameters.
  9. Understand why high-dimensional geometry is both powerful and dangerous.
  10. Describe how the ideas in this book combine into the mathematical grammar of AI.

25.3 25.1 The representation principle

AI systems compute with numbers. Before an AI model can use an object, the object must be represented numerically.

Real-world object Numerical representation Linear algebra object
Image pixel intensities matrix or tensor
Sentence token vectors sequence of vectors
Document embedding vector
User interaction history sparse vector
Movie/product feature or latent-factor vector vector
Dataset rows of examples matrix
Network/graph adjacency table matrix
Model parameters weights matrices and vectors

This gives the first principle of AI.

ImportantRepresentation principle

An AI system can only compute with an object after the object has been represented as numbers.

Most modern AI begins by turning objects into vectors, matrices, or tensors.

A representation is not just bookkeeping. It decides what the model can see. Two different representations of the same object may lead to very different behavior.

For example, the word “bank” can mean a financial institution or the side of a river. A simple word-count vector may not distinguish the two meanings well. A contextual embedding from a modern language model may place the word in different regions depending on the surrounding sentence.

25.4 25.2 Embeddings: coordinates for meaning

An embedding is a vector representation of an object. The object could be a word, sentence, image, user, item, protein, graph node, or mathematical expression.

For example,

\[ \text{``linear algebra''} \longmapsto \begin{bmatrix} 0.18 \\ -0.42 \\ 1.07 \\ \vdots \\ 0.31 \end{bmatrix} \in \mathbb{R}^{d}. \]

The individual coordinates of an embedding are usually not chosen by hand. They are learned from data. The goal is not that coordinate 17 has a simple human name. The goal is that the geometry of the space becomes useful.

NoteEmbedding principle

An embedding turns an object into a vector so that useful relationships become geometric relationships.

In a good embedding space:

  • similar documents point in similar directions;
  • related images lie near each other;
  • users with similar preferences have nearby vectors;
  • items with similar audiences have nearby vectors;
  • questions and relevant answers have high similarity.

This is why the earlier chapters on distance, angle, projection, SVD, and PCA are not separate from AI. They are the mathematical tools for studying embedding spaces.

25.5 25.3 Similarity: the dot product returns

Given two embedding vectors \(u\) and \(v\), one common measure of similarity is cosine similarity:

\[ \operatorname{cosine}(u,v) = \frac{u\cdot v}{\|u\|\|v\|}. \]

This quantity measures whether two vectors point in similar directions.

  • If \(\operatorname{cosine}(u,v) \approx 1\), the vectors point in nearly the same direction.
  • If \(\operatorname{cosine}(u,v) \approx 0\), the vectors are nearly orthogonal.
  • If \(\operatorname{cosine}(u,v) \approx -1\), the vectors point in opposite directions.

This simple formula appears throughout AI:

AI task What is compared? Linear algebra operation
Search query vs documents cosine similarity
Recommendation user vs item dot product
Classification point vs class direction dot product score
Attention token vs token dot product matrix
Clustering point vs centroid distance or cosine
Retrieval-augmented generation question vs stored chunks nearest-neighbor search
ImportantDot-product principle

The dot product is a basic comparison engine in AI.

It turns two vectors into a score.

25.5.1 Example 25.1: a tiny semantic search system

Suppose we store three document embeddings:

\[ d_1 = \begin{bmatrix} 1 \\ 0 \end{bmatrix}, \qquad d_2 = \begin{bmatrix} 0 \\ 1 \end{bmatrix}, \qquad d_3 = \begin{bmatrix} 0.8 \\ 0.6 \end{bmatrix}. \]

Suppose a query has embedding

\[ q = \begin{bmatrix} 0.9 \\ 0.4 \end{bmatrix}. \]

The system ranks documents by comparing \(q\) with each \(d_i\) using cosine similarity.

