Open Source.

The visual prelude to llm-from-scratch: building from a single vector all the way up to the attention mechanism, one geometric move at a time. Every page answers a single why: drag a slider, change a parameter, watch the geometry shift, then one line ties it straight back to LLMs. From vectors and matrices through SVD, softmax, and cross-entropy to self-attention and the Transformer block, in 35 interactive lessons across 9 parts, so you understand the math behind training, fine-tuning, and inference instead of just calling an API.

A from-scratch LLM course in Chinese: build a Transformer by hand, line by line, starting from a character-level bigram and working all the way up to reproducing GPT-2 124M, with interactive in-browser visualizations at every step. Aimed at readers who want to understand attention, training, and sampling by writing the code, not just calling an API.

The visual counterpart to the LLM courses, turned on algorithms themselves: drag a handle, watch a sorting pass or a graph traversal rearrange, and read off the geometric idea underneath. A living notebook in Chinese that grows one algorithm at a time, for readers who want to see why an algorithm works instead of memorizing its steps.

Treats IELTS not as an English exam but as a scoring system you can reverse-engineer: starting from the official band descriptors and working backwards to what each of Listening, Reading, Writing, and Speaking rewards. It covers score-allocation strategy, question-type maps, speaking and writing frameworks, and a diagnose → prescribe → retrain prep loop. Written in Chinese, for everyone heading out into the world.

A Chinese-English grammar reader for speakers of Chinese. The two languages do not share one system of tense, aspect, or state, so the book opens with that contrast and then walks sentence patterns, auxiliaries and tense, clauses, non-finite verbs, lexical chunks, and meeting language in 20 sections across 4 parts. Band descriptors and the prep loop are in ielts-whitepaper; this volume is the grammar underneath.

A guide to losing weight and keeping it off, built on one idea: body weight is a reading, behavior is the dial. Each number is tied to its source. The first section ships a calculator that takes current and target weight and returns where three plans settle (exercise only, diet only, both), using Hall et al.'s rule of about 24 kcal/day per kg and exercise compensation from 18% in short studies to 84% over about 80 weeks.

Audits whether a paper's references exist and, for claim citations, whether the cited paper supports the claim. The first run covered all 209,985 references in the ACL 2026 proceedings (4,459 papers): 2 were confirmed nonexistent (0.001%). The claim-support rate did not replicate: three runs of the same pipeline gave 0.95%, 5.66%, and 6.12%, traced to an unlogged judge model. The report publishes all three runs and treats the non-replication as the finding.

A curated field map of proactive AI assistance, organized around one question most agent work skips: should the agent act right now, unprompted? Each work is placed in a five-stage pipeline (signals, decision, action, memory, evaluation & self-improvement) by which part of the intervention trade-off it advances, with notes corrected from full-text reads rather than abstracts. Bilingual (EN / 中文), open to PRs.

A lightweight task-specification format for AI-assisted development. Instead of heavyweight specs or no spec at all, nano-spec captures just enough intent (goal, constraints, acceptance) in a tiny document the model and you can both hold in your head, so agents build the right thing on the first pass.

Scans your local Claude Code and Codex usage and renders a GitHub-style coding heatmap, both as an embeddable badge and a shareable HTML report, so you can see when and how much you actually code with AI assistants.

One-command TLS certificates for heterogeneous fleets. Auto-detects the nginx flavor (standard / LNMP / BT panel), gives you unified observability (doctor / status), and falls back across CAs (ZeroSSL → Let's Encrypt). Built for ops managing many servers that were never set up the same way twice.

• #2841Fixes figure URLs in the two Deep Research API notebooks so each image loads from ../../images/ at the repository root.

• #4336Renders nine East Asian DOCX numbering formats, including Chinese counting, Japanese counting, and circled digits, as list markers.

• #799Keeps an original Markdown list marker in AUTO mode when it contains any Unicode letter or digit, so markers such as 第九条 and ① stay in the export.

• #7278Writes certificate config to a temporary copy, checks it, and replaces the original only after the write succeeds, so a full disk leaves the conf intact.

• #5142Restores a paused nested agent-as-tool run against that tool’s own agent, so the saved agent references resolve to the agent they were serialized with.

• #539Measures and paints CJK text in the document’s eastAsia font (w:rFonts w:eastAsia and the theme’s East Asian faces).
• #540Wraps CJK text at the column width with UAX #14 ideographic breaks and basic kinsoku.

• #373Adds FIM Agent to the curated Next.js list.

• #196Adds FIM Agent to the curated agents list.

• #1800Removes a leftover sentence from the CPM-LM readme that said plato-mini does not support one-line prediction.

• #1675Fixes the VisualDL launch example so the command passes --logdir.