🛋️ lazeword (dsh-lazeword) — 躺着背单词 中文 lazeword is a word-anchored bilingual learning system — learning as simulation, trajectory as log (theory in docs/learning-as-simulation.md). It grew from a real family need: the author is an engineer with a rather wide range of experience — trained in Tsinghua's humanities program, worked on Chinese classics for nearly seven years, then moved to speech and autonomous-driving simulation (at Baidu Silicon Valley AI Lab he co-authored Deep Speech 2 with Dario Amodei and worked on Apollo simulation), then seven years at Weichai on heavy trucks and diesel engines. Not much to brag about — just sharing these experiences as a bit of material, for a child starting school in Hong Kong: from classics, to speech, to simulation, to manufacturing, to AI education, the path itself is part of the product (see "People & Words" and docs/ai-governance.md). Education is the key to how civilizations pass things on. While working on space AI, he made this for his daughter on the side — a bit of a willful experiment. But perhaps many fathers will like it too. - Open foundation: ships as a DeepSeek Harness (dsh, everything is a plugin) plugin; AI stories/explanations powered by DeepSeek (optional Cloudflare Worker keeps the key server-side; fully offline-capable) - Answers three questions of AI-era education: what to learn — words are the interface to knowledge (15,000+ entries: official EDB math/science/geography + Oxford 5000 + Chinese classics + engineers' words + AI literacy); how to remember — FSRS-5 deterministic scheduling (Anki's parameter family, replayable trajectory); what words are FOR — talking and creating with AI: "The hottest new programming language is English" (Karpathy); the AI literacy pack (prompt / context window / describe / verify… plus papers, people and rules) teaches kids to communicate with AI — for good, and verifiably - Core engine = scenario-driven + data-driven (modeled on autonomous-driving simulation): word = actor, sentence = scenario, trajectory = log, movie = scenario sequence — learning scenarios auto-generated from the child's own inputs and habits (see docs/scenario-engine.md); dsh's space-time determinism and the determinism required by AV simulation standards are the same thing - One file, zero dependencies: a dsh web plugin that also runs standalone (a single HTML file); five games + AI tutor; biliteracy & trilingualism; one-click 繁/简 - Adults and children are classmates in front of AI: this is not a top-down education product — the author learns and experiments together with his daughter (the AI-literacy pack's "verifiable doubt" applies to adults too). We don't know AI's "genetics" either — honest exploration here, no authoritative answers ## 30-second tour 🚀 Try it now (no install): https://xczhanjun.github.io/lazeword/ (GitHub Pages, single-file build; progress stays in your browser) 1. Learn 5 words — each card shows phonetics, morpheme breakdown, and the AI's real tokenization — see how an LLM reads words 2. Open "🎓 Tutor" — one lesson = diagnosis → lecture → practice → grading → written back into the memory schedule 3. Open "🧹 Chores" — pick a chore, write one English sentence, get praised by the AI teacher Learn card (token view) AI tutor Word racing (AV-simulation projection) dsh sidebar ## The name lazeword (躺着背单词) stays — the name is the answer: - Laze is not laziness — it's an attitude toward AI: comfort and trust. Lazy mode isn't about not studying; it's about letting learning happen in the most relaxed posture — good learning doesn't run on willpower, it runs on a system that remembers for you. - Word and token are the same thing: token in Chinese is 词元 ("word-unit") — the engineering form of a word. A child memorizes words; put those words into an AI and they become tokens. English has become a programming language ("The hottest new programming language is English"), and vocabulary is its syntax and lexicon — memorizing words is accumulating tokens for creating with AI. ## Features - 14,968 curated words (two modes) — basic mode by default: ~1,141 core words (947 everyday + HK subject/campus terms, fastest loading); opt into advanced mode for the full set: Oxford 5000 (KET/PET/FCE) + EDB math/science/geography + autodrive + Chinese classics (English ↔ 中文/繁體) - Spaced repetition (FSRS-5) — adaptive scheduling in Anki's parameter family (state = fold(events)) + wrong