The guide
The deep learning topic built attention. This topic builds the parts that turn attention into a modern language model, most of which are small changes that add up.
A causal mask lets one forward pass train on every position at once. RoPE rotates queries and keys so that attention sees relative position. RMSNorm and SwiGLU are cheaper or better versions of LayerNorm and the ReLU MLP, and nearly every recent open model uses them.
Then there is inference. A KV cache makes generation cost one new token's work per step instead of the whole sequence. Top-p sampling chooses what to say next. LoRA makes fine-tuning cheap by learning a small low-rank update to frozen weights.