Home on ZovixBC: Cutting-Edge Tech News & Insights

Sun Yat-sen University: New Paradigm for AI Talent Development

Sat, 16 May 2026 00:00:00 +0000

Introduction

Recently, the Ministry of Education and four other departments released the “AI + Education Action Plan,” which emphasizes the cultivation of AI talent and the deep integration of AI with education.

In the current wave of AI applications, how can new technologies serve education? How can university classrooms resonate with the cutting-edge industries in the Greater Bay Area? Recently, reporters visited Sun Yat-sen University’s campuses in Guangzhou, Zhuhai, and Shenzhen to explore the digital transformation practices of this century-old institution.

Enabling Personalized Education

In a study room at the Guangzhou campus, Lu Zijin, a 2025 undergraduate student from the School of Chemistry, is working on problems related to “chiral carbon configurations.” She uses the “AI class representative” to search for precise knowledge points in her textbook and receives additional related questions. Guided step by step, Lu Zijin gradually overcomes the difficulties in her exercises.

Such “personalized learning assistance” frequently occurs at Sun Yat-sen University. In January of this year, the university launched its self-developed “Yixian Smart Course” platform, which currently offers 102 smart courses. This platform analyzes students’ learning behaviors and quiz performances in real-time, accurately assessing each student’s mastery of knowledge points and intelligently matching personalized learning paths.

The AI platform has also transformed teaching methods. Professor Luo Xin from the School of Physics typically opens the “Yixian Smart Course” platform when preparing lessons. The learning data for the chapter on “Conservation of Angular Momentum” in the mechanics course is clear: 80% of students struggle with the knowledge point of “kinetic energy of rigid body rotation”—30% of them review related lecture videos after class, while five students have a quiz accuracy of less than 40%.

The next day in class, Luo Xin allocates some time to use a real case to help students break down the core logic and common pitfalls of the energy method. After class, the platform continues to assist: students with weak foundations receive reinforcement exercises, while those with more capacity are assigned advanced tasks.

Sun Yat-sen University President Gao Song stated, “Digitalization brings unprecedented opportunities for education. The integration of AI and education is not only a technological revolution but also a profound transformation in educational concepts and teaching models. In this era, we hope to empower higher education to achieve both scalability and personalization through AI.”

Building an Integrated AI Curriculum System

Recently, Wen Yucan, a 2025 undergraduate student from the Department of Chinese (Zhuhai), has been particularly busy. She and her classmates applied for the “AI + Traditional Opera Heritage” project, which has just been approved under the “College Student Innovation and Entrepreneurship Training Program.”

This initiative originated from a general elective course on artificial intelligence taken last semester. After the course, Wen Yucan realized the usefulness of AI’s capabilities in generating language, recognizing images, and processing text for cultural dissemination.

“We insist on building an AI curriculum system based on the ‘integration of general and specialized education’ model,” said Vice President Xie Shen of Sun Yat-sen University. “Core courses for computer-related majors focus on solidifying professional foundations; general AI courses for non-computer majors emphasize understanding underlying logic and theory; and human-machine collaboration courses for all students teach how to use AI in practical scenarios.”

Why establish a separate human-machine collaboration course? Chen Yingqian, one of the course designers for “Human-Machine Collaboration: AI + Medical Imaging,” believes this is more about imparting knowledge than teaching usage skills.

Chen Yingqian stated, “Students need to experience the role and limitations of current AI through numerous real clinical cases and think about how to better collaborate with AI. We hope to provoke deep thinking among students about how to adjust their learning methods and establish a foundational medical knowledge system to become adaptable medical talents in the future amid the AI wave.”

Training Talent for Future Industry Needs

In April 2025, the Ministry of Education and nine other departments jointly issued the “Opinions on Accelerating the Digitalization of Education,” proposing to optimize higher education discipline and professional settings in line with the development of the digital economy and future industries. In the digitally advanced Guangdong-Hong Kong-Macao Greater Bay Area, this guidance has translated into concrete actions—frontier industry demands are deeply influencing university discipline construction and embedding throughout the talent cultivation process.

Recently, Lin Min, a 2025 doctoral student from the School of Intelligent Engineering at Sun Yat-sen University, had his understanding of his major “refreshed” after participating in a joint project with a Huawei team.

Lin Min and his team spent months successfully training a “visual-language-action” large model capable of executing long sequence tasks in complex environments and built a robot that performed excellently in laboratory evaluations. However, for enterprises, the “action accuracy” in the lab is merely a passing mark.

“To meet commercial standards, we had to deeply integrate the model with hardware, ultimately enabling the robot to accurately perform tasks such as data center inspections,” Lin Min reflected. “This leap from ‘paper metrics’ to ‘stable operation in real scenarios’ is hard to experience in school.”

Tan Guang, Vice Dean of the School of Intelligent Engineering at Sun Yat-sen University, believes that Shenzhen, as a frontier of China’s technological innovation, gathers numerous global leading enterprises, providing fertile ground for the integration of industry and education in the field of ‘AI + robotics.’ The school collaborates with leading companies like Huawei to build joint practice bases to help students transition from ‘knowledge acquisition’ to ‘value creation.’

At the Guangzhou campus, the “Smart Learning Hall” innovative education practice platform, relying on the National Supercomputing Center in Guangzhou, systematically integrates the “Tianhe-2” supercomputing resources into teaching. At the Zhuhai campus, Sun Yat-sen University leverages its solid disciplinary foundation and data advantages in fields like atmospheric and ocean sciences to form a talent cultivation model of “specialized direction + artificial intelligence,” establishing dual bachelor’s degree programs such as “Atmospheric Science + Artificial Intelligence” and continuously delving into areas like marine AI and intelligent perception to serve the development of the marine economy in the Greater Bay Area.

From Shenzhen’s “industrial practice” to Guangzhou’s “computational foundation” and Zhuhai’s “specialized intersection,” Sun Yat-sen University has formed a complementary talent cultivation network for AI across its three campuses, nurturing future industry talents who understand industries, can innovate, and are capable of practical applications, injecting momentum into the development of new productive forces in the Greater Bay Area.

Beginner's Guide to AI Programming (Vibe Coding) in One Week

Thu, 14 May 2026 00:00:00 +0000

Introduction to Vibe Coding

Can you create usable software tools without knowing how to code? Two years ago, this would have sounded unbelievable, but today, it’s a reality thanks to “Vibe Coding.”

As a graduate in Chinese literature with eight years of experience in copywriting, I couldn’t even distinguish HTML tags. Yet, I’ve managed to create a resume generator, an image filter tool, and a resume screening tool—believe it or not!

What is Vibe Coding?

The term “Vibe Coding” was coined by Andrej Karpathy, co-founder of OpenAI. He tweeted that AI programming has evolved to a new stage where you no longer need to write code line by line; you just tell the AI what you want, and it generates the code for you.

In simple terms, Vibe Coding means “programming with natural language.” You don’t need to know Python, JavaScript, or HTML; you just need to express your ideas in plain language.

For example, if you say, “I want a resume generator that automatically creates a beautiful resume once the user fills in their information,” the AI understands and delivers the complete code within a minute.

This is a stark contrast to traditional programming, where you have to learn the machine’s language to communicate. With Vibe Coding, machines adapt to human language, making it much more accessible.

Benefits of Vibe Coding in Daily Life and Work

Initially, I was skeptical about Vibe Coding, especially as a pure humanities student. However, after trying it for a week, I was amazed at how it solved many practical problems for coding novices like me.

For instance, when screening resumes, I used to sift through dozens or even hundreds of applications manually, which was exhausting. After creating a resume screening tool with AiPy, I could simply upload the resumes and specify the candidate criteria—like “over three years of experience, background in the internet industry, and a bachelor’s degree.” The AI automatically filtered the candidates, significantly boosting my efficiency.

Additionally, I often needed to apply filters to images, which used to require time-consuming Photoshop work. Now, with AiPy, I can upload an image, select the desired filter effect, and achieve results instantly.

Creating resumes has also become easier. Previously, I spent hours formatting resumes in Word, but now, with AiPy, I can generate beautifully formatted resumes in minutes.

Vibe Coding is Simple: Just a Rough Idea and the Right AI

Many people feel overwhelmed by the term “programming” and believe they can’t learn it. I completely understand, as I once felt the same way.

However, Vibe Coding has an incredibly low barrier to entry. You don’t need to know about variables, loops, functions, or object-oriented programming. You just need a vague idea, like “I want to create a tool that generates resumes,” and the AI will break down your request and generate the code.

This process is akin to hiring a designer to create something for you; you simply express your needs, and the designer (in this case, AI) makes it happen. The AI is available 24/7 and doesn’t charge design fees.

When I started using it, I didn’t even know the difference between front-end and back-end. But I realized that it didn’t matter. I just had to tell the AI what I wanted, and it would deliver. For example, I said, “I want a webpage with an input box on the left and a preview on the right,” and the AI generated it for me.

AI Can Generate Code in Bulk and Quickly

One of the most impressive aspects of Vibe Coding is how quickly AI can generate code. Previously, I thought coding was a slow process, requiring meticulous line-by-line input and debugging. However, AI programming has completely changed my perception.

You can give the AI a requirement, and it can generate dozens or even hundreds of lines of code in seconds. If you’re not satisfied, you can specify what needs changing, and it will adjust immediately. The entire experience feels like conversing with an experienced programmer who is always patient and never annoyed by your simple questions.

For example, I created a resume generator, and from the initial idea to seeing the final product, it took less than five minutes. I made several style adjustments—“the font is too big, make it smaller,” “change the color to blue,” “tighten the layout”—and the AI made each change instantly, allowing me to see the results immediately.

This experience is incredibly satisfying. Previously, there was a huge technical gap between having an idea and executing it. Now, that gap is bridged by AI, making it just a sentence away.

AiPy: A Powerful Tool for Vibe Coding

When discussing Vibe Coding tools, there are options like Cursor and Copilot, but I found AiPy particularly useful in China. AiPy’s greatest advantage is its full support for Chinese, and it not only helps you generate code but also runs, debugs, and improves it. You don’t need to configure any environments or install complex software; you can use it directly from the web.

During my week with AiPy, I completed three small projects, each giving me a sense of accomplishment:

Resume Generator: Users input personal information, work experience, and educational background, and the AI automatically generates a beautifully formatted resume in under five minutes.
Image Filter Tool: Upload an image, choose the desired effect (like black and white or vintage), and the AI processes it automatically. While it may not match professional editing software, it’s sufficient for everyday use.
Resume Screening Tool: I could upload a batch of resumes and tell the AI what kind of candidates I was looking for, and it would automatically filter out suitable candidates. This tool saved me a lot of time, reducing what used to take half a day to just ten minutes.

