Follow the mechanism.
An opening systems notebook. Each explanation states its scope and points back to a source.
Attention connects positions.
The Transformer paper introduces an architecture based on attention rather than recurrence and convolution. In an explanation diagram, queries and keys can help show which positions interact; that diagram is a teaching model, not a literal display of a model’s understanding.
Inspect the source-scoped Transformer note ↗Context is an input, not a guarantee.
Retrieval makes external passages available to generation. Inspecting the passages and checking whether an answer is supported are separate tasks. Our retrieval demo stops at context selection so that this boundary stays visible.
Inspect selected context ↗Next studies
KV cache, quantization and mixture-of-experts are planned study topics. They are not yet source-reviewed articles or working simulations in this edition.