Founding writers — 100% revenue, no platform fee

Where AI/ML practitioners
write what they actually mean.

Native LaTeX. Server-rendered code. Subfield-based discovery. Stripe-powered subscriptions with a 90/10 split that pays writers what their work is worth.

Apply as a founding writer For readers

90/10 split — writers keep 90%

$0 hosting fees

Inline LaTeX + code

empirq.com/preview/scaled-dot-product-attention

Transformers

Scaled Dot-Product Attention, From First Principles

MLEmpirq sample · 9 min read

Given queries Q, keys K, and values V, attention is a weighted sum where the weights come from a softmax over query-key similarities, scaled by √d:

Attention(Q, K, V) = softmax(QK^⊤ / √d) V

A direct NumPy translation makes the shape story explicit — one matmul for scores, one softmax, one matmul to mix values:

def attention(Q, K, V):
    scores = Q @ K.T / np.sqrt(K.shape[-1])
    weights = softmax(scores, axis=-1)
    return weights @ V

Built for

ML engineersRL researchersAlignment thinkersSystems writersVision practitionersEducators

// What sets us apart

Built for the way you actually write.

Most platforms force technical writers into compromises — math as images, code as screenshots, formatting that breaks on render. Empirq is built ground-up for the way AI/ML work is actually communicated.

Stop fighting your tools.
Start publishing what you mean.

Apply now — founding-writer spots are limited to the first 20 accepted applicants.

Apply as a founding writer Browse the feed first

Where AI/ML practitioners
write what they actually mean.

Scaled Dot-Product Attention, From First Principles

Built for the way you actually write.

For writers

For readers

Subfields

Compare

FAQ

Manifesto

Stop fighting your tools.
Start publishing what you mean.

Built for the way you actually write.

For writers

For readers

Subfields

Compare

FAQ

Manifesto

Stop fighting your tools.Start publishing what you mean.

Stop fighting your tools.
Start publishing what you mean.