x^{T}

Transpose

Operator: it does something

In plain English

Flip a matrix across its diagonal: rows become columns. A column vector $x$ becomes a row vector $x^T$ . The point is bookkeeping: $x^T y$ multiplies a row by a column and produces a single number (the dot product), while $x y^T$ produces a whole matrix. In $QK^T$ (attention), transposing $K$ lines up every query with every key. The $T$ is an operation, not an exponent.

As code

const A = [[1, 2, 3],
           [4, 5, 6]];               // 2×3
const AT = A[0].map((_, j) => A.map((row) => row[j]));
// [[1,4],[2,5],[3,6]]              // 3×2

See it

\begin{bmatrix}1 & 2 & 3 \\ 4 & 5 & 6\end{bmatrix}

x^{T}

flips rows and columns. A 2×3 matrix becomes 3×2. This is why

QK^{T}

works: it lines up every query with every key.

What it means, field by field

Mathematics

Matrix transpose; symmetric matrices satisfy A = Aᵀ.

AI / ML

QKᵀ in attention; Wᵀ in tied embeddings and in backpropagation’s backward pass.

Physics

Rotation matrices satisfy Rᵀ = R⁻¹; the conjugate transpose † in quantum mechanics.

In a research paper

Scaled dot-product attention (Vaswani et al. 2017).

\mathrm{softmax}\!\left(\frac{QK^{T}}{\sqrt{d_k}}\right)V

Kᵀ makes every query-key pair meet in one matrix multiplication, producing an n×n grid of similarity scores.

Common misreadings

Reading xᵀ as "x to the power T".
Ignoring shapes. Most transpose confusion disappears if you write each matrix’s dimensions next to it.

Connected symbols

\mathbf{a}\cdot\mathbf{b}

Dot product (or multiplication)

\in

Element of (set membership)

\|\mathbf{x}\|

Norm (length of a vector)

Where it appears on this site

∂ Scaled dot-product attention § 6. The Mathematics Inside Modern AI