The diff: 2017 transformer → a modern LLM block

Click any row for the one-line reason. Press play to watch the 2017 design get rewritten one decision at a time.

Attention Is All You Need (2017)Llama / DeepSeek-style block (2026)

Not every model makes every one of these choices, but a Llama 3 block makes nearly all of them, and DeepSeek-V3 swaps in MLA and MoE on top.