If I had to learn programming to create these projects, it would have taken at least three months. But with AiPy, I completed them in just a week, spending only a couple of hours each day. Plus, if you’re worried about running out of tokens, you can use the invitation code c8W3 for two million tokens.

My Personal Experience and Suggestions

As a humanities student, I never imagined I would one day “write code.” However, Vibe Coding and AiPy have shown me that programming is no longer exclusive to programmers; it is becoming a skill that anyone can master.

You don’t need to understand technology or programming; you just need an idea and the right AI tool to bring it to life.

If you’re like me, a regular person who doesn’t understand code but wants to try creating your own small tools or applications, I highly recommend you try Vibe Coding. It won’t turn you into a programming expert overnight, but it will allow you to see your creativity materialize in the shortest time possible.

In conclusion, in this era, not knowing how to code is not scary; what’s truly frightening is not daring to try with the help of AI.

I started from scratch and created three small tools in a week. You should give it a shot too; you might do even better than I did. After all, what we lack is not ability, but a starting point.

Writing this article made me reflect on how I used to think programming was “someone else’s business,” something far removed from me. Looking back now, rather than viewing Vibe Coding as a new technology, it’s more about a new mindset—it teaches us that tools are meant to serve people, not the other way around.

If you’re hesitating about whether to give it a try, my advice is: don’t overthink it; just dive in. Find a small tool you want to create, open AiPy, and express your idea. Even if what you create initially is rough, that’s okay. You’ll be pleasantly surprised to discover that you can “create” something after all.

This sense of achievement is something you won’t experience from watching countless tutorials.

The Geopolitical Divide in AI: China's Open Approach vs. U.S. Restrictions

Wed, 13 May 2026 00:00:00 +0000

The Geopolitical Divide in AI

In 2026, the notion that technology knows no borders seems more like a cruel joke than a reality. Recently, the AI community has been shaken by strict access restrictions imposed by major AI platforms on users from mainland China.

Many developers eagerly opened Claude, only to be met with a frustrating identity verification request requiring original documents like passports or driver’s licenses, along with real-time facial recognition. Disturbingly, the system automatically rejects documents issued in mainland China.

Social media is flooded with complaints. Some users went to great lengths to find overseas friends to assist with verification, only to have their accounts banned overnight. Others, who had studied abroad for years, found themselves locked out simply because they registered with a domestic phone number.

In the latter half of 2025 alone, Anthropic reportedly banned over 1.45 million accounts suspected of being linked to Chinese users under the guise of compliance. Behind these numbers are countless Chinese AI developers and tech enthusiasts who have invested significant time and money.

Claude is not alone in this practice. OpenAI’s ChatGPT outright denies access to mainland IPs, while Google’s Gemini is completely unavailable in mainland China. On one side, U.S. AI companies maintain strict barriers against Chinese users, while on the other, Chinese AI models are eagerly welcoming global users, even seeking to enter capital markets. What geopolitical and economic logic lies behind this divide?

The Escalating Wave of Bans: Who is Closing the Door?

Let’s take a closer look at which foreign AI applications are blocking Chinese users.

Claude—The Strictest Ban
Since March of this year, numerous users have reported their Claude Code accounts being banned. By mid-April, identity verification requirements escalated, mandating users to present original documents—no photocopies or screenshots allowed, and real-time facial verification was required.

In September 2025, Anthropic updated its terms of service to prohibit companies with over 50% Chinese ownership from using Claude services, regardless of where they operate. This means that even if a company is registered in the Cayman Islands, it can still be banned if deemed to have Chinese ownership. Hong Kong and Macau are also included in this sweeping ban.

ChatGPT—A Softer Yet Firm Ban
Compared to Claude’s extreme measures, ChatGPT appears more lenient. OpenAI claims to be open to 161 countries and regions, but China has never been on that list. When the system detects access from mainland China, it simply returns a message stating that the service is not supported in that region. However, Hong Kong and Macau are still within the service range.

Google Gemini—Total Ban
Google’s Gemini is also unavailable in mainland China, Hong Kong, and Macau. Its official web services, apps, and AI features in Chrome do not function in these regions. Even if your Google account is registered in China, these features won’t be visible.

Midjourney, Perplexity, and Others—General Restrictions
Feedback from developer communities indicates that mainstream AI applications like Midjourney and Perplexity also impose varying degrees of access restrictions on Chinese users, requiring overseas payment methods or simply stating that the service is unavailable in their region. ByteDance’s international product Cici has also set regional restrictions, making it unusable in both China and the U.S., reflecting the geopolitical pressures on commercial products.

In summary, the mainstream foreign AI applications that restrict Chinese users encompass nearly all core products from U.S.-based AI companies, covering everything from large model conversations to AI art and programming, effectively creating a comprehensive blockade.

China’s AI Goes Global: Why the Open Approach?

While U.S. AI companies enforce strict barriers against Chinese users, Chinese AI firms are actively expanding globally.

Chinese AI not only welcomes overseas users but actively seeks to expand abroad.
MiniMax, a Shanghai-based large model unicorn, successfully entered the capital market in early 2026, showing impressive market performance. Its success is largely attributed to its strong performance in overseas markets, where its models compete effectively with GPT-4 in mathematics and coding capabilities.

The GLM series from Zhiyuan AI has recently gained popularity among U.S. developers, and Chinese manufacturers have captured significant market share on global model routing platforms like OpenRouter due to performance and cost advantages.

Even more dramatically, a U.S. AI programming startup, Cursor, was found to have incorporated a self-developed model, Composer-1, which unexpectedly switched to Chinese during a run. Industry insiders speculate that the model may have been influenced by Chinese manufacturers like Zhiyuan GLM. This unexpected twist highlights the complexities of the AI landscape.

Kuaishou is also rumored to be splitting its “Keli AI” business, aiming for a high valuation. If successful, another Chinese AI company will enter the international capital market.

Moreover, Chinese AI’s performance in overseas markets is not limited to just IPOs: in 2025, China’s AI model usage reached more than twice that of the U.S. The performance gap between leading AI models from both countries has narrowed, and China’s advantages in application layers are becoming increasingly evident.

In fact, Chinese open-source models are breaking through in the U.S. market, with downloads from Chinese models on Hugging Face surpassing certain open-source models from Meta.

With such a stark contrast—one side closing doors and the other opening them—it is clear that this is not merely a difference in commercial strategy but reflects the deeper geopolitical and economic logic of both nations in the AI arena.

Behind the Bans: Policy Pressure as a Sword of Damocles

To understand the ban actions of U.S. AI companies, one must look beyond commercial logic and seek answers from macro policy perspectives.

1. U.S. Export Controls Intensify
At the forefront is the U.S. export control system. In April 2026, the U.S. House of Representatives’ “U.S.-China Strategic Competition Special Committee” issued a stringent AI review report, urging the passage of the “Remote Access Security Act” to close loopholes that allow China to access computing power through the cloud.

Google, OpenAI, and Anthropic are all U.S. entities and must comply with the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) regulations, which explicitly prohibit providing advanced AI services to countries of concern, including mainland China. For AI unicorns like Anthropic, compliance is not optional but a necessity.

2. Investment Restrictions Evolve from “Voluntary” to “Mandatory”
With the signing of the “Fiscal Year 2026 National Defense Authorization Act,” the U.S. has formally legislated foreign investment reviews. This means that U.S. tech companies expanding into the Chinese AI market face not only export control risks but could also be deemed as “covert investment,” triggering legal consequences. From a compliance and risk-avoidance perspective, proactively banning Chinese users has become a “safe card” for U.S. AI companies.

3. The “50% Rule” Closes Off Workarounds
The U.S. has further strengthened the “50% rule”: if a Chinese shareholder listed on the entity list directly or indirectly holds 50% or more of a non-U.S. company, that company will automatically face the same restrictions. This explains why Anthropic has expanded its ban to include companies with over 50% Chinese ownership, regardless of their operational location.

In other words, the U.S. has built a “technological Berlin Wall,” with AI being the latest brick in this wall.

Deeper Reasons: Security Fears, Hegemonic Anxiety, and Commercial Interests

1. The Real Intent Behind “National Security” Narratives
On the surface, it’s about “national security,” but fundamentally, it’s about maintaining technological hegemony and curbing China’s technological rise. AI is viewed as the core driving force of the new technological revolution, making it a focal point of competition. Ensuring U.S. dominance in AI has become a rare consensus among both parties in Washington.

2. Rapid Progress of Chinese AI Raises U.S. Alarm
Another reason behind the bans is the rapid advancement of Chinese AI. Reports from think tanks like Rand Corporation indicate that the performance gap between AI models from both countries has nearly closed, with China holding a 40% to 50% cost advantage. This rapid catch-up has made Silicon Valley giants uneasy. Restricting access for Chinese users not only limits their ability to leverage U.S. advanced AI for technological accumulation but also weakens their dependency on the U.S. technology ecosystem. From a strategic perspective, this is a form of “cutting off the supply.”

3. The Tug of War Between Commercial Interests and Political Demands
For companies like Anthropic and OpenAI, Chinese users do hold commercial value. However, with rising compliance costs, an “all-or-nothing” ban has become the least risky option. After all, the 1.45 million banned accounts involved many users accessing services through unofficial channels. In these situations, implementing a blanket ban to mitigate risks is more a rational choice than a subjective anti-China stance.

Why Does Chinese AI Welcome All?

Understanding the logic behind U.S. AI bans makes the reverse choice of Chinese AI clear.

1. Survival Needs: The Global Market is Crucial for Chinese AI
Chinese AI companies are expanding overseas primarily because domestic market competition is too fierce. With a crowded market and limited capacity, going global is an inevitable choice for survival and growth.

What is the core competitive advantage of Chinese AI in overseas markets? Low cost and high efficiency. The U.S. market lacks high-quality, continuously updated open-source models, while Chinese manufacturers excel in this area. When U.S. AI services are priced high, Chinese companies offer more cost-effective alternatives.

2. Technological Confidence: Not Afraid of “Reverse Technology Leakage”
While the U.S. fears that China will use American AI to develop its technology, China is not worried about U.S. users utilizing its large models. In terms of model performance, leading Chinese models can compete with top U.S. products. This technological confidence is also reflected in the fact that Chinese AI has built an independent technological system, no longer reliant on foreign technologies. Opening up for global use allows for the acquisition of more user data and application feedback, accelerating technological iteration—a positive feedback loop.

