Svelte, September 2026: SvelteKit 3 RC, SvelteMap, and sv migration
Svelte 5.57 adds SvelteMap getOrInsert; SvelteKit 3 Release Candidate; ai-tools replaces mcp; sveltekit-3 task in sv migrate.
Stuzhuk Lab — Chemistry of Code
Chemistry of Code
← All posts
Tag
News
All blog posts with this tag.
Haiku R1/beta6: first release in two years, 25th anniversary
On August 26, 2026 Haiku R1/beta6 shipped — two years after beta5, a week after the project's 25th birthday.
CMP 90HX for LLM: why a mining GPU slows down and what “six-seven” fixes
MoE tolerates CMP 90HX; dense models don't — compute paths are capped. Driver mod via V67 nearly doubles throughput but doesn't erase mining-card limits.
Lovable: 400 routes from Next.js to TanStack Start without a big bang
Proxy dual-run, #shared core, adapters, and agents: lovable.dev left Vercel for its own stack at 42M MAU.
Cloudflare freed ~100 TB by reshaping the DNS cache
Five Rust layout changes in Big Pineapple cut per-entry memory 56% and sped inserts 43% for 1.1.1.1-scale DNS.
Bun 1.4: Rust runtime, closer Node.js compatibility, less memory
+1,517 Node tests, up to 35% less memory, ~2× faster Linux startup—first stable release after the Zig→Rust rewrite.
How to evaluate LLMs before production: GitHub secret scanning lessons
Start with the product decision and guardrails, then the model—offline eval as integration tests plus error analysis.
Ukrainian LLM Leaderboard: ranking models on Ukrainian
Ukraine’s Ministry of Digital Transformation, UCU, and lang-uk open an Hugging Face leaderboard for LLMs on Ukrainian: translation, summarization, retrieval, ZNO.
WordPress 7.0: PHP-only blocks without React or a build step
autoRegister in WordPress 7.0 lets you register a block in PHP alone—great for migrating legacy themes, weak for interactive new blocks.
Cheap LLMs for agents: why token savings raise task cost
Why distilled models spawn extra tool calls, how benchmarks mislead, and which metrics beat price-per-million for agent TCO.
NanoGPT Speedrun: 153 agent runs across frontier models
Prime Intellect’s nanoGPT speedrun: 18 models, validated top score for Kimi K3kimi-code, plus 41 full agent trajectories.
Problem Engineering: specs beat prompt tricks for AI-generated code
Five-layer problem specs for AI coding: invariants, data contracts, concurrency, fault topology, and observability—before you prompt.
Self-hosted LLM for 25 developers: prefix cache beats raw model choice
vLLM on Blackwell: 96.9% prefix cache hit rate, 452:1 input/output, and why teams keep inference in-perimeter for control—not token savings.
Vector databases for AI agents in 2026: beyond raw ANN speed
Qdrant, Pinecone, Weaviate, pgvector, and Milvus compared for agent memory: filtering, hybrid search, tenancy, and quantization.
CrewAI on Go with zero deps: agent orchestration on the standard library
crewai-go mirrors Agent, Task, and Crew in pure Go stdlib: one binary, fast cold start, and explicit task dependencies.
React Native in 2026: New Architecture, native code, and AI boundaries
A real-app walkthrough: Fabric and JSI, where work runs, bare vs Expo, keyboards on complex forms, and where AI actually helps in RN.
Semantic context for LLMs: how LSP helps a model understand code
Part 4 of a completion series: Language Server Protocol supplies types, definitions, and references — a layer on top of repository text search.
Bun's Zig-to-Rust rewrite: 'never rewrite from scratch' was a frozen cost model
64 agents moved Bun's core in 11 days. The rewrite taboo was about cost, not metaphysics — and review debt did not disappear.
JavaScript in the agent harness: when a script is cheaper than model reasoning
A sandboxed JS interpreter replaces some tool hops and LLM arithmetic. Notes from a Habr deep dive on LM Studio and LangChain.
AI video in a content pipeline: edit the frame, don’t invent the clip
Runway Aleph in production: shot manifests, one op per pass, preserve clauses, human QC, and prompt caching.