answers auto-queued + event-sourced trajectory (deterministic, replayable) - 6 quiz types — meaning, word, IPA, listening, spelling (letter grid), cloze; wrong answers can be retried, correct answers auto-pronounce and auto-advance, combo-streak animation - Pronunciation — English TTS + Cantonese + Putonghua read-aloud, IPA syllable & stress highlighting that follows the speech, repeat-after-me scoring - Word details — example sentences, synonyms, etymology, root/affix hints (offline) + AI explanation (DeepSeek, optional) - AI articles — 9 bilingual articles written from the word list (highlighted, clickable words) + AI turns today's learned/wrong words into a story - Reference — 111 irregular verbs, 24 grammar points, 43 phrasal verbs, 50 conversational sentences - Five games — space word-matching + memory flip + letter tracing (see the Chinese meaning, drag-connect letters on the grid to spell the word) + word racing (pseudo-3D road, steer into the right lane) + word minesweeper (classic rules + defuse mode: mines are your mistake-book words — answer their meaning to defuse them) - Family features — up to 4 learner profiles, parent/teacher report, daily 10 words, JSON backup, study streaks, study heatmap - Lazy mode — dark comfortable theme, big fonts, auto-pronounce + auto page-turn, voice control ("会 / 不会"), Space to pause - 繁/简 switch — one-click Traditional/Simplified for the whole UI - Anki ecosystem — one-click TSV export + AnkiConnect sync + review-log import (one deterministic timeline across platforms) - Dictionary lookup — any English word, even outside the library (phonetics/definitions/examples/synonyms/audio) - Shareable URLs — the URL is the state: tabs, filters and reference pages update the address bar live, copy to share; deep links: ?user=anna&tab=quiz&scene=math&word=integer&ref=ai-chat&advanced=1 - Pronunciation settings — Chinese entries default to Putonghua; optional Putonghua/Cantonese read-along on page turns; adjustable flip speed (slow by default) - Biliteracy & trilingualism — English/Cantonese/Putonghua read-aloud, one-click 繁/简, Chinese classics pack (三字經/唐詩/論語) - 🧑‍💻 People & Words — learn vocabulary from legendary engineers' real code & docs (21 people in two groups: Engineering Legends / AI Scientists — antirez, Linus, Bellard, Knuth, Hamilton, Hopper, Turing, Hinton, Fei-Fei Li, Kaiming He, Andrew Ng… verified quotes, GitHub links), and meet the people behind them - 🛡️ AI literacy & safety — 64 AI-literacy terms + 12 milestone papers + a kid & parent AI handbook + "Chat with AI" practice (four-element prompt teaching, 5 templates — describe / story with my learned words / ask why / verify a claim / fix my sentence, library words auto-highlighted in replies) - 🏭 Industrial manufacturing — vehicle/machinery/factory terms (truck, diesel engine, transmission, piston, welding, assembly line…) - 🎓 AI tutor — one lesson = diagnosis (trajectory picks weak words) → lecture (AI, three-tier fallback) → practice (deterministic questions) → grading (AI grades your sentence) → done (written back into FSRS, schedule adapts); English + math, reproducible questions per child/day, degrades to a fully deterministic lesson with no AI - 🛒 Ecosystem picks — curated dsh plugins & skills in Settings (offline voice input, desktop pet, themes, memory stack, token panel) with one-click install commands and a third-party-code safety note - Fully offline — one HTML file, zero dependencies (online only for example sentences / dictionary audio / AI features; optional Cloudflare Worker backend keeps the AI key server-side) ## Screenshots Running inside DeepSeek Harness (real installation test): lazeword in dsh sidebar The panel opens the full app (1094 words, fully offline): lazeword panel in dsh ## Why this design (theory) - The brain is a prediction machine: learning = updating a predictive model (Rao & Ballard 1999; Friston 2010; Clark 2013) - Words are knowledge anchors: 2,000 word families cover 87.8% of fiction / 89.4% of spoken English (Nation 2006); frequency follows a power law (Zipf 1949) - Forgetting is computable: Ebbinghaus 1885 → FSRS (Ye et al., KDD 2022); retention R(t,S) is a deterministic prediction - Trajectory = simulation log: event sourcing (Fowler 2005) + reproducible research (Buckheit & Donoho 1995) - Full argument, references and product decisions: docs/learning-as-simulation.md; industry