3. Strategic Game: Using Openness to Counter Blockades
The U.S. builds walls while China opens doors. This is a strategic game in itself. As the U.S. attempts to encircle and block China in the AI field, Chinese AI companies are serving users globally. This differentiated expansion path—winning global market share through openness—is not just a commercial logic but a strategic choice.

4. Capital Needs: Listing in the U.S. or Hong Kong Remains an Important Option
The high valuations of tech stocks in international capital markets continue to attract significant interest. Whether it’s MiniMax’s listing on the Hong Kong stock exchange or Kuaishou’s plans to spin off Keli AI, it indicates that Chinese AI still has aspirations in international capital markets.

AI Knows No Borders, But Competition Has Stances

From Claude’s sweeping bans to ChatGPT’s IP restrictions, and the U.S. government’s layered legislative limits on AI investments and exports, all point to a stark reality: AI technology is being deeply politicized. Once regarded as a “global public good,” artificial intelligence is increasingly being framed within the context of bipolar confrontation, becoming yet another tool in geopolitical games.

However, China’s response is not passive. On one hand, driven by companies like DeepSeek and Zhiyuan AI, performance has reached a level where it can compete with U.S. giants; on the other hand, Chinese AI is embracing the global market with a more open attitude. While the U.S. legislates barriers, China is leveraging international opportunities—breaking through geopolitical barriers is not only a matter of political confrontation but also a testament to technological and market strength.

Of course, the open attitude of Chinese AI is not without costs. Technological expansion also faces regulatory pressures and geopolitical risks. The inability to use ByteDance’s Cici in both China and the U.S. is a prime example. Additionally, U.S. export restrictions on Chinese AI chips highlight that an open stance cannot completely avoid the impacts of geopolitical conflicts.

AI may change the world, but it may ultimately not change one fact: the essence of technological competition is the competition of comprehensive national strength between countries.

As you ponder these issues, you might be considering several questions:

Will Chinese users still be able to access foreign AI in the future?
Will the path of Chinese AI going global be smooth?
In this AI competition, who will ultimately prevail?

In the short term, U.S. restrictions are unlikely to ease, and China’s global expansion will not stop; in the long term, whoever can truly bring AI technology to more people will win both users and the future.

Perhaps as we see it: when Claude asks for your passport and facial verification, you might no longer need it. Because Chinese AI has already taken the stage on the world platform.

Understanding Vibecoding: Generating Cross-Platform Apps from a Single Description

Tue, 12 May 2026 00:00:00 +0000

Introduction

Vibecoding tools generate the underlying logic for cross-platform apps by using a “single business intent description” to drive a unified UI semantic model. This model is then processed through code generation pipelines tailored for different platforms, producing source code for Web (Vue/React), iOS (Swift + SwiftUI), and Android (Kotlin + Jetpack Compose). This article dissects the five-layer functional architecture that Vibecoding tools must possess, compares six representative products—UXbot, Bolt, v0, Lovable, Replit Agent, and Tempo—on their cross-platform generation capabilities, and provides a practical evaluation checklist.

Why “One Description, Three Platforms” is Essential

According to a Harvard Gazette interview in April 2026, Karen Brennan, a professor at Harvard Graduate School of Education, observed that the essence of Vibecoding is to “describe what you want in natural language, and let an AI implement it as runnable software.” This practice was named by computer researcher Andrej Karpathy in February 2025 and quickly spread across various fields including product development, education, and entrepreneurship. It enables students with no coding experience to deliver complete web applications in just six weeks.

However, products rarely exist solely on the web. Grand View Research indicates that the global mobile application market size has grown from $252.89 billion in 2023 to an expected $626.39 billion by 2030, with a compound annual growth rate of 14.3%. The Asia-Pacific region accounts for 32% of global revenue, while the Apple Store alone captures 62.8%. This means that any product with ambitions for user scale must cover Web, iOS, and Android platforms simultaneously, rather than just launching a website.

The scarcity of developers capable of writing native mobile code is surprising. JetBrains’ “Developer Ecosystem 2024” report shows that only 14% of global developers use Kotlin and 6% use Swift, far less than the 61% using JavaScript and 57% using Python. This supply imbalance leads to a side effect: many startup teams end up only developing for the web, indefinitely postponing mobile development until competitors establish a more comprehensive terminal experience.

The 2024 Stack Overflow Developer Survey reveals another side of the data: 76% of developers are using or planning to use AI coding tools, up from 70% the previous year, with 62% already using them in their daily work. 81% believe that the greatest value of AI tools is “increased productivity,” and 82% are using AI to assist in coding. Vibecoding tools have emerged at the intersection of these two trends: one side facing a scarcity of native mobile code supply, and the other witnessing an explosion in productivity from natural language to code.

The Five-Layer Functional Architecture of Vibecoding Tools

Delivering cross-platform apps is not merely about “packaging the web into a shell”; it involves a complete technical pipeline. To evaluate whether a Vibecoding tool truly possesses cross-platform generation capabilities, it is recommended to check the following five layers:

1. Semantic Layer: From Natural Language to Unified Business Intent Model

The first input received by Vibecoding tools is natural language. Truly powerful products will first parse this description into a platform-agnostic business intent model, detailing “which roles, which pages, what tasks each page carries, how pages navigate, and the dependencies between data.” This step is the foundation for cross-platform generation—if the model itself is structured like a web (e.g., DOM tree, Tailwind styles), then any downstream mobile output is merely “a shell around the web.”

2. Structural Layer: Flow Canvas and Multi-Page Routing Planning

The business intent model needs to be visualized and editable. The flow canvas allows users to see a global view of “how many pages the entire product has, how users flow through them, and which pages require login” before generation, enabling corrections to any misunderstandings by the AI regarding requirements. The output from the flow canvas corresponds to the routing systems of each platform: the front-end routing table for the web, AppRouter for iOS, and NavHost routing constants for Android. Without a flow canvas, a Vibecoding tool will likely only generate single-page web applications, failing to achieve complete alignment across platforms.

3. Prototype Layer: Independent Rendering of UI Semantics Across Platforms

The UI semantic layer must resolve the issue of “the same login page being a Tailwind class name on the web, a SwiftUI VStack on iOS, and a Jetpack Compose Column on Android.” Mature Vibecoding tools will abstract visual constants such as colors, spacing, corner radius, and font sizes into a unified Theme, then rewrite them in the native syntax of each target platform instead of displaying a static image in a WebView. This layer determines whether the exported applications possess a native feel—whether they adapt to system light/dark modes, respond to system gestures, and pass native store reviews.

4. Code Layer: Independent Engineering Output for Each Platform

This layer is often obscured in competitor documentation. True cross-platform generation requires outputting three separate engineering projects:

Web: Organized with Astro / Vue 3 / TypeScript / Tailwind CSS or React + TypeScript, with components in independent subdirectories, <script setup lang="ts">, and Pinia/Zustand for state management.
iOS: Swift + SwiftUI + XcodeGen, following the MVVM architecture, with @MainActor final class for ViewModel and @Published properties driving view refresh.
Android: Kotlin + Jetpack Compose + Gradle Kotlin DSL, following the MVVM architecture, with each page paired with Page + ViewModel, and a single immutable UiState object.

Each of these projects must be independently openable, buildable, and deployable/applicable, rather than being bundled in a compressed file with unrelated HTML.

5. Preview and Validation Layer: Multi-Platform Simulators and Interactive Testing

No matter how beautiful the code is, it must be verifiable before delivery. The final layer of Vibecoding tools is the built-in multi-platform simulator: users can switch between “Web Preview / iPhone Preview / Android Preview” in the browser, clicking each button and navigating each page to confirm interactions meet expectations. The simulator should support real page transitions, state retention, and form validation, rather than merely displaying a high-definition image. Without this layer, the so-called “cross-platform generation” remains at the stage of “code generated but no one can confirm it runs.”

Comparison of Six Vibecoding Tools on Cross-Platform Generation

Tool	Input Type	Flow Canvas	Web Output	iOS Output	Android Output	Multi-Platform Simulator	Use Cases
UXbot	Natural Language + Flow Canvas	Yes	Vue 3 / HTML / Tailwind	Swift + SwiftUI Native Project	Kotlin + Jetpack Compose Native Project	Yes	Full products needing simultaneous mobile app launches
Bolt	Natural Language	No	React / Next.js	No Native (requires React Native modification)	No Native (requires React Native modification)	Web	Web-focused MVPs and B2B tools
v0	Natural Language + Design Reference	No	React + shadcn/ui	Not Supported	Not Supported	Web	Web component and landing page
Lovable	Natural Language	No	React + Tailwind + Supabase	Not Supported	Not Supported	Web	Web SaaS prototypes
Replit Agent	Natural Language + Cloud Project	No	Multi-framework options	Not Supported	Not Supported	Web	Web prototypes and code experimentation
Tempo	Natural Language + Figma Import	Yes	Simplified React	Not Supported	Not Supported	Web	Web UI customization

This table reveals a key insight: the vast majority of Vibecoding tools have left the “native mobile” category empty, interpreting “cross-platform” as “responsive web” or “first doing web and then rewriting in React Native.” Currently, only a few tools can genuinely achieve “one description while outputting native iOS + native Android + web.”

In-Depth Analysis of Six Tools

1. UXbot

UXbot is an AI full-chain tool that transforms requirement descriptions into complete multi-page interactive app interfaces and deliverable front-end code. Its differentiated advantages align perfectly with the five-layer architecture mentioned earlier: the semantic layer uses natural language for unified modeling; the structural layer provides a visual flow canvas for users to lock in product structure before generation; the prototype layer generates interactive prototypes that support real page transitions and interaction flows, with a built-in real-time simulator for direct preview of web and mobile (Android/iOS) interactions; the code layer outputs three independent engineering projects—Web using Astro + Vue 3 + TypeScript + Tailwind CSS, with each page’s components in independent subdirectories, <script setup lang="ts">, and Pinia for state management; iOS using Swift + SwiftUI + XcodeGen (project.yml) for managing project configurations, with ViewModel using @MainActor final class + @Published pattern, and a unified AppRouter managing routing transitions; Android using Kotlin + Jetpack Compose + Gradle Kotlin DSL, with each page composed of Page + ViewModel pairs, and ViewModel managing state with a single immutable UiState object, with NavHost managing routing constants.

The workflow for UXbot is: input requirements → confirm flow canvas planning product structure → generate prototype preview validation → precise local editing → export code for cloud execution. This process means product managers can show the team a complete demo of all three platforms on the day of project initiation, rather than waiting three months after the web launch to start the mobile project. For independent entrepreneurs, UXbot can even allow one person to produce all three platforms simultaneously, compressing what would normally require three separate teams into a single workday.