Claude system prompts: lessons from Haiku 3 to Opus 5
Anthropic publishes system prompts: not magic personas, but decision rules, tools, uncertainty handling, and UX for long agent tasks.
AI agents need a platform, not just faster typing
Why coding assistants don't shorten monolith release cycles without trackers, runtimes, data access, and MCP—SimpleOne's low-code cases.
HTML over WebSockets: SPAs with barely any JavaScript
The LiveView pattern: server-rendered HTML over a persistent channel—no separate API contract or heavy client framework required.
LLMs overturn old math claims—verification is the bottleneck
Counterexamples and proofs from models, a sandbox escape, Lean certificates: don't treat chain-of-thought as faithful reasoning.
Air-gapped MLOps on Kubernetes: MLflow, DVC, and OpenBao
A DevOps-oriented blueprint for training and serving models behind the perimeter—GitOps, secrets, and GPU outside the cluster.
React useLatest: fresh state in async callbacks without extra renders
useLatest from @reactuses/core: a layout-effect ref against stale closures after await, timers, and subscriptions.
Up to 8,000 Ukrainians: free Oracle IT training and certification
UK-Ukraine TechBridge and Diia City: Oracle University MyLearn until May 2027 — cloud, data, DevOps, Java, and a free exam on request.
SvelteKit Remote Functions, TypeScript 6, and CLI Plugins
A Dev.to walkthrough argues that typed remote functions and experimental CLI plugins can simplify SvelteKit feature delivery.
Giving an AI Agent MCP Access to the Yandex Webmaster API
A practical MCP server design for Yandex Webmaster: eight task-oriented tools, OAuth sign-in, and safe API actions for AI agents.
OpenAI brings frontier AI to US scientific research
OpenAI’s National Science Initiative funds model access for labs and universities, but validation, governance, and expert review remain decisive.
Why an AI Agent Can Cost Less Than RAG
More LLM calls do not necessarily cost more: support teams should measure the cost of resolving a question, not generating one answer.
Your Local LLM Is Not Automatically Private
CVE-2026-7482 in Ollama shows how a local LLM server can expose process memory when it accepts untrusted model files and network access.
CSS conic-gradient(): triangles, rays, and pie segments
conic-gradient() paints color around a point, enabling pie segments, rays, triangles, and interactive effects without extra markup.
Kimi K3: What Open Weights Mean for Developers
Moonshot AI has released Kimi K3 weights. Here is what its MoE design, 1.4 TB footprint, and licence mean before adoption.
Training a 6.4M-Parameter Transformer for Recipe Tasks
A developer built a compact PyTorch transformer from scratch, trained it for recipes, and runs it on CPU without external APIs.
Open-weight AI is reaching its Kubernetes moment
Open-weight models are becoming a portable platform for developers to run, adapt, and extend with an independent tooling ecosystem.
Arcovia maps architecture health in React and Next.js
Arcovia analyzes React and Next.js projects, maps dependencies, explains its architecture score, and highlights refactoring priorities.
How TypeScript Generic Constraints Prevent Runtime Errors
How `extends`, `keyof`, and indexed access let TypeScript catch invalid keys and values before an application runs.
Gemma 4 26B on a 13-Year-Old Xeon: What CPU-Only AI Can Really Do
A Dev.to author reports running Gemma 4 26B on a Xeon E5 v2 without a GPU. Here are the requirements, claims, and caveats.
Production RAG: Lessons from Processing 10,000 Listings a Day
A production RAG pipeline depends on document chunking, embeddings, pgvector, cost controls, and observability—not just an LLM.
Open LLMs in production: 8 lessons from llama.cpp, Gemma, and Qwen
Eight practical lessons for running open LLMs: inference control, task-specific model selection, prompt templates, and structured output.
Bun in Rust, TypeScript 7, and npm 12: JavaScript Tooling Shifts
Bun rebuilds on Rust, TypeScript 7 speeds up compilation, and an npm incident underscores the operational cost of dependencies.
