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

ATTENTION GLYPH

A UNIFIED VIEW OF INFORMATION FLOW ACROSS A SET OF VECTORS.

A=softmax⁡(QKTdk)A = \operatorname{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)

d_model = 768 n_heads = 12 seq_len = 1024

ALISSAJOUS

x=A sin⁡(at+δ)x = A\,\sin(at + \delta)
y=B sin⁡(bt)y = B\,\sin(bt)
a=3b=2δ=π/2a = 3 \qquad b = 2 \qquad \delta = \pi/2a=3b=2δ=π/2a = 3 \quad b = 2 \quad \delta = \pi/2

BPHYLLOTAXIS

θ=nφ\theta = n\varphi
r=cnr = c\sqrt{n}
φ=(1+5 ) / 2\varphi = (1 + \sqrt{5}\,)\ /\ 2

CVECTOR FIELD

F(x, y)=( −y, x )F(x,\ y) = (\,-y,\ x\,)
∇⋅F=0\nabla \cdot F = 0

DSIGNALS

x=A sin⁡(at+δ)x = A\,\sin(at + \delta)
∑t wt xt\textstyle\sum_t\ w_t\, x_t

LATENTHow AI Works

Motion Design Style-Frame Book

A VISUAL EXPLORATION OF THE MATHEMATICS, MECHANISMS AND BEAUTY BEHIND MODERN ARTIFICIAL INTELLIGENCE.

  1. OUTPUTyy
  2. TRANSFORMfθ(x)f_\theta(x)
  3. ATTENTIONA=softmax⁡(z)A = \operatorname{softmax}(z)
  4. EMBEDDINGx∈Rdmodelx \in \mathbb{R}^{d_{\text{model}}}
  5. LATENT SPACEz∈Rdz \in \mathbb{R}^d