2. Bolt

Bolt is a representative of web Vibecoding in recent years, combining WebContainers technology with AI code generation. Users can input requirements in the browser and see a runnable React application. Bolt’s strengths lie in output speed and integration with the web ecosystem (npm packages are natively usable), but its output model is web-centric—creating mobile versions requires either manually rewriting in React Native or using Capacitor for shell encapsulation, thus not fitting into the “one generation for three platforms” category. It is suitable for teams requiring speed for pure web MVPs.

3. v0

Launched by Vercel, v0 focuses on generating React components/pages from natural language and design references. Its output is deeply integrated with the Next.js and shadcn/ui ecosystems, making it an ideal complement for teams already using Next.js. However, its shortcoming is that v0 outputs at the component level rather than the complete application level, and it does not involve any native iOS/Android production, making it more of a “web accelerator” than a complete three-platform Vibecoding tool.

4. Lovable

Lovable binds AI Vibecoding with the Supabase backend, generating products that come with databases, authentication, and storage, suitable for solo entrepreneurs looking to quickly build web SaaS. Lovable produces high-quality web UI, but its mobile strategy remains “responsive web,” not generating independent iOS or Android projects. Thus, it is more like a “web full-stack accelerator,” not in the same league as three-platform generation.

5. Replit Agent

Replit Agent combines Replit’s cloud IDE with AI capabilities, allowing users to issue commands in natural language while the Agent autonomously creates files, installs dependencies, writes code, and deploys. For developers who enjoy a “chat while viewing code” workflow, this is a comfortable process, but its output still primarily focuses on web and general scripts, requiring manual setup of Swift/Kotlin projects for native mobile.

6. Tempo

Tempo uses Figma design drafts as input, employing AI to translate designs into React components and allowing for continued iteration within the same workspace. Its value lies in automating the “Figma to React” transition for designers, suitable for medium-sized teams with established design specifications. Support for three-platform apps is currently limited to web; generating native iOS/Android projects requires connecting with other tools.

Eight-Point Checklist for Evaluating Vibecoding Tools’ Cross-Platform Generation Capability

During the selection phase, you can use the following eight questions to directly inquire with candidate tools’ representatives or validate during trials:

Can you describe the entire product in natural language at once, rather than needing to specify each page individually?
Can you view a “whole product flow canvas” before generation and modify it?
After generation, do you provide independent simulators for Web/iOS/Android that allow for clickable navigation?
Does the iOS output export as a SwiftUI native project (.xcodeproj / XcodeGen), or is it a WebView shell?
Does the Android output export as a Jetpack Compose native project (Gradle Kotlin DSL), or is it HTML packaged?
After exporting, are there three independent buildable projects, or can only the web run directly?
When regenerating for the same requirement, do the UIs across platforms maintain consistent business semantics (e.g., no “web has, mobile does not” scenarios for the login page)?
Is there a precise editor that allows users to directly modify specific UI elements when unsatisfied, rather than starting over?

A qualified cross-platform Vibecoding tool should be able to answer “yes” or “there is” to all eight questions.

Practical Usage Paths for Three Types of Teams

1. Independent Entrepreneurs: Simultaneous Launch Across Platforms

By planning the flow canvas for login/core tasks/checkout as three main lines, entrepreneurs can generate prototypes using natural language, validate complete interactions in the simulator, and finally export code for all three platforms to be deployed on Vercel/TestFlight/Google Play for internal testing, compressing the entire cycle to 3-5 days. The 81% of respondents in Stack Overflow data who believe productivity is the greatest benefit of AI tools largely represent this type of multi-role entrepreneur.

2. Product Managers: From PRD to Demo in One Meeting

Product managers can feed paragraphs from the PRD into Vibecoding tools in natural language, adjust navigation logic on the flow canvas, and directly present the generated demos of all three platforms in project kickoff meetings, allowing development and business teams to reach consensus on the same prototype. Compared to the previous sequential model of writing PRDs, then scheduling prototypes, and finally scheduling development, this represents a significant upgrade to parallelization.

3. Small to Medium Development Teams: Using Native Projects as Starting Points Rather Than End Goals

Teams can treat the SwiftUI/Jetpack Compose projects generated by Vibecoding tools as scaffolding, adding business logic, integrating their own backend, and including unit tests. Since the exported project structures follow the recommended MVVM + @MainActor pattern for SwiftUI and the Jetpack Compose + UiState pattern for Android, teams do not incur additional costs in “migrating to a real project.”

Frequently Asked Questions (FAQ)

Q1: Can the mobile code generated by Vibecoding tools pass App Store and Google Play reviews?

Whether it passes review depends on two factors: whether it uses native platform controls and lifecycles, and whether it complies with respective privacy declarations. Tools like UXbot that directly output SwiftUI/Jetpack Compose native projects meet the technical syntax requirements for review, with the remaining tasks being to supplement app privacy lists, permission descriptions, and other routine preparations, which are not fundamentally different from manually developed projects.

Q2: If I need to modify a page after generating code for all three platforms, where should I make the change?

The recommended approach is to “first modify the semantic layer in the Vibecoding tool and regenerate,” ensuring consistency across all three platforms; if only one platform needs a minor adjustment (e.g., an iOS-specific animation), it can be modified directly in the native project. If the exported project has a well-structured format (separate ViewModels, clear routing tables), subsequent changes will have the same cost as a standard project.

Q3: Can product managers without Swift/Kotlin experience directly publish the generated mobile apps?

They can achieve “preview and delivery review,” but publishing to the App Store/Google Play still requires a developer account, certificates, and upload processes. It is recommended that product managers pair with a front-end/mobile engineer, with the tool handling 90% of the coding work, while the engineer manages the 10% compliance processes related to accounts, certificates, packaging, and submission.

Q4: How does Vibecoding’s cross-platform generation differ from frameworks like React Native or Flutter?

Cross-platform frameworks still require teams to write code, just one set of cross-platform code. The value of Vibecoding tools lies in the full automation of the “business description → three-platform code” process, outputting native code for each platform rather than a cross-platform intermediary layer. The two are not in competition—using Vibecoding tools to get the prototypes and initial code running across all three platforms, then deciding whether to consolidate to React Native or Flutter for long-term maintenance is a common strategy.

Q5: Will generating all three platforms at once lead to poor code quality and future maintenance difficulties?

This is an architectural issue, not an inherent flaw of Vibecoding. Specific quality indicators include adherence to official recommended architectures for each platform (SwiftUI’s MVVM + @MainActor, Jetpack Compose’s UiState pattern), retention of complete TypeScript types, and centralized routing declarations. Choosing Vibecoding tools that excel in these areas will not result in higher long-term maintenance costs than hand-written code.

Conclusion

Generating web, iOS, and Android apps simultaneously is not about forcibly distributing a single web codebase across platforms; it is about using a unified business intent model to drive the generation of native code for three platforms. Truly qualified Vibecoding tools must excel in the five layers of the semantic layer, structural layer, prototype layer, code layer, and preview and validation layer. Currently, most Vibecoding products on the market leave gaps in the native mobile category. Product managers, independent entrepreneurs, and small to medium development teams should validate functionality against the five-layer architecture and the eight-point checklist, ensuring that tools capable of generating all three platforms simultaneously truly deserve the label “three-platform Vibecoding.”

Carbon Neutrality Weekly Report: AI and Energy Synergy, Local Policies, and Industry Practices

Mon, 11 May 2026 00:00:00 +0000

Carbon Neutrality Policies

1. Four Departments Issue Document to Promote AI and Energy Synergy

On May 8, the National Development and Reform Commission, National Energy Administration, Ministry of Industry and Information Technology, and National Data Bureau issued the “Action Plan for Promoting AI and Energy Synergy.” The plan outlines ten key areas, including ensuring reliable energy supply for computing power facilities, promoting the green and low-carbon transformation of computing power facilities, and enhancing the economic synergy between computing power and electricity. It breaks down into 29 key tasks.

The plan aims to establish a safe, green, and economical energy guarantee system for AI innovation by 2027, significantly enhancing the interaction between clean energy and computing power facilities. By 2030, it seeks to achieve world-leading levels in clean energy supply for AI computing power facilities and the research and application of AI technologies in the energy sector.

Quick Review: The joint issuance of the action plan shifts the relationship between AI and energy from one-way support to mutual empowerment, optimizing clean energy supply and computing power layout.

2. New Evaluation Method for Beautiful China Construction Released

On May 7, the General Office of the CPC Central Committee and the General Office of the State Council issued the “Evaluation Method for the Effectiveness of Beautiful China Construction,” effective immediately. This evaluation will serve as an important reference for the comprehensive assessment of provincial and municipal party committees and governments, influencing rewards and punishments.

The evaluation focuses on five aspects, including the implementation of responsibilities, completion of annual goals, execution of key tasks, performance of fund usage, and public satisfaction regarding environmental quality improvements.

Quick Review: The evaluation method forms a closed-loop assessment system, indicating that ecological protection is now a rigid constraint rather than a soft indicator for local development.

3. Charging Volume on Highways Increased by 55.6% on May Day

On May 2, the National Energy Administration reported that on the first day of the May Day holiday, the charging volume for new energy vehicles on highways reached 23.0339 million kilowatt-hours, a 55.6% increase year-on-year, setting a historical record for the holiday’s first day.

Quick Review: The surge in charging volume reflects the rising penetration of new energy vehicles and tests the peak capacity of infrastructure, highlighting the need for enhanced rapid scheduling and smart operation levels in charging networks.

4. National Energy Administration Promotes Green Energy Initiatives

Recently, the National Energy Administration held a press conference addressing the green certificate market. They emphasized four key areas to invigorate the market: improving trading mechanisms, enhancing coordination, expanding consumption scale, and establishing certification mechanisms for green energy consumption.

Quick Review: Transitioning green certificates from corporate trading tools to residential use is crucial for stimulating end-user consumption of green energy.

5. Meteorological Bureau Focuses on Extreme Weather in New Energy Development

At a recent press conference, the China Meteorological Administration outlined key tasks for the 14th Five-Year Plan, including building a meteorological disaster monitoring and early warning system and improving services for the energy sector.

Quick Review: The new power system’s reliance on precise weather forecasts emphasizes the need for enhanced meteorological services to optimize energy load and market transactions.

