Miatz playground

The Spotlight

Watch AI decide which words matter to which.

Coming to the Build-LabAIE-101BYOK AIAI-Concept-Vizdev · data

The Demystify signature

The pronoun-resolution moment: the exact layer/head where 'it' lights up on 'trophy' not 'suitcase,' turned into a visible, replayable event instead of an assertion that the model understands context.

How it works

What goes in, what comes out

What it does

Learner picks an ambiguous sentence ('the trophy didn't fit in the suitcase because it was too big') and sees animated lines connecting each word to every other word, line thickness weighted by attention strength for the token currently selected. Scrubbing a layer/head slider changes which connections light up, showing that meaning is built through several different weighted lookups, not one.

You bring

A preset ambiguous sentence, or the learner's own short sentence.

You get

An animated weighted-graph overlay on the sentence, plus a plain-language caption naming which word 'it' resolved to and at which layer/head that resolution first appears.

You control

Layer slider, attention-head selector, sentence presets (ambiguous pronouns, negation, long-range references), click-to-isolate one token's attention row.

Where it's used

Demo inside AIE-101 (paired with the Prompt Lab playground) and the AIE-130 model-internals elective; secondary lead magnet given high shareability of attention art.

self-attentionQKVattention headscoreference resolution

Optional · watch & try

Go deeper, elsewhere

Hand-picked public explainers and open tools that complement this one — always optional, never required, never graded.

Attention in transformers, step-by-step | Deep Learning Chapter 6 (3Blue1Brown)Gold-standard visual explainer of the attention mechanism inside a transformer Transformer ExplainerLive in-browser GPT-2 running with every transformer stage visualized as you type How LLMs Work: A Visual GuideVisual walkthrough of the LLM pipeline from input text to output token How LLMs REALLY Work (Simple Explanation)Plain-language mental model for what an LLM is actually computing LLM Visualization (bbycroft)3D walkthrough of a GPT-style model's full forward pass, weight matrix by weight matrix A visual introduction to tokenization in LLMs | Byte Pair Encoding AlgorithmShows how byte-pair encoding splits text into subword tokens step by step

Concepts click when you open the machinery.

Three labs are already live and free — the same hands-on style this playground brings to its module.

Play the free labs Take the DSAT