Show solution

First compute the norms:

\[ \|q\|=\sqrt{0.9^2+0.4^2}=\sqrt{0.97}. \]

The first document has cosine similarity

\[ \frac{q\cdot d_1}{\|q\|\|d_1\|} = \frac{0.9}{\sqrt{0.97}}. \]

The second document has cosine similarity

\[ \frac{q\cdot d_2}{\|q\|\|d_2\|} = \frac{0.4}{\sqrt{0.97}}. \]

The third document has

\[ q\cdot d_3 = 0.9(0.8)+0.4(0.6)=0.96. \]

Since \(\|d_3\|=1\), the cosine similarity is

\[ \frac{0.96}{\sqrt{0.97}}. \]

Thus \(d_3\) is most similar to the query.

25.6 25.4 Vector search: memory as geometry

A modern AI assistant often uses stored documents. The system may work like this:

  1. Split documents into smaller chunks.
  2. Convert each chunk into an embedding vector.
  3. Store all embedding vectors in a vector database.
  4. Convert the user question into an embedding vector.
  5. Retrieve nearby chunks.
  6. Use those chunks as context for a language model.

This is called retrieval-augmented generation, often abbreviated as RAG.

The key linear algebra step is vector search:

\[ \text{find documents } d_i \text{ with large } \operatorname{cosine}(q,d_i). \]

This is just nearest-neighbor search in a high-dimensional space.

NoteRAG as linear algebra

A retrieval system uses geometry to decide which pieces of text are relevant to a question.

The question and the stored text chunks are vectors.

25.7 25.5 Matrices as batches of embeddings

A single embedding is a vector. A collection of embeddings is a matrix.

Suppose we have \(n\) objects, each represented by a \(d\)-dimensional vector. We can place them into a matrix

\[ X = \begin{bmatrix} --- x_1^T --- \\ --- x_2^T --- \\ \vdots \\ --- x_n^T --- \end{bmatrix} \in \mathbb{R}^{n\times d}. \]

Each row is an object. Each column is a coordinate in embedding space.

This row-matrix view appears everywhere:

  • a batch of images becomes an input matrix;
  • a set of documents becomes a document-embedding matrix;
  • a sentence becomes a token-embedding matrix;
  • a user-item rating table becomes a matrix with many missing entries;
  • a dataset becomes a design matrix for learning.

Now matrix multiplication can process many objects at once.

25.8 25.6 A neural network layer is a matrix machine

A basic neural network layer has the form

\[ h = \sigma(Wx+b). \]

Here:

  • \(x\) is the input vector;
  • \(W\) is a weight matrix;
  • \(b\) is a bias vector;
  • \(Wx+b\) is an affine transformation;
  • \(\sigma\) is a nonlinear activation function;
  • \(h\) is the output vector.

Without \(\sigma\), a stack of layers would still be just one large linear transformation. The nonlinear activation is what allows neural networks to represent curved decision boundaries and complex functions.

ImportantNeural network layer

A neural network layer is a matrix machine followed by a nonlinear gate.

\[ h = \sigma(Wx+b). \]

25.8.1 Row view and feature detection

If

\[ W = \begin{bmatrix} --- w_1^T --- \\ --- w_2^T --- \\ \vdots \\ --- w_m^T --- \end{bmatrix}, \]

then

\[ Wx = \begin{bmatrix} w_1\cdot x \\ w_2\cdot x \\ \vdots \\ w_m\cdot x \end{bmatrix}. \]

Each row \(w_i\) acts like a detector. It asks: “How much does the input look like this pattern?”

This connects neural networks to cosine similarity, projection, and dot products.

25.8.2 Batch computation

If \(X\) contains many input vectors as rows, then a layer can process the whole batch by

\[ H = \sigma(XW^T + \mathbf{1}b^T). \]

This is why matrix multiplication is the computational heart of deep learning.

25.9 25.7 Training is optimization over matrices

A neural network has many parameters. These parameters are entries of matrices and vectors.

Training means choosing those numbers so that the model performs well on examples.

Given data points \((x_i,y_i)\) and a model \(f_\theta\), training often means solving

\[ \min_\theta \; \frac{1}{n}\sum_{i=1}^n L(f_\theta(x_i),y_i), \]

where \(\theta\) represents all trainable weights and biases.