A self-updating open-data site with GitHub Actions, Pages, and data branches
Village Finder uses GitHub Pages, Actions, releases, and isolated branches to publish changing open data without an application server.
A 744B AI Model in 25GB of RAM: Inside Colibri
Colibri runs the GLM-5.2 MoE model without a GPU by keeping dense weights in RAM and streaming experts from NVMe. Here are the trade-offs.
Xiaomi Released a Frontier AI Model Anonymously—and Developers Chose It
Hunter Alpha was an early MiMo-V2-Pro build. Its uptake shows how price, output quality, and data jurisdiction shape model selection.
Turn a Job Description Into a Focused Mock Interview Pack
A small Node.js script extracts signals from a job description and turns them into technical prompts, story gaps, and a seven-day practice plan.
Meta open-sources its Astryx design system for React
Astryx combines React, StyleX, 160+ components, themes, and templates, offering a look at Meta's approach to large-scale UI consistency.
Frontend without the wait: TSGO, Oxlint, Rsbuild, and React Compiler
A practical frontend tooling roundup: faster type checks, linting, builds, API contracts, and React optimisation.
Git 2.55 brings incremental maintenance and easier history fixes
Git 2.55 adds incremental multi-pack indexes, a direct way to fix earlier commits, and parallel configured hooks.
React: debounce searches and autosaves without extra requests
How to choose useDebounce or useDebounceFn in React, avoid stale closures, and safely control pending work.
Astro 7: Rust core, route caching, and AI-friendly tooling
Astro 7 moves its compiler and Markdown processing to Rust, speeds up builds, and adds route caching and agent-aware development tools.
How to Stop an AI from Inventing Numbers in Your Data
A practical Claude Desktop-inspired setup: DuckDB performs calculations while the model works through schemas, tools, and guarded access.
From Local LLM to a Safer API with FastAPI and Ollama
A FastAPI wrapper for a local SQL-generating LLM shows why read-only execution and compact schema context matter.
Fable 5 and qa-mcp: reverse-engineering the 1C TestClient protocol with AI
Habr case study: from Vanessa Automation to native TestClient in Python — where Codex stalled for a week, Fable 5 mapped the protocol in a day for qa-mcp.
YouTube transcripts for summarization and RAG: REST instead of scraping
Fetch JSON transcripts with timestamps via a RapidAPI endpoint — for LLM summarization, vector search, and subtitles without brittle page parsing.
Shumai: an open-source Frame.io alternative for creative teams
Shumai offers media storage, frame-accurate review, collaboration, and an AI agent with semantic search. Deploy in minutes with Docker Compose or npm.
Fine-tuning Qwen 3 0.6B: from 10% to 92% question classification accuracy
A home chatbot experiment: fine-tune a tiny local LLM with Unsloth and two-letter category codes instead of label names for metadata-aware RAG.
From pip install to production: a self-healing AI agent in 10 minutes
Practical guide to NeuralBridge SDK: multi-provider failover, cascade self-healing, and observability for OpenAI-compatible LLM calls in Python.
euv: a Rust and WebAssembly UI framework for the browser
Intro guide to euv: virtual DOM, reactive signals, the html! macro, and mount — declarative frontend with Rust type safety compiled to WASM.
I gave Tailwind typed props. Then it ate React Hook Form.
How the DashForge author turned Tailwind classes into typed props and built application-aware components: forms, visibleWhen, RBAC — without Controller per field.
React Router v8: boring, but good
React Status #479: React Router v8 as an intentionally boring release, Rust React Compiler port, RN 0.86, TypeScript 7 RC, and React ecosystem news.
TypeScript 7.0 RC: Go-based compiler and up to 10× speed
Microsoft shipped TypeScript 7.0 RC on a new Go foundation: 6.0 compatibility, parallel builds, @typescript/typescript6 side-by-side, and a rebuilt --watch.
React 19: concurrent rendering, Actions, and Suspense in production
Deep dive into React 19 concurrent rendering: urgent vs transition updates, useTransition, useOptimistic, and Suspense boundaries without blocking the UI.
