Atlas — the pipeline, node by node

The Atlas is a map of the inside of the pipeline. Where the other P10 tracks teach each stage one module at a time, the Atlas lays every stage's internal wiring side by side — the same diagrams, the same demo corpus, the same numbers — so you can see the whole machine at once and zoom into any part of it.

Every diagram uses one shared grammar of three marks: learned matrices (the parameters training changes), tensors (the data flowing through) and operations (the steps applied). Every node is clickable — the inspector under each diagram shows the actual numbers computed from the demo corpus: real token ids, real Q/K/V matrices, real attention weights.

Two things here go beyond static diagrams:

• Architecture knobs — width, heads, layers, segment embeddings, the causal mask, Pre- vs Post-LN. Change a knob and every affected inspector recomputes; the shape badges show what grew or shrank.
• A live trainer — train the full encoder stack with one of three objectives (BERT's masked-token, GPT's next-token, or an illustrative MSE loss) and watch the loss fall and the weight matrices drift step by step. Save the weights you trained as a checkpoint (stored in your account) and fine-tune from it later — the classic transfer-learning story, in demo scale.

All of it is pure numpy at demo scale (d_model = 16, a 24-token vocabulary) — small enough to display every number, real enough that the mathematics is exactly the one the live model uses. When you want the real thing, the Console and Inspect modules run an actual PyTorch BERT behind the same concepts.

Checking access…