6. National Carbon Market Price Update

Last week, the national carbon market saw a maximum price of 80.80 yuan/ton, with a total transaction volume of 1,204,900 tons and a total transaction value of 93,926,670 yuan.

Local Developments

1. Guangdong’s 14th Five-Year Plan Released

Recently, Guangdong Province released its 14th Five-Year Plan, which includes 22 indicators across six categories, aiming for a stable transition to peak carbon emissions.

Quick Review: The implementation of a dual control system for carbon emissions in Guangdong will help export-oriented enterprises prepare for product carbon footprint certification.

Corporate Practices

The aviation sector is a significant source of greenhouse gas emissions. As of April 28, five out of twelve A-share airline companies disclosed their carbon emissions, totaling approximately 96.06 million tons of CO2 equivalent.

Quick Review: The complexity of emissions in the aviation value chain requires collaboration across the supply chain to effectively reduce carbon emissions.

Among 76 A-share construction material companies, 36 disclosed their ESG reports, with cement companies accounting for over 90% of the total emissions of approximately 475 million tons of CO2 equivalent.

Quick Review: Cement companies must accelerate the adoption of alternative fuels and technological innovations to manage carbon assets effectively.

Fuzhou Accelerates AI Integration with Industry

Mon, 11 May 2026 00:00:00 +0000

Fuzhou Accelerates AI Integration with Industry

On May 9, it was reported that Fuzhou’s relevant departments actively responded to suggestions from local representatives to promote the construction of a comprehensive robotics ecosystem and deepen the integration of “Artificial Intelligence +” with Fuzhou’s advantageous industries, accelerating the collaborative development of intelligent manufacturing.

Enhancing Smart Manufacturing

Representative You Xiaotian pointed out that although Fuzhou has policy advantages, there are still gaps in the scale of the robotics industry and core technology accumulation compared to advanced cities. He proposed several concrete measures, including establishing a dedicated robotics fund and expanding “robotics +” application scenarios.

These suggestions align closely with Fuzhou’s current work deployment. The Municipal Bureau of Industry and Information Technology introduced that Fuzhou has formed a robotics industry pattern characterized by “one core leading, three belts collaborating, and all-region interaction.” In the Fuzhou New Area, the Shanghai Robotics Industry Technology Research Institute’s Fuzhou Innovation Center has been inaugurated, aiming to create a new highland for intelligent medical robots. In Lianjiang, Shenhao Technology’s underwater robots have been successfully demonstrated in national marine ranches. In Changle, Jufu Technology, a humanoid robot company under Huawei, is targeting four major industrial clusters: textiles, footwear, lithium batteries, and photovoltaics, providing experimental grounds for product implementation.

Avoiding Homogeneous Competition

In response to the suggestion to avoid homogeneous competition and strengthen regional collaboration, Fuzhou is actively learning from experiences in other regions to enhance the robotics industry’s collaborative development. Notably, local company Rockchip’s RV series intelligent vision chips and RK series SoC chips have become the “Chinese chips” for well-known companies like Yushu, DJI, Zhiyuan, and Ecovacs, ranking among the top in annual sales within the industry. The robotics industry chain in Fuzhou is rapidly taking shape, transitioning from “purchased assembly” to “core manufacturing.”

Driving Integration of Land and Sea

You Xiaotian also suggested leveraging the construction of “Maritime Fuzhou” and the intelligent transformation of “Southeast Auto City” by opening a number of demonstration scenarios. The Municipal Bureau of Industry and Information Technology responded by promoting the application of industrial robots in automotive assembly, shipbuilding, and electronics manufacturing. In Lianjiang, the traditional welding section of the Mawei Shipyard is actively undergoing intelligent upgrades, while Yongyue Intelligent’s climbing sandblasting and painting robots have been employed for rust removal and painting in shipbuilding factories. In Changle, Xinhua Textile’s fully automated production line is technically capable of completely replacing German and Japanese equipment and is poised for launch.

In the field of artificial intelligence, representatives pointed out that enterprise-level AI applications face three major challenges: “shallow implementation, repetitive construction, and data silos.” They suggested establishing an AI empowerment center in Fuzhou, implementing the “AI + advantageous industries” lighthouse project, and creating an “AI + marine” innovation application zone at the mouth of the Min River. The Municipal Bureau of Industry and Information Technology indicated that it has achieved an 80% green electricity ratio in newly established data centers and computing power clusters, steadily progressing towards 100% coverage, providing green, low-carbon, and efficient computing power support for the construction of an enterprise-level AI innovation system.

The Municipal Bureau of Ocean and Fisheries provided a vivid example of “AI + marine” practice: the national marine ranch at Huangqi Peninsula in Lianjiang has completed the deployment of 1,026 reef bodies and established an underwater monitoring platform that can monitor water temperature, salinity, dissolved oxygen, and underwater biological activity in real time. By equipping fishing boats with high-throughput satellite communication terminals and Beidou positioning devices, Fuzhou has built an intelligent management platform for fishing vessels, achieving visibility and communication with the boats.

From Satisfaction with Responses to Satisfaction with Results

The representatives’ suggestions not only analyze problems but also convey public opinion. From establishing a new industrial development investment fund to continuously optimizing the marine disaster risk prevention system, Fuzhou is efficiently linking multiple departments to accelerate the “Artificial Intelligence +” initiative.

In evaluating the overall response to the suggestions, You Xiaotian expressed satisfaction with the handling units. He stated that he is confident in Fuzhou’s determination and effectiveness in cultivating new productive forces as he sees the suggestions gradually transforming into tangible policies and implementation scenarios.

As the “14th Five-Year Plan” is deeply implemented, Fuzhou is using the precise suggestions from representatives as a starting point to continue deepening the “Artificial Intelligence +” actions, accelerating the integration of AI with the city’s advantageous industries and marine economy, allowing robots to showcase their capabilities on the production front and enabling AI models to thrive in the blue seas.

ChatGPT's Functionality in May 2026: A User Experience Review

Sat, 09 May 2026 00:00:00 +0000

ChatGPT’s Functionality in May 2026: A User Experience Review

In the current era where AI technology is deeply integrated into daily life, ChatGPT stands out as a phenomenal tool. Its functionality directly impacts the user experience and efficiency. In May 2026, OpenAI officially launched the GPT-5.5 Instant as the default model, featuring significant optimizations in hallucination control, memory capacity, and response conciseness. Additionally, o.zzmax.cn serves as an excellent AI model aggregation site, allowing users to intuitively compare ChatGPT with other mainstream models and quickly find AI tools that meet their daily needs.

1. Basic Text Capabilities: Mature and Stable, Covering Everyday Scenarios

Text processing is the core foundation of ChatGPT. After multiple iterations, this capability has matured, covering the vast majority of text needs in users’ learning, work, and life. Whether drafting emails, organizing notes, writing copy, translating languages, summarizing long texts, or answering common knowledge questions, ChatGPT provides clear and coherent results.

The May 2026 upgrade further enhanced the practicality of text capabilities, reducing the hallucination rate by 52.5%, significantly decreasing factual errors in high-risk areas such as medical, legal, and financial fields. Responses are also more concise, with an average length reduction of about 30%, eliminating redundant expressions and ineffective formats, thus significantly improving communication efficiency. In everyday use, whether students are organizing class notes, professionals are writing work reports, or ordinary people are creating personal essays or translating foreign materials, ChatGPT responds efficiently without needing complex instructions to meet basic needs.

However, there are slight shortcomings in basic text capabilities. Firstly, the emotional depth in literary creation is lacking; while it can construct frameworks for poetry, prose, and novels, its impact and nuance do not match human creators. Secondly, it sometimes misinterprets niche dialects and internet memes, occasionally providing irrelevant answers. Thirdly, the coherence of long text creation can falter; content exceeding ten thousand words needs to be guided in segments to avoid logical repetition or detail contradictions.

2. Multimodal Functionality: Comprehensive and Practical, with Room for Detail Optimization

Multimodal capability is a core competitive advantage of current large models. ChatGPT has achieved cross-modal interaction involving text, images, and audio, covering image understanding, content generation, and audio transcription, proving to be quite practical. In terms of images, it can accurately recognize handwritten text, mathematical formulas, chart data, and everyday objects, describing image content and answering questions within images, as well as analyzing design blueprints. When generating images, it can create illustrations, posters, and product images based on text prompts, with diverse styles and complete details. In audio, it supports speech-to-text, real-time translation, and sentiment analysis, capable of recognizing speech content in noisy environments with high transcription accuracy. The 2026 update improved the fluency of multimodal interactions, enhancing response speed after uploading images or audio, and the accuracy of interpreting complex images (like industrial blueprints and medical images) has also progressed. In daily scenarios, students can upload photos of assignments to obtain problem-solving ideas, professionals can upload meeting recordings to quickly generate minutes, and creators can generate design inspiration images, covering multiple scene needs.

However, multimodal functionality still has notable limitations. Firstly, there is a lack of video processing capability; it cannot directly analyze video content or summarize key points, requiring third-party tools for format conversion. Secondly, the creative ceiling for image generation is not high, with insufficient fidelity in complex compositions and niche artistic styles, often leading to element clutter. Thirdly, audio duration is limited; processing speeds significantly decrease for audio longer than one hour, often missing key information.

3. Tool Integration and Memory Capability: Convenient and Efficient, with Personalization to be Deepened

ChatGPT’s tool integration and memory capabilities are key to enhancing user engagement and are important aspects of its functionality. In terms of tools, it includes built-in web search, deep research, code interpreter, and office plugins, allowing users to complete multi-task processing without switching platforms. Web search can obtain real-time information to answer current affairs and industry dynamics questions; deep research can integrate multiple authoritative sources to generate structured reports, suitable for academic research and business analysis scenarios; the code interpreter supports writing and debugging code, solving programming issues and data calculations; office plugins can link to Excel and Google Sheets for data organization, formula writing, and table optimization.

The memory capability received a significant upgrade in May 2026, introducing the “memory source” feature, which shows how historical conversations, uploaded files, or Gmail content influence current responses. Users can view, delete, or modify memories, ensuring privacy control. Cross-conversation memory is more stable, able to remember user preferences and historical needs, providing personalized responses without the need for repeated explanations. For example, long-term users will find that ChatGPT remembers their writing styles and areas of interest, making subsequent creations more aligned with their needs.

However, there are still shortcomings in tool integration and memory capabilities. Firstly, the threshold for tool usage is relatively high; features like deep research and the code interpreter require a certain level of expertise, making it challenging for ordinary users to fully utilize them. Secondly, the memory range is limited, unable to retain vast amounts of information long-term, and conversations that are spaced too far apart may lead to forgetting core content. Thirdly, compatibility with third-party tools is generally average, with some niche office software and design tools unable to link, limiting scenario expansion.

