What this is
Decode Notes is a static notebook about inference systems: the part of an AI stack that turns weights into tokens under a latency budget.
The posts stay close to mechanisms. A number is included only when the formula that produced it is on the page. There is no product announcement, no model ranking, and no claim that a public benchmark transfers to an unmeasured workload.
Topics that belong here:
- decode loops, speculative verification, and why a faster draft model can still lose
- KV-cache bytes, grouping, paging, and what those choices do to batch size
- routing, communication, and capacity in mixture-of-experts servers
- evaluation hygiene: contamination, prompt sensitivity, and decoding settings