Biome vs Oxlint in 2026: Replacing ESLint
Rust-powered linters run 50–100× faster than ESLint, but choosing Biome or Oxlint depends on custom plugins, type-checking, and whether you want one tool instead of ESLint plus Prettier.
GitHub Opened a Multilingual Repositories Dataset for AI
80M+ classification rows across 40M public repos — README, issue, and PR language signals under CC0 for evaluating multilingual developer AI tools.
Zero-JS was not enough: edge-caching SSR HTML on Cloudflare Workers
How the author stabilized Core Web Vitals after dropping React: Cache API, cookie-gated HTML, removing Zod from the browser, and refactoring Astro forms.
The new SDLC: vibe coding vs agentic engineering
A senior dev's honest take: Agent = Model + Scaffold, context engineering, the factory model, and why vibe coding does not scale past prototypes.
npm and pnpm Introduce Staged Package Publishing
Staged publishing lets you upload to the registry, verify, then promote — reducing supply-chain blast radius. Plus pnpm 11.3 trustLockfile and npm 12 prerelease.
Six Laravel performance lessons from real client projects
Eloquent abuse, missing indexes, cache, code structure, and client communication — what actually slows Laravel apps down.
Five container runtimes: when runc is not the answer
runc, gVisor, Kata, Firecracker, and WASM on one Go service — cold start, memory, and threat-model fit.
Why a million Rust coroutines weigh less than 100k Python ones
async fn becomes a state machine everywhere — but Rust puts it on the stack while CPython puts it on the heap. Memory benchmarks and practice.
What CSS is still missing in 2026
State of CSS, layout gaps, centering, declarative partial updates, and how AI reshapes site value.
layercache: stopping cache stampedes in Node.js
L1 + Redis, single-flight, and graceful degradation — layercache benchmarks against thundering herds.
Smart document registry: free Google Drive OCR and Gemini quota tricks
Google Apps Script, hidden Drive OCR, Gemini key rotation, and LockService — a Habr pipeline for cataloging huge archives.
Extend UI — open-source React components for PDF, DOCX, and e-sign
14 MIT-licensed components from extend.ai: document viewers, bounding box citations, file upload, and e-signature.
Vue 3: reactivity and Composition API pitfalls
ref vs reactive, destructuring traps, watchEffect loops, Teleport context — common Vue 3 mistakes from Dev.to.
Beyond the hype: how top engineering teams actually use AI
Platform9, Monday.com, and PlayStation on restructuring the SDLC around AI, orchestration engineers, and review bottlenecks.
DeepSeek V4 Pro beats GPT-5.5 Pro on precision
Head-to-head on schema and instruction tasks: DeepSeek 38–33, one regex vs split patterns, fewer unsolicited extras.
The smallest brain you can build: a perceptron in Python
One input, weight and bias, training epochs, and a linear decision boundary — neural network fundamentals without frameworks.
Five products in one FastAPI monolith: HTMX, Telegram Mini App, and Stars
getdoday.ru: ~76k lines without React, themeParams and bot DNS pitfalls, Telegram Stars billing with entitlements.
React 19 Form Actions with FastAPI: server validation without duplication
useActionState + Pydantic: one schema, field-level errors, built-in isPending — roughly half the form code.
TypeScript path aliases in monorepos: Astro, React, and shared types
baseUrl, paths, and project references: @shared/types instead of ../../../ and silent build failures without workspace:*.
Genkit in Go: building AI apps beyond a single-prompt demo
Google Genkit for Go—structured outputs, multi-step flows, tools, observability, and model swaps without rewriting your logic.
The frontend knowledge gap AI won't close for you
Rendering, networks, component architecture, and Core Web Vitals—what button tutorials skip and production demands.
Why BullMQ beat in-API file conversion
Sync conversion in HTTP vs a Redis queue: p95 dropped from 800ms to 3.8ms, throughput rose 4.8×—plus the complexity trade-off.
Database Optimization: Six Strategies That Cut Query Time by 60%
Profiling, indexes, ORM refactors, materialized views, connection pooling, and cache — a practical PostgreSQL and MySQL guide.
