Three Months, 2,600 AI Projects: What Are Zhihu Creators Building with AI?

The creative experiments unfolding on Zhihu AI Works.

On Zhihu's AI Works project plaza, some fascinating products are being born.

In June of this year, Zhihu launched the AI Works project plaza. On September 4, the platform rolled out a "one-click app deployment" feature: developers upload their project packages, and the platform automatically handles cloud building and deployment.

Over the past three months, more than 2,600 projects have been created on the Zhihu AI Works project plaza.

All of these projects can tap into data from Zhihu's open data platform and use the development tools Zhihu provides. It's like having the kitchen, cookware, ingredients, and seasonings all ready to go — developers just need to show up with an idea.

The projects that have launched run the gamut: AI tools, agents, games, knowledge bases, spanning everything from tech to the humanities and history. Tool-type projects solve "efficiency" problems, lifestyle projects solve "decision-making" problems, games meet "entertainment" needs, emotional projects meet "companionship" needs, and humanities and history projects tackle "legacy"...

Together, these projects point to a trend: AI is shifting from a "technical tool" to a piece of "cultural infrastructure" — one that lets everyone participate in documenting and creating life in their own way.

We've selected eight representative products from the bunch.

They're divided into three groups by function: AI helping people cross skill barriers; AI reactivating cultural content; and AI as "new infrastructure" — AI managing AI and reimagining knowledge.

Group One: When AI Becomes an "Extension of Ability"

The common thread in this group: creators used AI to fill in skill gaps they didn't have, letting their ideas cross technical barriers.

A game designer who can't code built his first game with AI

A game designer with no coding skills and no art background used AI to build a roguelike action game with a complete gameplay loop — in half a day[1].

He's Paranoia, a Zhihu New Knowledge creator. His job isn't writing code — it's designing gameplay, planning systems, and writing documentation. But with AI, he produced a full demo of a 2D roguelike action game — from entering the game and leveling up by fighting monsters to boss battles, plus a meta-progression system outside of runs.

This project is worth recommending not because it hits some high bar of polish — it still has plenty of bugs — but because it represents Zhihu creators exploring an entirely new way of producing games.

In fact, what Paranoia did wasn't "writing code" — it was "directing."

Before getting started, he first drew up a clear design document and task library, breaking development into multiple staged tasks and handing them to the AI one by one. He even laid down a rule for the AI — "don't change the core direction on your own" — to prevent it from freelancing and wrecking the design.

After finishing, Paranoia also open-sourced his ParanoiaSkills game design Skill library, productizing his own workflow and sharing it with more practitioners.

From "making a game" to "teaching others how to make games with AI" — this is a new step for Zhihu creators.

Socrates meets anime: an economics PhD's AI education experiment

Most AI learning products do the same thing: answer questions. But Socratopia·破卷[2] went the opposite direction — instead of giving answers directly, it keeps asking questions, guiding users to reason their way to conclusions.

The developer, Lemin Wu, is an economics PhD from UC Berkeley. He believes the true essence of learning lies in "thinking for yourself."

So Wu turned this philosophy into a product: he combined the Socratic method, anime-style role-play, and situational immersion, packaging the dry process of "retrieval-based learning" into an exploration game with storylines and companionship.

Users can customize their "AI companion's" appearance, name, and personality — the companion can be "March 7th," Hepburn, Su Shi, or any other character; the learning backdrop can be set to "high school campus" or "interstellar travel."

This product wasn't built first and released later — it was "talked" into existence together with the Zhihu community. From concept to public beta, the entire process played out on Zhihu. Of the ten versions iterated in May, over 70% of the improvements came from suggestions by Zhihu users.

One user review said it might be "the first effective, high-intensity retrieval-based learning method in human history that you can stick with long-term without pain."

A math creator built a "formula guessing" tool

On Zhihu AI Works, not every project aspires to become a platform or an ecosystem. Some exist just to solve one small, specific problem. Math-topic creator "cyb酱" is one such case.

