The Highest Information Density in the Chinese AI Community Isn't on Twitter

linuxdo is likely the best large-scale Chinese technical forum currently available. It uses an application-based entry system—you have to write a sincere essay of over 50 words to get in. A significant portion of the material for the long articles I write on Chinese Twitter comes from this forum. My domestic AI comparison piece cited real-world feedback from linuxdo users on DeepSeek and Kimi, and I first saw a complete comparison of Claude Code cost-saving solutions there.

This article explains what you can find on this forum and why this content only exists there.

Why this content isn't on Chinese Twitter

Three reasons.

First, the application system. New users must write a 50-word essay to enter, and AI-written ones don't count. This step filters out a large number of bot and advertising accounts. Members call each other "Lao You" (Old Friends), and most posters are developers, indie devs, or heavy users of AI tools.

Second, a ban on AI-generated content. This is the strictest anti-AI sentiment I've seen in a Chinese community—the posting rules are clear: posts obviously written by AI are deleted, and accounts are banned. Whatever you see on linuxdo, high quality or low, won't have a strong AI "smell."

Third, the forum format. A tweet on Chinese Twitter is 280 characters and sinks quickly. linuxdo uses the Discourse system, where a post can be updated for months, with hundreds of replies adding corrections. The most typical example is the "Cost-Saving Series"—the author wrote from part 1 to part 9, updating each one, with community members adding their own test data in the comments. This kind of accumulation is impossible on Twitter or Zhihu.

The linuxdo community once made a "Summary of AI Communities Across the Web." When listing Chinese AI forums, they only wrote three: linuxdo itself, Alibaba's ModelScope, and Baidu's AI Developer Community. Their evaluation of CSDN was "outdated technology and concepts," and for Zhihu, "occasionally has experts posting articles."

How to use Claude most cheaply—The Cost-Saving Series

One of the most viewed post types on linuxdo is cost-saving solutions for AI tools. The most representative is a serialized post called the "Cost-Saving Series," which has 9 parts so far, with a single post reaching 19,000 views and over 1,400 likes.

This series does something no one on Twitter does: it takes every conceivable channel for using Claude Code—carpooling, API relays, official subscriptions, Cursor, Antigravity—and creates a comparison table based on monthly budget tiers. Each tier calculates actual usable quota, cache rates, and target audience. From 100 to 600 RMB per month, it clarifies which plan to choose and why.

One insight I haven't seen elsewhere is about cache rates. Many people think API relays are cheap and buy a lot of tokens, only to run out quickly. It's not fake token counts—it's the difference in caching mechanisms that causes the actual usage time for the same amount of money to vary greatly across channels. This kind of analysis, which requires diving deep into billing mechanisms, isn't something tweeters spend time on, but forum posts can.

The forum also has an "Indie Developer Poor Man's Kit"—a community-maintained document listing all free or low-cost AI programming solutions. From using Zhipu's GLM 4.5 for free via 'zai', to using Gemini 2.5 Pro for free on Google AI Studio, to free GPT-4.1 quotas on GitHub Copilot—over a dozen platforms' free plans are in one table. This post has over 10,000 views, 591 likes, and is continuously updated as community members add new free options in the comments.

Claude Anti-Ban—From linuxdo Premiere to Twitter Reposts

Another type of high-value content on linuxdo is Claude anti-ban guides.

One "Core Guide to Claude Anti-Banning" received over 1,500 views and 146 likes on the forum before being reposted verbatim by Twitter users. The author, an operator of a Claude Code API relay, combined leaked Anthropic code with their own operational experience to write 12 anti-ban rules.

The most valuable part of this guide isn't the specific rules—you can see those in the original post—but the mental framework it proposes: Anthropic's risk control isn't a fixed set of rules, but a continuously running probability model judging "whether this account looks like a normal person." Understanding bans this way brings logic to seemingly random events. For example, why someone might be fine for three months and then suddenly get banned—it's likely their behavior pattern triggered an anomaly threshold at that point.

The post has hundreds of replies where users share their specific ban (or non-ban) situations—IP, subscription method, duration, and actions taken. This density of real cases is impossible on Twitter because no one describes their detailed Claude environment on public social media.

The Real World of Indie Developers

Most indie developer stories on Twitter are success narratives—how much they earn per month, growing from zero to many users. linuxdo is different; there are both success and failure posts, and the failures are more detailed.

One developer wrote "Quit my job for full-time indie dev for a year, how much did I earn?", with 7,300 views and 858 likes. He quit to build an AI financial report analysis tool, taking three months from research to launch. After a year, he earned less than $1,000. He didn't even make back the server costs.

In the post, he wrote the complete process of failure—the promotion methods tried, how much each cost, how many users they brought, and why they were stopped. His final conclusion: a person's energy is very limited, indie development cannot have obvious weaknesses, and he decided to stop full-time indie dev and return to the workplace.

Another post, "Continuously updating until I make it," by an author born in '96 with 2 million in debt, has 4,600 views and 463 likes. He listed every project he's done—Pinduoduo e-commerce for two months with 10k revenue but all spent on promotion, a self-built e-commerce mini-program banned for lacking medical device certificates, a UGC forum product that died naturally because the logic wasn't thought through, and a lukewarm local community life project.

The value of these posts isn't in teaching you what to do, but in letting you see real paths to failure—you don't see these on Twitter because no one wants to write about earning less than $1,000 on public social media.

Deep Technical Posts—Living Documents

There's a category of content on linuxdo with no equivalent on Twitter: deep technical posts updated over several months.

The "Large Model Series" serialization is a typical example. The author writes long reviews around models like GPT-5 and DeepSeek-V3.1, including Aider scores, Werewolf tests (testing the model's ability to lie and disguise), LiveBench scores, and FictionBench long-context scores. Every time new test results come out, the original post is updated, and the comments section adds other people's test data.

The "10x Speed Writing Series" did a full-platform ranking of AI voice input methods—4o-transcribe, whisper, gemini-2.5, Sogou, iFlytek—ranked into 6 tiers with specific testing methods and cost calculations. A single post has over 10,000 views.

There's also a "Comparison Study of Top 10 IDEs" post, comparing AI programming tools like Codex, Antigravity, and Cursor across multiple dimensions: Rules/Skills/Quota/Speed/Price. The author calls it "Comparative Learning—learning multiple software programs on the same topic at once."

The commonality of these posts is: the information volume is too large for a single tweet, and they are continuously updated. A post's lifecycle might be months or longer, with every update appended to the original, unlike Twitter where posts sink once published.

How to Use

linuxdo is application-based; you need to go to the registration page and write a sincere essay of over 50 words. Once passed, there's a trust level system—new users start at Level 0 and progress to Level 1 and 2 through reading and interaction. Different levels can see different sections and features.

For those who don't want to register, many posts on linuxdo can be viewed without logging in and can be found via search engines.

Sections worth following: Development Tuning (AI programming), Frontier News (AI news and model releases), Resource Collection (tools and open-source projects), Welfare & Perks (public APIs and free quotas), Document Co-construction (community-maintained tutorials).

A reminder: linuxdo also has a lot of fluff and casual chat. Like all forums, high information density doesn't mean every post is worth reading. Use search and tag filters—the "Featured Posts" (精华神帖) tag is quite useful; prioritize those.