This is an optimization problem.

The earlier chapter on energy landscapes now reappears: the loss function is a landscape over parameter space. Gradient descent moves through that landscape.

NoteLearning principle

Learning means adjusting matrices and vectors so that a loss function becomes small.

25.10 25.8 Attention: a matrix of comparisons

Attention is one of the most important ideas in modern language models.

The basic idea is simple:

Each token asks which other tokens are relevant to it.

Suppose a sentence has token vectors arranged in a matrix \(X\). A transformer layer forms three new matrices:

\[ Q = XW_Q, \qquad K = XW_K, \qquad V = XW_V. \]

These are called queries, keys, and values.

The attention score matrix is

\[ S = QK^T. \]

The entry \(S_{ij}\) is a dot product between the query vector of token \(i\) and the key vector of token \(j\).

After scaling and applying softmax row by row, we get attention weights:

\[ A = \operatorname{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right). \]

Then the output is

\[ Z = AV. \]

This says: each output token is a weighted combination of value vectors.

ImportantAttention as linear algebra

Attention is built from matrix multiplication, dot products, softmax normalization, and weighted averages.

\[ Z = \operatorname{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V. \]

25.10.1 Why this is powerful

Attention lets a model decide which parts of the input matter for each position.

For example, in the sentence

The student opened the notebook because she wanted to study.

A language model may use attention to connect “she” with “student” or “notebook” with “study,” depending on the task.

Mathematically, this is a learned geometry of relevance.

25.11 25.9 Low-rank structure and hidden factors

Many AI systems discover that data has hidden lower-dimensional structure.

This idea appeared throughout the book:

  • PCA finds important directions in a data cloud.
  • SVD decomposes a matrix into rank-one layers.
  • Image compression keeps the strongest singular directions.
  • Recommendation systems learn user and item factors.
  • Text models discover topic-like directions.
  • Neural networks learn hidden representations.

A low-rank approximation has the form

\[ A \approx U_k\Sigma_kV_k^T. \]

This says that a large table of data can often be explained by a smaller number of hidden patterns.

NoteHidden-factor principle

Large data often has hidden structure.

Linear algebra helps reveal it through projections, eigenvectors, SVD, and low-rank approximation.

25.12 25.10 AI as a chain of representations

An AI model often transforms data through many representation spaces.

For an image classifier:

\[ \text{pixels} \longrightarrow \text{edges} \longrightarrow \text{textures} \longrightarrow \text{parts} \longrightarrow \text{object class}. \]

For a language model:

\[ \text{tokens} \longrightarrow \text{embeddings} \longrightarrow \text{contextual vectors} \longrightarrow \text{next-token scores}. \]

For a recommender system:

\[ \text{ratings/clicks} \longrightarrow \text{user factors and item factors} \longrightarrow \text{predicted preference}. \]

Each stage changes the coordinate system. Each stage tries to make the next task easier.

This is why representation learning is so important.

25.13 25.11 High-dimensional geometry: power and risk

AI works in high-dimensional spaces. An embedding vector may have hundreds, thousands, or even more coordinates.

High dimensions are powerful because they provide room to separate many different patterns. But they are also risky because geometry behaves differently.

Important high-dimensional phenomena include:

  1. Random vectors are often nearly orthogonal.
  2. Distances can concentrate.
  3. Nearest neighbors can become less meaningful without good representation.
  4. Sparse vectors can be very far apart in Euclidean distance but meaningful under cosine similarity.
  5. Small perturbations can sometimes change model outputs.
WarningHigh-dimensional warning

In high dimensions, intuition from the plane can fail.

AI systems need good representations, normalization, and evaluation because high-dimensional geometry can be surprising.

25.14 25.12 A map of the book inside AI

The chapters of this book form a sequence of ideas that now fit together.