research: docs/research-physics-simulation.md; core engine (scenario-driven + data-driven): docs/scenario-engine.md; AI governance position & open-source history: docs/ai-governance.md; roadmap: docs/roadmap.md; attributions & licenses: ATTRIBUTIONS.md ## Core design: dsh ecosystem & learning-as-simulation lazeword shares dsh's philosophy — everything is a plugin, everything is deterministic in space & time: | dsh concept | lazeword counterpart | |---|---| | Plugin mechanism | subject packs (math ships quiz-type code; geography/science/culture/autodrive are word packs), statically composed at build time, zero runtime loading | | Event sourcing | learning trajectory (append-only log, 20k cap + deterministic compaction); state = fold(events), replayable bit-for-bit | | Determinism | FSRS-5 scheduler (Anki's parameter family): retention R(t,S) is a deterministic prediction; seeded shuffles; golden-vector tests | | Host interop | /api/progress/:user sync at boot, sharing one progress record with the dsh AI teacher | Anki ecosystem: one-click TSV export (any Anki version) → AnkiConnect sync into a deck → review-log import (Anki reviews join the same timeline) → FSRS shares Anki's parameters and algorithm. Whether reviewing in lazeword or Anki, it is one deterministic learning trajectory. Theory (predictive processing, Nation's coverage data, the FSRS paper) and product decisions: docs/learning-as-simulation.md; industry research: docs/research-physics-simulation.md. ## Install on DSH Desktop (recommended for families) DeepSeek Harness Desktop (macOS / Windows, no Node required) wraps dsh into a native desktop app. Two steps: 1. Download and launch DSH Desktop (dshdesktop.cn / deepseekdesktop.com) 2. Tray → Open DSH Terminal → run: sh dsh plugin add dsh-lazeword # once published to npm; or install from a local path: dsh plugin add /path/to/lazeword Then restart DSH Desktop (so the new bundle enters the Loader composition) — the 🛋️ lazeword button appears in the sidebar. The desktop exposes desktopProfiles / desktopPnpm plugin services; lazeword is a regular dsh plugin and works out of the box in compatibility mode. ## Install as a dsh plugin sh dsh plugin --profile web add dsh-lazeword Then look for 🛋️ 躺着背单词 in the sidebar. The panel opens the standalone app inside the harness (blob-URL iframe — no server, no network, deterministic). ## Run standalone Open app/lazeword.html directly in any browser (double-click). standalone app Or serve it for phones on the same LAN: sh python3 -m http.server 8000 # open http://<your-ip>:8000/app/lazeword.html ## Cloud deployment (optional) A Cloudflare Worker backend ships in the repo (worker/index.js): hosts the DeepSeek key server-side, AI-story endpoint, and a dictionary proxy (rate-limited, input-sanitized). bash npx wrangler secret put DEEPSEEK_API_KEY # paste your DeepSeek key npx wrangler deploy # deploys to workers.dev; bind a custom domain via wrangler routes # static site: enable [site] in wrangler.toml pointing at ./app to serve on the same origin Serve app/lazeword.html from the same origin (or via the dsh host) to unlock AI stories and dictionary lookup. ## Develop Requires Node ≥ 22.19. Zero npm dependencies (tests use the built-in node:test). sh npm test # unit tests (core functions) node scripts/build.mjs # build the standalone app node scripts/build-client.mjs # embed the app into the dsh client bundle npm run check # syntax-check plugin entrypoints ## Architecture src/core.mjs deterministic pure core (single source of truth): SRS scheduling, IPA syllable/stress parsing, orthographic syllabification, grading, seeded RNG, event log (append-only learning trajectory) app/template.html the app UI shell (__CORE__ / __WORDS__ markers) data/*.json word lists (Vocabineer 947 + HK subjects) lib/index.js dsh host entry (plugin registration) lib/client.js dsh web client bundle (embeds the app; generated) scripts/ build + extraction tooling tests/ node:test unit tests (20 assertions, zero deps) Design principles: deterministic, offline, auditable — the app is statically composed at build time (no runtime plugin loading), and the core event log makes the whole learning trajectory reproducible ("spatiotemporal determinism"). ## License MIT ## Disclaimer DeepSeek Harness is in developer preview and may introduce breaking changes. This plugin is community-maintained and not affiliated with DeepSeek.