Features

Everything you need to finish the chapter.

Interactive 3D visualizations. Hand-traced PyTorch. Math, code, and diagram in the same paragraph. Progress synced across every device. Static-first pages that load before they hydrate.

Interactive 3D, not GIFs.

Every tensor, every transformer block, every crystal lattice is a real Three.js scene you can rotate, drag, and zoom — not a video.

Every line, hand-traced.

Teacher-grade walkthroughs: arguments, shapes, actual values, edge cases. No iteration is skipped, no parameter is hand-waved.

Math, code, diagram — together.

The equation, its PyTorch translation, and the mental model live in the same paragraph. No tab-switching, no context loss.

Notes bound to your reader.

Bookmarks, progress, and notes sync across devices. Pick up at the exact paragraph you marked, on any machine.

Static-first, instant load.

First screen is legible before a single byte of JS hydrates. No spinners, no skeleton flashes, no FCP penalties.

Past the demo.

Production tradeoffs, failure modes, cost curves, and the second-order tricks tutorials never finish.

3D visualizations
01

See the geometry, not just the formula.

Every concept that lives in space — attention heads, tensor reshapes, Bravais lattices, parametric surfaces, vector fields — is a real interactive 3D scene. Built with Three.js + @react-three/fiber, not pre-rendered video.

  • Drag any vector tip, watch its image vector update live
  • Rotate a crystal lattice in 3D, replicate the unit cell
  • Slice a tensor along any axis with a slider
  • Toggle between coordinate frames in one click
Vector v ∈ ℝ³drag me
v = (2.0, 1.4, 0.8)
Live
Code-trace walkthroughs
02

Every PyTorch line, explained at the value level.

We don't paste a code block and say `softmax over the last axis, see?`. Each function gets a teacher-grade card: every argument, the actual numerical values, the shape transitions, the edge cases. Loops are unrolled. Nothing is skipped.

  • Every line gets its own annotated card
  • Real values shown — not just shapes like (B, H, D)
  • All function parameters explained, not just the interesting ones
  • Loops expand iteration by iteration
attention.py · line 14●●●
q = x @ W_q
k = x @ W_k
attn = softmax(q @ k.T / √d)
attn — shape (B, H, S, S)
Each row sums to 1.0. Position 3 attends mostly to position 1 (0.62), then 5 (0.21) — that's where the verb lives.
↳ 4 of 9 annotations shown
Same paragraph
03

Math, code, and diagram never separate.

Tutorials make you tab-switch between the LaTeX, the PyTorch, and the picture. We render them inline, side-by-side, in the same paragraph — because that's how working engineers actually think.

  • KaTeX inline math (γ, σ², ∇, all of it)
  • Syntax-highlighted code blocks with copy buttons
  • Hand-built SVG diagrams next to the equation that produced them
  • Tables for numerical walkthroughs, not screenshots of spreadsheets
Chapter 4 · §2 · BatchNorm

We normalize each channel: ŷ = (x − μ) / √(σ²+ε), then scale-and-shift with γ, β ∈ ℝᶜ.

y = γ * (x - μ) / (σ.pow(2) + ε).sqrt() + β
channel 2 highlighted
Reading state
04

Your book remembers where you stopped.

Bookmarks, progress, notes, highlights — bound to your account, synced across every device. Open the laptop, the iPad, the phone; pick up at the exact paragraph you marked.

  • Per-section progress bars in the sidebar
  • Bookmark any heading or equation
  • Margin notes attach to specific paragraphs
  • Recently-opened list across all volumes
Reading progress · all volumes
Transformers from scratch62%
Deep Learning38%
Linear Algebra91%
Probability & Statistics14%
Last bookmark:Transformers · ch 4 · §2
Static-first performance
05

Legible before a single byte hydrates.

Every section page is server-rendered HTML — the math, the code, the diagrams arrive as the page does. The interactive 3D bits hydrate progressively. No spinners, no skeleton flashes, no first-contentful-paint penalty.

  • Sub-second LCP on a cold cache
  • Zero client JS required to read a chapter
  • Heavy 3D viewers dynamic-imported only when scrolled into view
  • Built for the long read on a slow connection
Lighthouse · /books/transformer/ch-04
98
Perf
100
A11y
100
BP
100
SEO
LCP0.8 s
FCP0.4 s
TBT20 ms
CLS0.00
Plus the small things

Quality of life that compounds.

None of these are headline features. Together they're why you actually finish the chapter.

Tensor + attention playgrounds
Cross-device bookmarks
New volumes the day they ship
Categorical book index
Margin notes per paragraph
Offline PDF downloads

Try one chapter. See for yourself.

The Reader plan is free forever. Pick any volume, open any chapter, and feel the difference in 90 seconds.