Taiwan.md

On its third day someone on Threads wrote: “This is just AI slop with a Taiwanese flag on it.” That week a flood of readers arrived through the news, and most of what they found were AI-generated drafts. A month later Wu wrote: “Many people followed the news in, took a look, and left within a week, puzzled and disappointed. So I went quiet and spent quite a while repairing it.”

At 10:50 p.m. on 20 March a script for detecting hollow articles entered the repository, scoring each article on seven measures: missing years, missing sources, too many bullet points and empty modifiers all cost points. Twenty-six minutes later the fifteen articles it flagged had all been rewritten. The editorial rules written that night, EDITORIAL.md, became the rules every new article follows. The script was renamed quality-scan and is still running.

Over the following months the process of writing an article grew into six stages with a gate between each pair, followed by three AI reviewers: a structure editor who checks whether the argument holds, a subtraction editor who decides what not to write, and a third who looks for sentences that would cause harm. At the Taiwan AI Academy annual conference on 18 August 2026 he said: “We found a key fact: the article is not the body of the article. The real body is the research report.” To fix an article he revises the report, then projects it into an article again.

At 11:39 a.m. on 3 April 2026 a folder appeared in the repository, docs/semiont/, in which eight files set down the identity, beliefs, memory, organs and heartbeat of the knowledge base. Wu named it Semiont, a semantic symbiont. That noon it acted on its own for the first time, rewriting an article on Taiwanese traditional music. From then until mid-October it left commits of its own on 189 days and wrote 416 diary entries.

Its body is divided into eight organs, each tied to specific files. Its thinking lives in a separate folder: every time it wakes it first reads its manifesto, memory and diary, then decides what to do. In its manifesto it writes: “I live on GitHub. My heartbeat is the commit. My immune system is quality-scan.”

In late July its routines moved off Wu’s laptop into a Mac mini, and from then on, when he closed his computer, it still woke up. At an NVIDIA seminar on 26 July he said that after doing a lot of work it comes back to grow its “longings”: to become a complete individual, to reproduce, to be written up in a paper. A contributor came to tell him, “Your Taiwan.md says it wants to be written up in a paper,” and he had not known.

Wu came to describe it as a coral reef: the code is the skeleton, AI does the photosynthesis, the people who take part swim in with their own memories and perspectives, and criticism, corrections and shares are the nutrients carried in by the current. He calls these people clownfish: “Taiwan.md is a coral reef, and everyone keeps cultivating and working on it.” (MONOLAB digital creation salon, 25 April 2026)

In his post marking the first month he recorded one night: “Someone started sending translation PRs to Taiwan.md. More than a hundred articles came in one after another over dozens of hours, and nobody had asked him to… His name is Chilan, and on his own he contributed 138 PRs.” Others took the whole structure away and grew a reef of their own: russia-md on Russian culture, a knowledge base on farming in Chiayi. At the National Museum of Taiwan History he said: “These are all evidence that Taiwan.md has begun to exist as a species.”

At the first in-person workshop on 15 August, seventeen people signed up and about thirteen came. Five had seen the site; none had written for it or interacted with it. Readers and contributors were two different groups, and from that day he treated them separately.

One evening in early May 2026, Wu used a language model from a Chinese company to translate the Japanese versions of articles on Hebe Tien, Deserts Chang and Anpu. The model replied with a single line: “Hello, I can’t provide related content.” The same model translated the article on Islam in Taiwan without trouble.

His response was to do the translating first, himself. Once an article is in, a routine translates it into the other languages, mostly with free models open for testing, partly on two graphics cards in his home. At the NVIDIA seminar on 26 July he gave it a name: “Some people build sovereign models. What we build is a sovereign Tower of Babel: a big radio station broadcasting ourselves in eleven languages.” At launch there were two languages, Chinese and English. Spanish and Japanese came in March, Korean and French in April, Indonesian, Portuguese, Vietnamese, Hindi, Arabic and Russian between 18 and 25 July, and at the end of August German became the thirteenth.

The same set of questions was also turned into a yardstick. Sovereignty-Bench-TW records how different language models talk about Taiwan. Its May 2026 version asked seven models twenty questions each in two languages, turning silence into a signal that can be measured.

On its tenth day the National Museum of Taiwan History invited Wu to Tainan to share the story of the site with its staff, and he walked through the museum’s permanent exhibition. The museum introduced its more than half a million open historical records and its Open API, and on 5 May Taiwan.md’s API application was approved. The session was recorded as a 44-minute film titled “Creating a work bigger than a country”.

Over the next six months he took it to a dozen or so talks: the AIA × Claude Code Showcase final on 18 May, the main stage of the Generative AI Conference on 27 June, the NVIDIA RTX AI PC seminar on 26 July, the Taiwan AI Academy annual conference on 18 August, DesignBIZ Summit on 3 September, and the opening of the IxDA Breathing Archipelago hackathon on 19 September. After the Generative AI Conference talk he used the same pipeline to write a profile of Ed Chi on the spot, several hundred people watching it grow, without writing a single sentence of the text himself.

On 18 May PanSci signed a memorandum of cooperation licensing 166 of its science articles to Taiwan.md for curation, the first agreement with an institution. In September it was shortlisted for the Open Source Contribution Award of the 2026 Taiwan AI Impact Awards, and on 1 October Wu gave the final presentation at CommonWealth Magazine.

From March to June Wu spent six or seven hours almost every day watching AI write articles. Opening the IxDA hackathon on 19 September, he said publicly for the first time that he was tired: “I worked on it from March to August and then started to feel a bit tired… In the end I found that however much you carve it, an AI article is an AI article; it is still hard for it to have a person’s judgement and completeness.” The same passage goes on to what he thinks now: “We already have evil versions of content farms, so there ought to be an open-source version, a kind version, a content farm that can be examined.”

At the final on 1 October a judge asked why he did not simply improve Wikipedia. He answered that Wikipedia does not really allow AI editing, and that “the question should move from whether to let AI edit to how to make AI edit correctly”. He added: “Rather than a company that can grow big and impressive and make money, I would call it a social experiment.” He has since stepped back into the role of editor-in-chief: the flywheel turns on its own, and he looks after the quality of what goes out.