Book idea AI interpretation
World as numbers data representation
Vectors features and embeddings
Linear combinations mixtures and learned features
Data as points datasets in feature space
Matrix machine model layers and transformations
Geometry of matrices learned transformations
Solving systems inverse problems and fitting
Information loss compression and non-invertibility
Length and distance errors, nearest neighbors
Angles and similarity cosine search, attention scores
Projection approximation and least squares
Orthogonality stable coordinates and decompositions
Eigenvectors stable directions and ranking
Iteration Markov chains, PageRank, dynamics
Energy landscapes loss functions and optimization
SVD compression, PCA, latent factors
Image compression low-rank visual structure
PCA dimension reduction and visualization
Fourier/Haar signal and image features
Images as matrices computer vision
Text as vectors NLP and semantic search
Neural networks trainable matrix machines
Recommendation systems matrix completion and latent geometry

The subject has been linear algebra all along. The applications changed, but the grammar stayed consistent.

25.15 25.13 Worked example: a tiny attention calculation

Suppose a sequence has two token vectors. Let

\[ Q = \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}, \qquad K = \begin{bmatrix} 1 & 1 \\ 1 & 0 \end{bmatrix}, \qquad V = \begin{bmatrix} 2 & 0 \\ 0 & 3 \end{bmatrix}. \]

Compute the unnormalized attention score matrix \(S=QK^T\).

Show solution

Since

\[ K^T = \begin{bmatrix} 1 & 1 \\ 1 & 0 \end{bmatrix}, \]

we get

\[ S = QK^T = \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \begin{bmatrix} 1 & 1 \\ 1 & 0 \end{bmatrix} = \begin{bmatrix} 1 & 1 \\ 1 & 0 \end{bmatrix}. \]

The first token gives equal raw attention score to both tokens. The second token gives higher raw score to the first token than to the second.

25.16 25.14 Worked example: a tiny neural layer

Let

\[ W= \begin{bmatrix} 1 & -1 \\ 2 & 1 \end{bmatrix}, \qquad b= \begin{bmatrix} 0.5 \\ -1 \end{bmatrix}, \qquad x= \begin{bmatrix} 2 \\ 1 \end{bmatrix}. \]

Let \(\sigma(t)=\max(0,t)\) be ReLU applied coordinatewise. Compute

\[ h=\sigma(Wx+b). \]

Show solution

First compute

\[ Wx= \begin{bmatrix} 1 & -1 \\ 2 & 1 \end{bmatrix} \begin{bmatrix} 2 \\ 1 \end{bmatrix} = \begin{bmatrix} 1 \\ 5 \end{bmatrix}. \]

Then

\[ Wx+b= \begin{bmatrix} 1 \\ 5 \end{bmatrix} + \begin{bmatrix} 0.5 \\ -1 \end{bmatrix} = \begin{bmatrix} 1.5 \\ 4 \end{bmatrix}. \]

Applying ReLU gives

\[ h= \begin{bmatrix} 1.5 \\ 4 \end{bmatrix}. \]

25.17 25.15 Python: embedding search from scratch

Code
import numpy as np

# Toy embeddings for short documents
labels = np.array([
    "linear algebra and matrices",
    "dogs and cats",
    "machine learning and neural networks",
    "eigenvectors and PCA",
    "cooking pasta"
])

X = np.array([
    [1.0, 0.9, 0.0, 0.1],
    [0.0, 0.1, 1.0, 0.8],
    [0.9, 0.8, 0.1, 0.2],
    [1.0, 0.7, 0.0, 0.0],
    [0.0, 0.0, 0.2, 0.1]
])

query = np.array([1.0, 0.8, 0.0, 0.1])

def cosine_similarity_matrix(X, q):
    X_norm = X / np.linalg.norm(X, axis=1, keepdims=True)
    q_norm = q / np.linalg.norm(q)
    return X_norm @ q_norm

scores = cosine_similarity_matrix(X, query)
order = np.argsort(-scores)

for i in order:
    print(f"{scores[i]:.3f}  {labels[i]}")
0.998  linear algebra and matrices
0.995  eigenvectors and PCA
0.992  machine learning and neural networks
0.097  dogs and cats
0.035  cooking pasta

This is a tiny version of vector search. Real systems use much larger vectors and specialized algorithms, but the underlying geometry is the same.