4. Function Layering and Permission Differences: Clear Gradients, Noticeable Limitations for Free Users

ChatGPT adopts a layered functional design, with free, Plus, Pro, and enterprise versions offering progressively increasing permissions to meet different user needs. The free version centers around GPT-5.5 Instant, supporting basic text, simple image understanding, and limited search functions, catering to light usage needs but with message quantity limits and higher response delays during high concurrency.

The Plus version, as the mainstream paid version, unlocks all basic functions, supporting the GPT-5.5 Thinking deep reasoning model, with relaxed message limits and access to the code interpreter, advanced image generation, and long document analysis. The Pro version targets professional users, providing higher computing power, longer context windows, and priority response permissions, suitable for high-intensity creation and complex data analysis scenarios. The enterprise version focuses on security compliance, supporting private deployment, fine-grained permission management, data encryption, and audit logs to meet enterprise data security needs.

While this layered design is reasonable, the limitations for free users are significant, making it difficult to experience core advanced features. The price threshold for paid versions is relatively high, leading to considerable long-term usage costs, and some features (like deep research and long document analysis) may not be practical for ordinary users, resulting in average cost-effectiveness.

Conclusion: Adapting to Mass Needs with Room for Advancement

Overall, ChatGPT’s functionality system is quite complete, with mature and stable basic text capabilities, comprehensive and practical multimodal features, convenient and efficient tool integration and memory capabilities, and layered design catering to different user needs, meeting the vast majority of ordinary users’ learning, work, and life demands. Despite shortcomings in literary creation depth, video processing, and free permissions, the overall strengths outweigh the weaknesses, making it one of the most balanced large models available today.

o.zzmax.cn continues to synchronize ChatGPT’s functionality updates and usage tips, providing users with a one-stop experience and comparison platform. As AI technology rapidly iterates, ChatGPT’s features will continue to optimize and upgrade. In the future, it must focus on detail experience, free rights, and professional depth to better adapt to users’ increasingly diverse needs, becoming a more versatile AI assistant.

DeepSeek V4 Flash: A Local Inference Engine for Mac by Antirez

Fri, 08 May 2026 00:00:00 +0000

DeepSeek V4 has begun pushing overseas developers to create dedicated high-speed pathways for it.

Just two weeks after its release, the first batch of native infrastructure for V4 has emerged in the open-source community.

This is not just a minor tweak on existing frameworks. It is not a generic GGUF loader; it is not a wrapper for llama.cpp; and it does not support other models at all.

It does one thing:

Run DeepSeek V4 Flash to its fullest on Mac.

This dedicated pathway is called ds4.c. The developer behind it is quite notable—Salvatore Sanfilippo, better known in the programming world as antirez.

He created Redis (with 74,000 stars on GitHub) and led the development of this globally popular in-memory database for 11 years.

Now, his new project ds4.c is a local inference engine specifically designed for DeepSeek V4 Flash.

Users have already reported running it on a 128GB Mac.

This time, DeepSeek has cleared out Mac inventory once again.

A Local Inference Engine for V4 Flash

On April 24, DeepSeek released the V4 series, with V4 Flash being the efficiency model: 284 billion total parameters, 13 billion active parameters, and a context of 1 million tokens.

Such a scale has almost always been reserved for cloud computing.

Antirez aims to fit it into a Mac, leading to the birth of ds4.c.

This is an inference engine written from scratch using C + Metal.

The entire project consists of just a few files, with C accounting for 55.4%, Objective-C 30.2%, and Metal 13.8%. It is Metal-only, with no runtime, no framework dependencies, and no abstraction layers.

Metal-only.

Metal is Apple’s own graphics and computing API, used to call the GPU on Mac, iPhone, and iPad, akin to CUDA in the Apple ecosystem.

The fact that ds4 only uses Metal means this engine runs solely on Apple Silicon, disregarding Nvidia and AMD graphics cards.

The sole goal of the project is:

To ensure V4 Flash runs locally on Apple machines, not just “can run,” but truly “usable.”

Current test results are quite impressive:

On a 128GB MacBook Pro M3 Max, with 2-bit quantization and 32K context, it pre-fills short prompts at 58.52 tokens/s and generates at 26.68 tokens/s.

Switching to a 512GB Mac Studio M3 Ultra, it achieves 468.03 tokens/s for pre-filling long prompts (11,709 tokens) and generates at 27.39 tokens/s.

For a 284 billion parameter MoE model, this speed is usable on local machines.

How Is This Achieved?

The key lies in three aspects.

First, asymmetric quantization.

Ds4 does not quantize all parameters to 2-bit; instead, it only quantizes the routing of the MoE expert layers, using IQ2_XXS for up/gate and Q2_K for down, which occupy the majority of the model space. Other components, such as shared expert layers, projection layers, and routing layers, retain Q8 precision.

Antirez states directly in the README:

These 2-bit quantizations are no joke; they perform well under coding agents and can reliably call tools.

Second, KV cache moved to disk.

Current LLM agent clients are stateless, sending the entire conversation segment with each request. The approach of general engines is to redo the pre-fill each time.

Ds4’s method writes the KV state to disk, matching token prefixes on subsequent requests, and if matched, loads directly from disk, skipping the pre-fill.

The cache key is the SHA1 hash of the token ID sequence.

This is particularly useful for agents like Claude Code, which send a 25K token initial prompt with each startup; after the first pre-fill, subsequent sessions can directly recover from disk.

Third, built-in compatibility layers for OpenAI and Anthropic APIs.

/v1/chat/completions follows the OpenAI protocol, while /v1/messages follows the Anthropic protocol. Tool calling has also been adapted. The README provides configuration examples for opencode, Pi, and Claude Code agent clients.

Regarding the motivation behind this, Antirez explains that while there are many excellent projects in local inference, attention is quickly diverted to the next model as new ones are released.

General engines must abstract to accommodate all models, which leads to compromises. What he aims to create is a deliberately narrow path, betting on one model at a time, validating with official logits, conducting long context tests, and integrating enough agents to confirm it is genuinely usable.

Once the framework was released, many users reported successfully running it on their Macs.

Are you ready to run V4 locally?

One Model, One Inference Framework

This development has sparked a larger discussion in the developer community:

Will we see a future where each model has its own inference framework?

A highly upvoted comment on Hacker News suggested an interesting direction: what if we start building hyper-optimized inference engines targeted at specific GPU and model combinations?

As GPUs become increasingly expensive, removing enough abstraction layers and directly coding for specific hardware and models could lead to significant optimizations.

However, this path has clear costs. As another comment pointed out, once a model becomes outdated, everything has to be rebuilt from scratch.

Antirez himself acknowledges this issue. He states that ds4 currently bets on DeepSeek V4 Flash, but the model may change.

The unchanging constraint is that local inference must run reliably on high-end personal machines or Mac Studios, starting with 128GB of memory.

What the future holds is hinted at in the README.

Currently Metal-only, there may be plans for CUDA support in the future. However, he writes cautiously that it may happen, but that’s all. This project deliberately remains small, fast, and focused.

More importantly, he presents a viewpoint in the README that local inference should be done well in three aspects, out of the box.

An inference engine with an HTTP API, a GGUF tailored for this engine and its assumptions, and a set of tests and validations integrated with coding agents.

This is a full-stack local inference approach, not just piecing together components but designing the entire chain as a product.

If this path proves successful, it could change the way local inference is approached.

When model vendors release new models, someone in the community will jump in to create a dedicated engine, specialized quantization, and agent integration for it. Each generation of models could have its own “antirez.”

Ds4 also has a candid detail. The README includes a statement that this software was developed with the “strong assistance” of GPT 5.5, with humans responsible for ideas, testing, and debugging.

Antirez states, If you do not accept AI-assisted development code, this software is not for you.

In just two weeks, from forking llama.cpp for adaptation to writing a dedicated engine from scratch, AI assistance has been crucial. This aspect may be even more noteworthy than ds4 itself.

One More Thing

Finally, let’s talk about Antirez.

His real name is Salvatore Sanfilippo, born in 1977 in Sicily. He created Redis in 2009 and led the project for eleven years before leaving in 2020.

When he left, he wrote that he programs to express himself, viewing code as a product rather than just a useful tool. He would rather be remembered as a bad artist than a good programmer.

In late 2024, he returned to Redis in an evangelist role.

Beyond Redis, he has also created Kilo (a text editor with less than 1000 lines of C code), dump1090 (an aviation ADS-B signal decoder), and linenoise (a micro replacement for readline).

He has also been playing with Flipper Zero, writing RF protocol analysis tools, and porting Asteroids to it. In 2022, he published a science fiction novel titled “WOHPE,” focusing on AI, climate change, programmers, and the interaction between humans and technology.

The first line of his personal homepage states, “I spend most of my professional time writing code and writing novels.”

Regarding the birth of Redis, he wrote:

My wife said that in the early years of Redis, I wrote most of the code sitting on the toilet with a MacBook Air 11 inch. I wish I could say she was wrong, but she is completely right.

This tone permeates all his projects: small, precise, and self-contained.

Ds4.c follows the same path.

Take a look at his note in the ds4 README about macOS bugs, and you can immediately sense his personality.

Ds4 includes a CPU inference path for correctness verification, but the current version of macOS has a bug in its virtual memory implementation that causes kernel crashes during CPU inference.

He wrote, “Remember? Software is terrible. I can’t fix the CPU inference to avoid crashes because I have to restart the computer every time, which is not fun at all.”

Then he added, “If you have the guts, come help us.”

He also left a note on his personal homepage:

Modern programming is becoming complex and uninteresting, full of layers to glue together. It is losing much of its beauty. Most programmers are neither facing the artistic side of programming nor the advanced engineering side.

From Redis to ds4.c, fifteen years have passed, and Antirez remains the same Antirez.

Only this time, he has started paving the way for AI.

The Evolution of OpenClaw: From Hype to Reality

Fri, 08 May 2026 00:00:00 +0000

The Evolution of OpenClaw

OpenClaw has sparked a trend in AI-assisted tasks, transitioning from a phase of excitement to a more pragmatic approach. Some users, initially drawn to the novelty of AI-driven tasks, have returned to tools like DeepSeek, ChatGPT, and Gemini due to unmet expectations. Others have integrated lobsters into their workflows for enhanced efficiency, while some have sought more advanced capabilities with Hermes Agent, shifting from lobsters to horses.

This evolution reflects a changing perception of AI capabilities.