MamayLM v2.0: Ukrainian LLMs and the Free Mamay AI Chat
INSAIT released Gemma 3–based 12B and 27B models with Ukrainian support, multimodal features, and infrastructure hosted in Ukraine.
Three SQL Injection Patterns That Still Ship in Node.js
Concatenation, template literals, and innocent-looking sql variables — why reviews miss them and how eslint-plugin-pg catches them statically.
Node.js: Factory and Strategy for a cross-platform npm package
EADDRINUSE on Windows, macOS, and Linux without OS if/else chains. Strategy, Factory, TypeScript contracts, and shell-injection-safe runners.
DeepSeek V4 Flash vs GPT-4o: latency, cost, and hybrid routing
Dev.to benchmark: ~0.9s vs ~3.2s p99, order-of-magnitude API economics, and routing most traffic to a cheap model with premium fallback.
Post-mortem: how Firebase RTDB array coercion bricked a React frontend
Deleting one RTDB log entry returned an object instead of an array — white screen and items.map is not a function. Client-side normalization fixes it.
Sandbox security for agentic AI: Docker, microVMs, and CVE-2026-31431
40% of teams cite security as the top barrier to scaling agentic AI. Docker Sandboxes, Engine hardening, and governance for autonomous code.
JavaScript performance: 8 fixes that actually matter in 2026
Measure first, then optimize: debounce, lazy load, Web Workers, bundle diet, and a quick-wins checklist.
Why JSON is becoming a bottleneck for AI agents — and ULMEN
Million-token context windows still move JSON between planner, tools, and memory. ULMEN LLM claims 44% fewer tokens in the author's tests.
A PostgreSQL-backed job queue in Go: the Swig example
Jobs table, row locks, JSONB payloads, LISTEN/NOTIFY, and goroutine workers — no Redis if Postgres is already in your stack.
Near-instant Next.js: a practical App Router checklist
Server Components by default, less client JS, next/image, no CLS, caching, next/script, and measuring on throttled mobile—not your laptop.
react-native-qr-camera-pro: a native QR/barcode scanner for New Architecture
Swift/Kotlin, TurboModules and Fabric, BarcodeThrottler, and zero JS frame processing—a lean module instead of a heavy camera SDK.
Build an AI Data Analyst That Needs No SQL
Natural language to SQL via LLM, in-process DuckDB, a three-stage pipeline, and read-only guardrails for safe analytics.
SQLite Is All You Need for Durable Workflows
DBOS and Obelisk: local SQLite plus Litestream to S3 for workflow state; Postgres when you need HA and shared scale.
Fast Client-Side Puzzle Solvers with Next.js and Web Workers
Next.js static export, Web Workers for heavy algorithms, bitboards for Block Blast, trie for word search — $0 hosting.
Semantic layer between storage and BI: why AI analytics needs it
Metrics Layer / Headless BI as an executable contract for metrics—not a data catalog—and why LLMs should not invent SQL on their own.
Node.js and banning AI code in core: the 19,000-line debate
A huge model-written PR, CLAs, and maintainer burnout — not «anti-AI», but review process for critical infrastructure.
Open WebUI on Ubuntu: Docker Compose and HTTPS with Traefik
Your own ChatGPT-style UI for Ollama and OpenAI-compatible APIs — a step-by-step deploy with Let's Encrypt.
Validating environment variables in TypeScript
A schema for .env, CI checks, and generated types — catch PORT=abc and empty API_KEY before deploy.
React.lazy and the blank screen after deploy
Stale HTML in the tab, new bundle on the CDN — chunk missing. A global listener and cache-bust in ~15 lines.
Why transcription alone fails clinical AI in healthcare
Whisper + LLM summary breaks in real practice: healthcare needs clinical reasoning engines and EMR integration, not plain text.
Antigravity 2.0: a $1000 OS built by 93 AI agents in 12 hours
Google I/O Writing Challenge: multi-agent orchestration assembled an OS core with no hand-written code — and shifts the developer role.
Cross-document View Transitions: gotchas without a framework
Deprecated meta tags, @view-transition { navigation: auto }, a 4s timeout, and pageswap/pagereveal—what many tutorials skip.
