Inside the Transformer
Attention Lab
Preset: Coreference ("it" -> "animal")
Preset: Context Polysemy ("bank")
Preset: Syntax & Modifiers ("fast compiler")
Head 1: Coreference
Head 2: Syntactic Dependencies
Head 3: Positional / Local
Head 4: Semantic Associations
Step Stage
Reset
1. Token Sequence & Attention Arcs
Focus: it [Token #7]
Attention(Q, K, V) = Softmax( (Q · Kᵀ) / √dₖ ) · V
2. Attention Weight Matrix (N × N Heatmap)
Softmax Probabilities
3. Projected Vectors for Focused Token
dₖ = 4
Query (Q)
W_q·x
[+0.42, -0.81, +0.15, +0.94]
Key (K)
W_k·x
[+0.38, -0.79, +0.19, +0.88]
Value (V)
W_v·x
[+0.12, +0.65, -0.44, +0.31]
Cosine Positional Waves merged with token embedding:
PE(pos, 2i) = sin(pos / 10000^(2i/d))
4. Feed-Forward & Next Token Prediction
LayerNorm + MLP
Softmax Temperature (τ):
0.7