25.18 25.16 Python: attention from scratch

Code
import numpy as np

np.set_printoptions(precision=3, suppress=True)

def softmax_rows(S):
    S_shifted = S - S.max(axis=1, keepdims=True)
    E = np.exp(S_shifted)
    return E / E.sum(axis=1, keepdims=True)

X = np.array([
    [1.0, 0.0, 0.5],   # token 1
    [0.0, 1.0, 0.5],   # token 2
    [1.0, 1.0, 0.0]    # token 3
])

WQ = np.array([[1,0],[0,1],[1,1]])
WK = np.array([[1,1],[1,0],[0,1]])
WV = np.array([[1,0],[0,1],[1,-1]])

Q = X @ WQ
K = X @ WK
V = X @ WV

S = Q @ K.T / np.sqrt(Q.shape[1])
A = softmax_rows(S)
Z = A @ V

print("Attention scores S:")
print(S)
print("\nAttention weights A:")
print(A)
print("\nOutput Z:")
print(Z)
Attention scores S:
[[1.591 1.237 2.475]
 [1.945 0.884 1.768]
 [1.768 1.061 2.121]]

Attention weights A:
[[0.243 0.17  0.587]
 [0.458 0.159 0.384]
 [0.343 0.169 0.488]]

Output Z:
[[1.036 0.551]
 [1.15  0.234]
 [1.087 0.401]]

The row \(A_{i,:}\) tells us how token \(i\) combines information from all value vectors.

25.19 25.17 Practice problems

25.19.1 Problem 1: representation

Give three possible vector representations for a movie. Which representation would be useful for recommendation? Which would be useful for content search?

Show solution

Possible representations include:

  1. A hand-designed feature vector: genre, year, runtime, language, rating.
  2. A user-rating vector: ratings from many users, with many missing entries.
  3. A learned latent-factor vector: hidden dimensions such as action level, romance level, humor level, or more abstract learned patterns.

Content search may use the hand-designed feature vector or a text embedding of the movie description. Recommendation often uses user-rating vectors or learned latent-factor vectors.

25.19.2 Problem 2: cosine similarity

Let

\[ u = \begin{bmatrix} 1 \\ 2 \\ 0 \end{bmatrix}, \qquad v = \begin{bmatrix} 2 \\ 1 \\ 0 \end{bmatrix}. \]

Compute the cosine similarity between \(u\) and \(v\).

Show solution

We have

\[ u\cdot v = 1(2)+2(1)+0(0)=4. \]

Also

\[ \|u\|=\sqrt{1^2+2^2}=\sqrt{5}, \qquad \|v\|=\sqrt{2^2+1^2}=\sqrt{5}. \]

Therefore

\[ \operatorname{cosine}(u,v)=\frac{4}{5}=0.8. \]

25.19.3 Problem 3: neural layer

Let

\[ W= \begin{bmatrix} 1 & 2 \\ -1 & 1 \end{bmatrix}, \qquad b=\begin{bmatrix}0\\1\end{bmatrix}, \qquad x=\begin{bmatrix}1\\3\end{bmatrix}. \]

Compute \(\sigma(Wx+b)\) where \(\sigma\) is ReLU.

Show solution

First

\[ Wx= \begin{bmatrix} 1 & 2 \\ -1 & 1 \end{bmatrix} \begin{bmatrix}1\\3\end{bmatrix} = \begin{bmatrix}7\\2\end{bmatrix}. \]

Then

\[ Wx+b=\begin{bmatrix}7\\3\end{bmatrix}. \]

Both entries are positive, so ReLU does not change them:

\[ \sigma(Wx+b)=\begin{bmatrix}7\\3\end{bmatrix}. \]

25.19.4 Problem 4: attention shape

Suppose \(X\in\mathbb{R}^{6\times 10}\), \(W_Q,W_K\in\mathbb{R}^{10\times 4}\), and \(W_V\in\mathbb{R}^{10\times 8}\). What are the shapes of \(Q\), \(K\), \(V\), \(QK^T\), and the final attention output?