Python vector databases: what breaks after the notebook
Chroma, Pinecone, Qdrant, Weaviate—not QPS charts, but packaging, API stability, and the path from dev to production.
JWT in Node.js: five mistakes that break your API
Base64 is not encryption, alg=none, weak secrets, and refresh without rotation — a pre-production checklist for JWT auth.
npm supply chain attacks: hardening installs at project level
ignore-scripts, npm ci, provenance, and lockfile review—layered defenses when node_modules holds thousands of transitive packages.
Docker as a safety net for AI-generated frontend code
Copilot/Cursor code compiles but breaks in the browser — why to run the app and Playwright/Cypress in one compose stack.
7 hidden security risks in Node.js applications
ReDoS, npm supply chain, prototype pollution, secrets in env, and NoSQL injection — a checklist teams often skip after a fast deploy.
10 facts about UX ROI: design is a business lever, not decoration
Speed, first impressions in ~50 ms, cost of fixes before launch — numbers from Smashing Magazine for talking to stakeholders.
The HTML `<install>` element: PWA install without JavaScript
Chrome and Edge are testing a native install button for web apps — simpler than manifest plus custom JS alone.
Bulletproof React: taming frontend chaos in production apps
Feature-based folders, four state types, and ESLint boundaries — a practical look at the Bulletproof React approach for production teams.
React stack 2026: a map of libraries by category
Vite, Next.js, Zustand, TanStack, shadcn/ui — what to pick in 2026 instead of a 2022 stack and why the defaults shifted.
Shrinking a 3 GB SQLite dictionary to ~10 MB with an FST
The tsk case: finite state transducers in Rust, shared Finnish suffixes, and why SQLite FTS was only a stepping stone.
Chrome Prompt API: run Gemini Nano locally from JavaScript
LanguageModel.create() with no server or API key—limits, ~4K context, and why a hosted LLM fallback is mandatory.
Svelte May 2026: TypeScript 6, remote functions, CLI plugins
SvelteKit 2.56—field.as() for forms, a breaking remote query refresh change, experimental community plugins, ThoughtWorks Radar.
Snyk vs Checkov 2.3 for Terraform 1.9: what the benchmark shows
Across 127 modules Checkov reported ~20% more misconfigurations and scanned faster—yet Snyk wins on PR integration simplicity.
MDN rebuilds its frontend with web components instead of a React wrapper
How MDN moved from a React SPA shell to Lit, server components, and interactive islands—and why content sites should care.
JavaScript backends: the full-stack tax
Node.js vs Go/Rust throughput and memory, npm risk, and when a single language stack still wins.
CSS 2026: what the browser can do instead of JavaScript
Frontend Focus roundup: Grid Lanes in Safari, anchor positioning, sticky headers in pure CSS—and where native features still lose to JS.
TypeScript 6.0, Next.js 16.2, and Node.js debates on AI-written code
JavaScript Weekly #778: strict by default, types=[], TS 7.0 prep; faster Next.js 16.2; Node.js nine CVEs and pushback on AI commits.
tRPC in 2026: full-stack TypeScript without the boilerplate
Shared types end-to-end, Zod validation, request batching — and when tRPC is the wrong tool.
From Next.js to Astro: when a blog does not need client-side React
A developer moved a portfolio from Next.js 15 to Astro 6: ~230 KB JS per post page became ~0 KB; Lighthouse 99–100.
TG Unblock: bypassing Telegram restrictions with Rust and WebSocket
Open-source desktop app wraps MTProto in HTTPS to web.telegram.org — no VPN, servers, or subscription. ~6 MB, one click.
Next.js deploys without 502: PM2 reload vs Kubernetes
Zero-downtime on a VPS: cluster mode, reload not restart, and why you should not delete .next mid-deploy.
A modern React stack: Vite, Tailwind, i18next, and TanStack Router
Step-by-step setup without dependency chaos—routing, styling, and i18n in one minimal but production-ready scaffold.
QA in 24 hours: an autonomous AI agent instead of manual regression
RankCaster AI built an autonomous tester with Claude Code, agent-browser, and read-only psql — regression dropped from 48 hours to 40 minutes.