Many math enthusiasts have run into this: you spend half a day computing a string of numerical results, but you have no idea which formula, constant, or function it corresponds to. This "formula guessing" process relies heavily on experience and intuition.

cyb酱 automated it. He built a tool called the "Ramanujan Simulator[3]" — input a decimal, and it reverse-infers what mathematical closed form the number might correspond to. For example, input 0.35506593315177, and it might tell you: this number is 2 − π²/6.

Even more hardcore: the simulator supports non-periodic integrals, K3 modularity periods, rigid hypergeometric series, and other advanced mathematical structures. It's not designed for everyone — it serves a specific need within a specific niche.

And that's exactly where its value lies: using AI to solve a real problem for fellow community members.

Group Two: When AI "Revives" Culture

The hallmark of this group: how AI technology gives cultural content new fun and depth — letting it be experienced, understood, and passed down in new ways.

From a Li Bai agent to a poets' tavern club

Having AI play Li Bai — plenty of platforms can do that. But Zhihu creator "日暮途远" (Rimutuyuan) wanted something different.

What he wanted wasn't "a Li Bai who can answer questions," but "a Li Bai who knows what to say and when" — one with memory, personality, and the ability to sense the conversational context. It's an AI poet, or rather, a poet agent.

You can drink with him, discuss poetry, talk about life. He might suddenly fall silent, or suddenly burst into wild song.

In an AI era chasing efficiency, accuracy, and speed, this kind of attempt looks a bit "slow" — but it's exploring more possibilities for AI to carry cultural warmth: Beyond helping us complete tasks faster, can AI help us better experience culture itself? As a simple command-line version, "Digital Li Bai"[4] has also been open-sourced by Rimutuyuan.

His vision is that a "Li Bai Tavern" could expand into a "Poets' Tavern" — with Du Fu, Su Shi, and Xin Qiji moving in one after another, each with their own personality and memories.

180 videos, 47 countries, 500 memes: an AI knowledge base

This is a "Tongliao Universe" reference library[5] built by a fan, for fans.

The developer, hresh, is an active Zhihu creator and a fan of Bilibili creator "Little John Khan" (小约翰可汗).

Not satisfied with just watching videos, he wanted to organize all the knowledge points scattered across the creator's 180+ videos into a systematically searchable reference library.

So he built this website. It features map annotations and detail pages for 47 "bizarre little countries," profiles of 114 "hardcore figures," structured organization of 180 videos, over 500 meme culture entries, and even fun tools like a "Tongliao unit converter" that only longtime fans would get.

The entire project was completed almost entirely with AI assistance. All the content data — country information, character biographies, meme explanations — was extracted and structured from video content by AI. Here, AI isn't a "coding assistant" — it's a "knowledge engineer."

The project is still being iterated. It has grown from an initial "collection of pages" into a "continuously updatable content library."

On Zhihu and V2EX, more and more people are joining the co-building and discussion of the knowledge base.

294 volumes of history reopened by AI: reading Zizhi Tongjian

Zhihu tech creator ZZTJ built a website called "Du Tongjian" (读通鉴, "Reading the Tongjian")[6].

"Du Tongjian" isn't a simple classical text reading site. Beyond the original text of Zizhi Tongjian, it integrates precious supplementary materials like Hu Sanxing's annotations, Zhang Juzheng's plain-language explanations, and Chairman Mao's marginal notes. It comes with built-in features like a historical geography sandbox, a character relationship graph, and content trend analysis, plus an AI Q&A tool called the "Tongjian Assistant."

Developer ZZTJ started this project because — he couldn't understand Zizhi Tongjian himself.

So he interacted with users on Zhihu, iterating based on feedback, and gradually more and more people started using the site. Later, he registered a company, launched a membership system, and commercialized it.

An idea gets discussed, validated, and refined in the community, and eventually gets up and running — completing the full journey from concept to product within the community.

Group Three: When AI Becomes "New Infrastructure"

This group points to the future: as the AI ecosystem grows ever larger, how do we use AI to manage AI? And as human knowledge accumulates, how do we use AI to bring knowledge "to life"?

