https://notes.likegeek.top/ Applied ML systems, written slowly. en-us https://notes.likegeek.top/posts/serving-small-models-on-one-gpu/ https://notes.likegeek.top/posts/serving-small-models-on-one-gpu/ 2026-09-18 Most of the cost sits in the KV cache and in bad batching, not in the parameter count you quote in a blog post. https://notes.likegeek.top/posts/eval-harness-that-survives-a-prompt-change/ https://notes.likegeek.top/posts/eval-harness-that-survives-a-prompt-change/ 2026-10-02 If a prompt tweak can move the score more than a model swap, the leaderboard is measuring your wording, not the system. https://notes.likegeek.top/posts/when-rag-is-just-a-slow-grep/ https://notes.likegeek.top/posts/when-rag-is-just-a-slow-grep/ 2026-10-14 Embeddings help when the query and the document do not share words. They do not excuse a messy corpus. https://notes.likegeek.top/posts/cuda-memory-is-a-product-constraint/ https://notes.likegeek.top/posts/cuda-memory-is-a-product-constraint/ 2026-10-27 The feature list is often decided by what still fits after the cache, the image, and the safety window. https://notes.likegeek.top/posts/logs-you-can-debug-at-2am/ https://notes.likegeek.top/posts/logs-you-can-debug-at-2am/ 2026-11-04 A request id, a model name, a token count, and a reason for the retry will save you more time than another dashboard. https://notes.likegeek.top/posts/shipping-a-model-change-without-a-mystery-reg/ https://notes.likegeek.top/posts/shipping-a-model-change-without-a-mystery-reg/ 2026-11-09 Treat a new checkpoint like a schema change: small, reversible, and compared on the same traffic.