Notes & Essays

Before Ask Yeremia on /lab went public, I worried about stolen API keys, runaway bills, and visitors sweet-talking the model past its rules — here is the four-part checklist I actually ran.

Retrieve top-k facts, refuse below 0.6 without calling the LLM, otherwise answer from a grounded prompt with citations — plus model fallback when backoff on one flash model is not enough.

Embeddings turn text into 3072-number vectors of meaning. Cosine similarity scored two coffee sentences at 0.801 and an unrelated sentence at 0.526 — far from zero.

ChatGPT is an app. The API is for programs: a streaming TypeScript call with @google/genai, a backoff retry for 503 and 429, and an API key that stays out of git.

Seven facts embedded once, ranked by cosine similarity. Mazecare ranked first on a healthcare résumé question without the word health — FHIR and clinical notes were enough. KiriminAja at 0.627 is the warning. Rendang topped out at 0.550.

An LLM is a next-token predictor. Tokens set cost and limits, sampling and temperature change how varied the wording is, and a sourced answer is retrieval before generation.

Working implementations of debounce, abortable fetch, backoff polling, and windowed lists — plus the JavaScript event loop and React engine that decide whether those patterns actually hold.

Forget the engine. Picture a search box and a 2,000-row list. Fiber is why the letters can appear now while the slow list is allowed to catch up.

A home-only overlay, sessionStorage (not localStorage), useSyncExternalStore, and Motion's AnimatePresence — with the hero painted underneath so LCP is not gated on the intro.

A 91 mobile score was LCP, not a missing sitemap. How this Next.js portfolio split Motion, labs, and the mobile sheet — and the traps that look like dead code but aren’t.

XSS, cookies, CSRF, CORS, CSP, and lockfiles — what React and Vue already escape, where v-html and dangerouslySetInnerHTML still bite, and how that maps onto App Router and Nuxt.

Call stack, microtasks, closures, this, and prototypes — then the TypeScript you use every day: narrowing and discriminated unions. Why it behaves that way, not just the syntax.

A custom font is a family name plus real files for each weight. The browser only uses bold, medium, or italic if you registered those faces — usually as WOFF2 via @font-face.

Share state across related UI pieces without prop wiring — and when a single component or shadcn is still the better choice.

Both run after render, but timing decides whether the user sees a flicker. Start with useEffect; reach for useLayoutEffect only when you must measure or sync layout before paint.

Reconciliation is how React compares the next UI tree with the previous one so it can update the real DOM as little as possible.

Re-renders are normal. Unnecessary ones usually come from unstable props, broad state, or overused context — fix the cause before sprinkling memo everywhere.

Controlled inputs keep form data in React state. Uncontrolled inputs keep it in the DOM. Prefer controlled for app forms; use uncontrolled for simple cases and file inputs.

memo caches a component, useMemo caches a value, useCallback caches a function. Use them together when prop identity matters — not on every line of code.

Learning in public: Hermes Agent on Biznet GIO Cloud, DeepSeek API, Discord gateway live (checkpoint + /sethome) — stack, install path, and what part 2 will harden.