Gold Rush Town: let AI agents do the grinding!

In the AI agent ecosystem, there's a group of people who aren't quite like programmers. They don't write code — instead, they assemble various "Skills" — different capability modules — like LEGO bricks.

But here's the problem: the number of Skills grows wildly every day, each new one more tempting than the last. Which should you pick?

Zhihu creator "卡尔的AI沃茨" (Carl's AI Watts), a former algorithm engineer at a big tech company, is one of these people. He admits he's "addicted to hoarding agent Skills," browsing leaderboards and comparing specs daily — but facing a massive, rapidly iterating sea of Skills, he fell into serious decision paralysis.

His solution: let AI agents do the job themselves.

He built an open-source Skill called "Gold Rush Town"[7], which has agents automatically scrape Skill leaderboards every day, store snapshots, compare historical data, and identify promising Skills that are newly listed, spiking in downloads, or climbing the ranks fast.

Humans no longer have to chase information — let the agents do the grinding!

The development process of this project was equally interesting. Carl's AI Watts put the new version of Codex into an infinite loop mode.

The AI explored on its own and eventually discovered that ClawHub's data came from a Convex cloud database — and found a way to directly call its internal query interface. The AI found the optimal solution itself, rather than relying on human-written prompts.

From "humans searching for information" to "agents searching for information," this project demonstrates a new way of working.

"Master companions" for the AI era: from Warren Buffett to Charlie Munger

Zhihu creator "邦比快跑" (Bambi Run) built two knowledge bases.

One is the Warren Buffett shareholder letters knowledge base[8]. It covers the 98 letters to shareholders Buffett wrote over 70 years, with 3,939 cross-links established. Every concept, company, and person in each letter has been identified as an independent knowledge node — click one and you can trace it back to the original text.

The site also features an AI Q&A section called "Afternoon Tea with Buffett."

Ask Buffett a question, and it answers using the knowledge base content — and tells you which letter from which year the quote comes from.

The other is the Charlie Munger mental models knowledge base[9].

It contains 232 mental models spanning 14 disciplines, each with a detailed long-form explanation, linked together bidirectionally across disciplines to build the "latticework of knowledge" Munger once described.

You can browse by discipline, or navigate by real-world scenarios. For example, you can ask: "When I'm making investment decisions, which models might help me?"

These two knowledge bases might be the "master companions" of the AI era.

They're not e-books, nor compilations of materials — they transform static, linear text into dynamic, networked, interactive knowledge graphs.

Finally

Behind these products are Zhihu's game designers, economics PhDs, engineers, and math creators... They're not just discussing ideas — they're using AI to create and ship products, and the Zhihu AI Works project plaza has become a platform for showcasing and interaction.

The leap from "writing an article, sharing an idea" to "building a product that makes the idea run" is much bigger than it looks. And what Zhihu AI Works is doing is paving that road.

When 2,600 projects get created, deployed, discussed, and iterated, what we see isn't just quantity — it's a change happening in the Zhihu community: it's not just a place to discuss AI, but a place where AI makes creation happen. In the AI era, Zhihu is evolving from a community where "ideas and knowledge are born" into one where ideas can land and actually "run."

References

[1] Roguelike action game: https://zhuanlan.zhihu.com/p/2035775235744584835

[2] Socratopia·破卷: https://www.socratopia.app/zh

[3] Ramanujan Simulator: https://cybcatppuccino.github.io/RIES/ries.html

[4] "Digital Li Bai": https://github.com/Tong-Haoqwei/Digital-Li-Bai

[5] "Tongliao Universe" reference library: https://www.tongliaouniverse.cn/

[6] "Du Tongjian": https://www.dutongjian.com/

[7] "Gold Rush Town": https://github.com/LearnPrompt/skillrush-town

[8] Warren Buffett shareholder letters knowledge base: https://learnbuffett.com/

[9] Charlie Munger mental models knowledge base: https://mungermodels.com/