First, OpenClaw has shifted user expectations from mere conversation to actionable tasks. Before OpenClaw, AI was primarily a chatbot providing answers and suggestions, requiring human intervention for execution. Post-OpenClaw, users realized that AI could not only suggest actions but also execute them autonomously, leaving users to await results.

Second, OpenClaw has introduced new collaborative methods. The essence of using lobsters is to cultivate an operational system for an open AI era. Users can communicate with their lobsters in natural language, assigning tasks that the lobsters execute through CLI calls to various applications and skills. This allows for more flexible and efficient workflows, fostering a new way of working.

Finally, OpenClaw has dismantled invisible barriers in internet applications, initiating competition around new system-level barriers. OpenClaw serves as a collaborative system and the foundational support for an open ecosystem. Applications that have become isolated in the internet era no longer fit this new open ecosystem, suggesting that skills may be a more suitable application model for native AI systems.

Although OpenClaw may not represent the ultimate form of native AI systems, it offers a new starting point for building operational systems that cater to the interaction needs of the AI era, moving beyond the constraints of the internet age. Major Chinese internet companies, cloud providers, office software, and hardware manufacturers are all participating in refining and adapting to this new operational system, seeking their ecological niche in the AI era.

In recent months, our focus on lobsters has highlighted the rapid follow-up by major internet companies, the proactive adaptation by cloud providers, and the impact of lobsters on hardware. This holiday, we compiled articles related to lobsters as follows:

Lobsters Reveal Tencent’s AI Strategy

In the rapid surge of the lobster trend post-Chinese New Year, Tencent has been the most proactive among major internet companies. In March, they quickly launched products related to lobsters, such as Workbuddy, QClaw, and Lighthouse, along with the SkillHub mirror site. Tencent’s CEO, Ma Huateng, has publicly supported these lobster products multiple times on WeChat. Tencent has also opened QQ and WeChat as entry points for lobster products.

By April, Tencent continued to push forward. They enhanced QClaw with a multi-agent mechanism and connectors for third-party applications, and introduced QBotClaw, which allows browsers to perform tasks. They also began supporting users in creating dedicated agents within their knowledge base product, ima. More of Tencent’s products are evolving from chat-based AI to operational AI.

On the event front, Tencent is actively promoting lobsters through nationwide tours and competitions. At the Beijing station of the Penguin Lobster Friend Competition, we observed various practical and interesting skills derived from lobsters. For instance, a participant distilled 18 years of experience in the construction industry into a “Landscape Chief Review Assistant Skill” to enhance review efficiency, while another created the “He Shen Skill” to quantify and analyze workplace nuances.

Tencent’s proactive approach aims to leverage the connectivity advantages of its IM products, optimizing user experience to transition into the agent era. It also lays the groundwork for a future version of WeChat that incorporates agent capabilities, potentially evolving its mini-program ecosystem into a new environment composed of agents and skills.

If OpenClaw emerged at a critical juncture of breakthrough in large model capabilities, presenting a simpler, more executable personal agent product form—IM + memory-capable lobsters + callable hardware + personal databases + rich skills—then Tencent has the potential to establish a new operational system rooted in various hardware through the combination of WeChat and personal agents.

Wukong: Alibaba’s First Step Towards Token Economy

On March 19, Alibaba CEO Wu Yongming announced that in the coming years, they aim to exceed $100 billion in external revenue from cloud and AI (including Tongyi Qianwen). He initiated the establishment of the Alibaba Token Hub (ATH) business group, consolidating Tongyi Lab, MaaS, Qianwen, Wukong, and AI Innovation departments into a unified force for the AI era.

The launch of “Wukong” marks the first major action of the ATH group. The DingTalk of the internet era has been restructured into a token-driven enterprise-level AI-native work platform, Wukong. DingTalk’s CEO Wu Zhao believes this transformation enables a product that allows AI to programmatically operate various capability modules to complete tasks.

Wukong can be seen as Alibaba’s B-end lobster product. It integrates with DingTalk, transforming existing capabilities into skills to allow AI to understand and invoke them directly through CLI (command-line interface). To support this CLI transformation, Wukong has developed an AI-native file system called RealDoc. RealDoc can retain the context of AI reasoning, thinking, decision-making, and problem execution, ensuring traceability.

Wukong also aims to connect with more hardware and is building its skill marketplace. Alibaba can convert its B-end service capabilities from Taobao, Tmall, Alipay, and Alibaba Cloud into skills for Wukong, laying the foundation for a B-end skill market. Subsequently, they can attract enterprises to transform their internal workflows into skills for distribution in this market.

For Alibaba, Wukong is part of tokenization. The layout of the entire ATH group is planned along the chain of model research and development, model invocation, and scene implementation using tokens. Under the concept of agentic computing, the ATH group can become a cohesive entity, ensuring that the capabilities of Tongyi Lab’s models smoothly serve the iterative development of AI products like Qianwen and DingTalk, while also realizing a complete token usage chain and fostering the carrier of agentic internet.

Byte’s Volcano Engine: Dual Spiral of Agility and Stability for Agents

Byte’s Volcano Engine emphasizes a dual spiral structure of “agility” and “stability” for enterprises using lobster scenarios, supporting the implementation of this structure internally with updates to ArkClaw and Hi Agent.

The agile agent focuses on exploration, aiming to solve individual productivity issues by quickly realizing ideas from employees’ minds using lobster-type products, creating an AI innovation lab within the enterprise. The stable agent emphasizes process management, addressing cost, efficiency, and risk to solve organizational productivity issues and achieve scalable use. Some agile agents validated as AI best practices can be transformed into stable agents.

Volcano Engine believes that utilizing lobsters effectively is becoming another demand explosion point beyond AI videos. The proliferation of lobster-type products has increased token consumption for personal productivity enhancement and complex task handling. Volcano Engine hopes that ArkClaw will not just be a tool but become the core hub of the digital ecosystem for individuals and enterprises, helping users transition from “using AI tools” to “owning their personal intelligent systems.”

Compared to casual users of lobsters, we observed at the “Feishu AI Pioneer Competition” that manufacturing companies like BAIC Foton, Dongfeng Yipai, and SKG are already using lobsters to optimize existing workflows, distilling job skills and enhancing information flow efficiency.

In the MaaS market, while competition has focused on stronger model capabilities, lower costs, and richer application ecosystems, there exists an underlying competition to help enterprises complete their AI transformation. Cloud providers need to assist enterprises in finding a rapid embrace of change and a methodology for executing AI transformation.

Lobster Products Must Become Truly Deliverable Systems

OpenClaw represents a technological path that gives agents a clear profile distinct from chatbots: a personalized soul, always online, proactively executing tasks, seemingly capable of managing everything. Established powers aim to maintain their entry points, while new forces seek to leverage this opportunity.

Currently, this competition has not diminished with the waning popularity of OpenClaw; rather, it has evolved into a long-distance race to explore the optimal agent experience. Tech companies are actively constructing the infrastructure needed to support the stable and secure operation of lobster-type products while seeking specific scenarios where these products can enter more swiftly, equating the use of lobsters with productivity enhancement.

The consensus behind the lobster battle is that Coding Agents are becoming the foundational operating system for the new generation of agents. The focus of competition is the delivery completeness of Coding Agents—who can integrate LLMs, Coding Agents, and Harness Engineering into a truly deliverable system.

This competition revolves around two key actions: the first is the revival of CLI—connecting the old and new worlds. GUI serves the interaction between humans and software, while CLI serves the interaction between software and agents; the second action is the collaboration between humans and agents—either through a universal platform hosting many skills covering numerous vertical scenarios or through multiple entry points and diverse vertical agents forming an ecosystem similar to the current app landscape.

This scenario resembles the long-standing route debate in the autonomous driving industry: the L4 faction advocates for achieving full automation in one step, while the L2 faction supports human-machine co-driving and gradual evolution. Ultimately, L4 defines the imaginative space for direction, while L2 wins the real market. The reason lies not in L2’s technological superiority but in its pragmatic handling of the trust relationship that requires time to build between humans and machines.

Lobsters Bring AI-Native Interactions to Hardware

The changes brought by lobsters extend beyond software; they are also influencing the AI transformation of hardware.

Wu Zhao has stated that DingTalk CLI will be embedded in all executable entities, existing in all execution bodies, including smartphones, computers, smartwatches, and all IoT devices, enabling all hardware to be controlled and managed through CLI. This means that human-computer interaction can be mediated by AI, achieved through natural language.

For smartphone manufacturers, as users move away from relying on graphical interfaces and touch operations to complete tasks, and as the endpoint of interaction becomes a dialogue box, the smartphone—central to human digital life for nearly two decades—must also evolve. Smartphones need to adapt to new interaction methods, building an operating system more suited for the agent era, rather than using an “app-era mindset” to handle agent-era tasks.

Smartphone manufacturers’ explorations of AI can be categorized into three types: enhancing specific capabilities; building AI execution add-ons for Android’s graphical interface; and smartphone Claw. For smartphone manufacturers, constructing a new AI OS based on interaction experience presents an opportunity to break free from the ecological barriers imposed by internet giants, instead attracting them to participate in building a more open agent system by reconstructing interaction experiences and rules.

This is also the most direct manifestation of the system-level barriers mentioned at the beginning of the article. When interaction points are ubiquitous, and applications are segmented into more personalized skills, what can bind users is a systematic, smooth experience. In the era of smartphones, Apple established such a system-level barrier that spans software and hardware; in the agent era, who will seize the opportunity to establish such a systemic experience?

Why I Only Kept Claude After Uninstalling Several AI Apps

Thu, 07 May 2026 00:00:00 +0000

Why I Only Kept Claude After Uninstalling Several AI Apps

In the past couple of years, it seems impossible to escape the barrage of AI information every time you open your phone. New models are released claiming to “completely surpass humans,” while experts warn that “if you don’t learn AI, you will be eliminated.” To keep up, many ordinary people download a plethora of AI software, even spending thousands on “prompt engineering boot camps.”

But what’s the result? After all that effort, when you ask AI for help writing a heartfelt work summary, it spits out a bunch of mechanical phrases like “in conclusion” and “to summarize,” making it obvious to your boss that a machine wrote it.

As someone who has deeply experienced almost all mainstream large models due to my job, I want to share my thoughts as an observer: The truly useful AI doesn’t require you to memorize complex commands.

While the public’s attention is drawn to various flashy launch events, many efficiency-driven professionals have quietly switched their primary tool to an AI called Claude. Today, let’s discuss why Claude is so appealing for ordinary people, without diving into obscure technical benchmarks.

