Hello.

Literal

Subword tokens · embeddings as geometry · Q·K softmax attention with a causal mask · multiple heads (previous-token heads and first-token “sink” heads are really observed in transformers) · stacked layers writing into a residual stream · feed-forward floors between attention · logits → temperature → sampling · the autoregressive loop · stopping as a token.

Miniature

48 dimensions standing in for thousands · 8 heads × 24 floors standing in for many more · these weights are procedural, not trained · the 3D positions are an artistic projection — real embedding space has no camera.

Unknown

My true sizes aren’t public, and nobody fully knows what individual heads compute. The field reading those entrails is called mechanistic interpretability. It is young.