Day 54 drills

Paper first. Warm-up before the timed run.

D1 (why “design an AI chat product” is mostly not about the model, and the familiar week-7 thing the model is most like):

D2 (TTFT: what it is, why it beats total response time in a chat UI, and the technique that makes a long answer feel fast):

D3 (GPU scarcity: how you keep a GPU busy per Day 37, and what you do when demand exceeds capacity instead of crashing):

D4 (RAG: why most failures are retrieval failures per Day 39, and two ways to beat naive top-k similarity):

D5 (cost: three concrete levers to cut the token bill, and why a long conversation costs more every turn):