
Tokenizer Split
7.5sA sentence breaks apart into coloured token chips that settle into rows, each labelled with its own token id.
Attention Heads

Each word in turn attends to the others, its links thickening with the weight it gives them as the heads cycle through.
Animate transformer attention with each word weighting the others as the heads cycle, for LLM, NLP and AI architecture explainers.
Part of Networks & AI. New to this kind of graphic? Read How to Animate a Neural Network Diagram for a Video.
Transparent exports and editor imports are explained in the transparent video guide, and every download is free under the commercial license.
Content: title and labels. Typography: font. Style: background, text color, primary, secondary, accent and thickness. Layout: aspect ratio and scale. Motion: duration, speed and variation.
Attention Heads runs 10 seconds by default, and you can set any length up to 60 seconds before you export. It plays once.
Attention Heads helps explain why transformer models handle context well. Watching one word's links thicken toward the words it depends on makes a hard idea visual, which suits lecture videos, technical blog companions and conference talks on model architecture.

A sentence breaks apart into coloured token chips that settle into rows, each labelled with its own token id.

An editor panel types your lines, then a dimmed AI suggestion streams in after the caret and turns solid when it is accepted.

A prompt types into the input box, the send button taps and the answer streams back token by token under a moving shimmer.

A labelled node graph builds itself edge by edge with arrowed links, then packets run the routes between agents.

A glowing assistant orb moves through idle, listening, thinking and speaking, its waveform rising and falling with each state.

A metric traces along inside its normal band until one reading spikes clear of it and gets ringed and flagged.

A request packet carries the endpoint across to the API, the server pulses while it works, and the response returns with its status.

New blocks with hash labels slide in, pulse and chain onto a scrolling ledger while data flows through the links.