Show solution

We have

\[ Q=XW_Q\in\mathbb{R}^{6\times 4}, \qquad K=XW_K\in\mathbb{R}^{6\times 4}, \qquad V=XW_V\in\mathbb{R}^{6\times 8}. \]

Thus

\[ QK^T\in\mathbb{R}^{6\times 6}. \]

After row-softmax, the attention matrix is still \(6\times 6\). The output is

\[ AV\in\mathbb{R}^{6\times 8}. \]

25.19.5 Problem 5: low-rank meaning

Explain why a low-rank approximation can be useful in recommendation systems and image compression.

Show solution

In recommendation systems, a low-rank model assumes that user preferences and item properties can be explained by a smaller number of hidden factors. Instead of storing every user-item rating independently, the model stores user vectors and item vectors.

In image compression, a low-rank approximation keeps the strongest large-scale patterns in the image while discarding weaker details. This can reduce storage while preserving the most important visual structure.

In both cases, low-rank structure means that a large object has hidden simpler structure.

25.20 25.18 Challenge questions

  1. Why is cosine similarity often preferred over Euclidean distance for text embeddings?
  2. Explain attention using only the words “query,” “key,” “value,” and “weighted average.”
  3. Why would stacking only linear layers without nonlinear activations fail to create a truly deep model?
  4. In what sense is a recommender system a form of matrix completion?
  5. How does PCA help us understand hidden directions in a learned representation?
  6. Why might high-dimensional embeddings make search powerful but also difficult to interpret?

25.21 25.19 AI companion activities

Use an AI assistant as a learning partner, not as a replacement for your own reasoning.

25.21.1 Activity A: explain the formula

Ask:

Explain the attention formula \(\operatorname{softmax}(QK^T/\sqrt{d})V\) using only ideas from linear algebra.

Then check whether the answer mentions dot products, normalization, and weighted averages.

25.21.2 Activity B: create a toy embedding space

Ask:

Create a small 2D embedding example with five words. Show how cosine similarity ranks them for a query word.

Then compute the cosine similarities yourself.

25.21.3 Activity C: connect all chapters

Ask:

Make a concept map connecting vectors, matrices, projections, eigenvectors, SVD, PCA, Fourier, Haar, images, text, neural networks, and recommendation systems.

Then edit the map so it matches your own understanding.

25.22 25.20 Summary

This chapter gathered the main ideas of the book into one AI-centered story.

AI begins by representing the world numerically. Once objects become vectors and matrices, linear algebra becomes the grammar for computation.

  • Embeddings turn meaning into geometry.
  • Dot products compare vectors.
  • Cosine similarity powers search.
  • Matrices transform representations.
  • Neural network layers are trainable matrix machines with nonlinear gates.
  • Attention is a matrix of dot-product comparisons followed by weighted averaging.
  • Training is optimization over matrices and vectors.
  • SVD, PCA, and low-rank structure reveal hidden patterns.
  • High-dimensional geometry makes AI powerful but difficult to interpret.

The story of the book can now be stated in one sentence:

Linear algebra is the language that lets machines represent, transform, compare, compress, and learn from the world.

25.23 25.21 Looking forward

This book began with simple numbers and ended with the grammar of AI. But the subject does not end here.

From this point, you can continue in many directions:

  • numerical linear algebra,
  • optimization,
  • machine learning,
  • deep learning,
  • signal processing,
  • computer vision,
  • natural language processing,
  • recommendation systems,
  • scientific computing,
  • data geometry,
  • and mathematical foundations of intelligence.

The main lesson is not that every AI system is simple. Modern AI is complex. But its complexity is built from structures you now know how to recognize.

Vectors are not just lists of numbers.
Matrices are not just tables.
Projections are not just shadows.
SVD is not just a factorization.
Attention is not magic.

They are parts of a mathematical language.

And now you can read that language.