A sentence becomes a constellation of nodes. Drag, touch, fire — then scroll inside.
You hand the model text: “The cat sat on the mat.” It won’t read letters — it prepares to see meaningful pieces.
The sentence is chopped into pieces, each mapped to a number (ID). From here the model works in numbers, not words.
Every ID turns into a point in a high-dimensional meaning space. Similar meanings land close together.
Each token computes how much to attend to the others. With a causal mask, a token only looks at the ones before it.
Attention + feed-forward repeats, stacked dozens deep (e.g. ×32). Every pass refines the representation.
The final layer scores the entire vocabulary into a probability distribution. The top token is picked (or sampled).
The chosen token is appended and the whole process runs again. Word by word, that’s how text is generated.
This tiny network is really billions of parameters — which is exactly why VRAM matters. Does your model fit a GPU?