Modern architectures
What changes when context gets expensive?
A new attention name can describe a different storage layout, a different set of visible entries, or a faster way to execute the same mathematics. Compare the design choice before comparing the headline.
The article
Related reading
02
03
Inside attention, one tensor at a time
Follow a complete decoder through Q, K, V, rotary positions, grouped-query attention, residual connections, and next-token logits. Every shape has a job.
KV cache: what can a model safely reuse?
Derive why old keys and values remain valid, inspect what each layer stores, and prove that cached decoding agrees with a full forward pass.