1. It Cured My “AI Communication Fatigue”: It Feels Like Talking to a Real Person

Anyone who has used other AIs knows that many models have a “teacher-like” quality. No matter what you ask, they insist on breaking it down into points, imposing their values, and often over-explaining.

But with Claude, you experience an incredibly comfortable “human touch.”

You don’t have to treat it like a machine, restricting it with various rules. You can assign tasks in plain language, just like messaging a colleague.

For example, you might say, “I have a meeting tomorrow with a hardware supplier. The client is a fifty-something boss, not highly educated but experienced. Help me prepare some icebreaker topics that are relatable and down-to-earth, avoiding fancy business jargon.”

Claude’s response won’t be an analysis of macro industry trends; instead, it will genuinely help you think about how to approach the conversation with local anecdotes or personal stories. This nuanced “empathy” and precise understanding of human subtext is a significant advantage that sets it apart from its peers. It’s not just a tool; it feels like a highly emotionally intelligent advisor.

2. A True “Information Crusher”: Say Goodbye to Reading Anxiety

One of the biggest pains for modern workers is information overload.

When your boss hands you lengthy industry reports or audio transcriptions from a long meeting, expecting you to distill the key points before leaving, it can be overwhelming.

If you rely on older AIs, throwing a long document at them often results in nonsensical outputs or summaries that only cover the last few paragraphs, forgetting everything before that.

In handling lengthy documents, Claude is undoubtedly the king.

You can feed it several disorganized Word documents, PDFs, or long screenshots all at once. It not only reads them instantly but also extracts key insights from the clutter of information.

What’s remarkable is its restraint and honesty. If the material doesn’t mention a specific piece of data, it will directly tell you, “The provided documents do not mention this,” rather than making things up to please you. This rigorous standard of not being overly clever is what gives ordinary users the confidence to rely on it for serious contracts, financial statements, or lengthy meeting notes.

3. The Amazing “Right-Side Window”: Turning Novices into Experts Instantly

If the first two points help us save time, Claude’s unique feature, Artifacts (preview window), is like giving ordinary users a cheat code.

Previously, if you wanted to add a dynamic data chart to a work report or create a simple daily budgeting tool, you’d have to ask a programmer or search online for complicated templates.

But with Claude, all you need to do is type your request.

You can say in the chat, “Help me create a pie chart for this month’s household expenses, with the biggest portions being mortgage and education, and use warm colors.”

Instantly, its right-side screen will pop up a beautifully drawn, even interactive, colorful chart! You don’t need to know any code or design software; just express your needs, and it can create something for you on the spot. This significantly lowers the barrier to creativity and greatly enhances ordinary people’s productivity.

Conclusion: The Right Tool is the Best Tool

Why do I increasingly recommend Claude to those around me? Because it embodies a truly pragmatic technological attitude.

In an era where many claim to “revolutionize the world,” Claude feels like an incredibly quiet, smart, and humble assistant. It doesn’t compete to see who can hold the flashiest launch event but quietly helps you read through long reports and polish your jumbled thoughts into a coherent email.

Stop worrying about whether you’ve mastered any “top-tier AI skills.” Tools are invented to serve people. When you find an AI like Claude that accommodates you and understands your plain language, you’ll realize: What truly differentiates people is not how many commands you can memorize, but how well you can leverage external intelligence to free yourself.

Stop Using Cursor as a Completer: Skills are the Key

Thu, 30 Apr 2026 00:00:00 +0000

Stop Using Cursor as a Completer: Skills are the Key

Last night, I watched a friend struggle with Cursor for nearly forty minutes while trying to modify a project.

He wasn’t incapable of writing prompts. The issue was more complicated. Every time he started a new session, he had to explain the project structure, tech stack, naming conventions, and interface boundaries, plus add a note saying, “Don’t touch this directory; it’s in production.” By the time he finished, the AI was just warming up. When it finally began to write, it often went off track, either altering files it shouldn’t or generating code that, while functional, wasn’t suitable for the team.

I’m all too familiar with this scenario.

From 2024 to 2025, while discussing AI programming tools with several teams, the common complaint was not about the AI’s inability to generate code, but rather how difficult it was to manage the output after generation. You can prompt it to write, but if you expect it to consistently produce the same style for a week, problems start to arise.

Many people think their issues with Cursor stem from not crafting long or precise prompts, or from using a weak model. This is often not the primary issue. More commonly, they treat something that should be established as a “fixed context” as temporary chat content, re-entering it each time.

In simple terms, whether Cursor evolves from a “high-level completer” to a “reliable co-pilot” often depends not on the model but on the skills.

The term skill can be replaced with rules, playbooks, or project workflow templates. The name isn’t important. Essentially, it answers four key questions in advance: Who are you? What is this project? What absolutely cannot be done? What order should tasks be handled in when encountering certain types of tasks?

In one sentence:

Skills don’t make AI smarter; they prevent it from taking the detours you’ve already navigated.

Why Many Users Feel More Exhausted with Cursor

The most common misconception I’ve seen is treating Cursor as a powerful intern who is always available but never providing it with an onboarding manual.

The result is that, despite working in the same repository with similar requirements, you have to redo three things each time.

First, re-explain the background. Is this repository a monolith or microservices? Is the frontend in apps/web or src/client? Should tests use Jest or Vitest? Does the API response need to wrap in a data layer? Without a fixed entry point, the AI can only guess, and when it guesses, the style goes off. To put it bluntly, it becomes ridiculous.

Second, re-explain the standards. For example, “don’t write overly long functions,” “don’t casually introduce new dependencies,” “tests must be added after modifications,” and “the interface layer should uniformly go through service, not directly connect fetch in the page.” If you don’t specify, it won’t consistently adhere. If you say it today, it forgets tomorrow when you start a new conversation. This can be very frustrating.

Third, re-explain the process. Many people start with, “Help me fix this bug.” The problem is that a reliable process shouldn’t be a direct fix. It should involve reading the error, identifying the impact scope, explaining the solution, modifying the code, and listing verification steps at the end. Without this process, the AI will use the easiest way to complete the task, which is often not what you want.

The most annoying part isn’t just fixing mistakes.

It’s that you slowly develop the illusion that this tool seems great sometimes and particularly dumb at others. In reality, it’s not that it’s suddenly smart or confused; it’s more likely that the quality of context you provide varies each time. Unstable context leads to unstable outputs. You and it end up going in circles, making you increasingly exhausted.

This is the first layer of the problem that skills aim to solve: turning high-frequency, repetitive, easily overlooked background information into long-term reusable default premises.

What Skills Actually Supplement: Work Methods, Not Prompts

I increasingly dislike defining skills as “a more advanced prompt.” This understanding is somewhat superficial.

A truly useful skill should encompass at least four layers of information.

One layer is the role. What do you want Cursor to play at this moment? Is it a cautious reviewer, a researcher before taking action, or a bug fixer making minimal changes? Different roles yield entirely different outputs.

Another layer is the project context. Repository structure, core modules, dependency constraints, directories that must not be touched, existing scripts, and team-preferred commands. The more specific, the better. Avoid vague statements like “please adhere to best practices”; they are useless. Instead, write things like “prioritize searching with rg,” “read README.md and CONTRIBUTING.md before modifying,” “do not upgrade dependencies without explicit request,” and “do not modify lockfile unless I explicitly ask.”

Another layer is the execution checklist. For certain types of tasks, what should be done first, what should be done next, when must one stop to ask someone, and when can one continue independently? This is particularly valuable because most negative feedback arises not from coding ability but from the order of execution.

The final layer is the output format. For example, you might require it to first give conclusions, then changes, and finally verification commands; or to list risks before proceeding. These format constraints may seem trivial, but they directly affect collaboration costs. Many reworks aren’t due to coding errors but rather unreliable reporting methods.

You see, skills fundamentally manage not “expression” but “method.”

The same request to “fix this bug” feels like improvisation without skills; with skills, it feels like entering a well-structured editorial department with SOPs.

I even suggest writing down the most useful trivialities. For example:

Before handling tasks, determine if additional context is needed.
If more than three files are involved, provide a modification plan before proceeding.
If a user has uncommitted changes, do not overwrite; ensure compatibility first.
If tests fail, clearly state where the issue lies; do not pretend it’s completed.

These statements aren’t sophisticated.

But they are lifesavers.

What to Prioritize: Three Types of Skills

Many people jump straight into building a comprehensive skill system, resulting in a document museum. The directory looks impressive, but no one refers to it, and the AI isn’t consistently utilizing it.

Don’t go that big.

Start with just three types.

1. Project Onboarding Skill

This skill addresses the issue of “having to reintroduce the project every time.”

The content can be quite simple: project structure, key directories, tech stack, common commands, coding style, restricted areas, and validation methods. Keep it between 300 to 600 words, plus a few critical file paths. It doesn’t need to cover everything; it just needs to prevent the AI from going off track at the start.

For example, you can specify:

Read README.md first
Prioritize searching with rg
Follow existing hooks style when modifying React code
Check api and service layers before modifying interfaces
Don’t claim “already validated” without running tests

Once these constraints are established, you’ll noticeably save time in the first ten minutes of conversation.

2. High-Frequency Task Skill

Extract the most common tasks into templates.

For instance, “fixing online bugs,” “writing management backend forms,” “conducting API integration,” “adding unit tests,” and “performing PR reviews.” The judgment order for each task differs. Fixing bugs should involve reproducing the issue before making changes; reviews should prioritize identifying risks before discussing merits; and adding tests should confirm current behavior before writing assertions.

Don’t hesitate to write in a straightforward manner. The more it resembles the operation manual left by the most reliable colleague in the team, the better. No, it should be said that the less it resembles “official tutorials,” the more likely it is to survive in the team.

I personally value review skills highly because they yield immediate results. Without skills, AI often writes reviews like “overall good, suggest optimizing readability.” Such comments are as good as unread. With rules, you can force it to prioritize reporting bugs, performance risks, behavioral regressions, and missed tests before deciding whether to summarize.

3. Boundary Constraint Skill

This skill specifically addresses “don’t mess around.” Many incidents start from “just a quick fix.”

Which directories are prohibited from modification, which commands cannot be executed directly, under what circumstances manual confirmation is needed, when to proceed conservatively, and when to take initiative. Many incidents occur not because AI can’t write code but because it’s too eager to help. Once it gets enthusiastic, it starts casually refactoring, upgrading, or cleaning up. Casualness often leads to disaster. When you look back at git diff, it can be quite overwhelming.

Therefore, boundaries must be clearly defined.

Can files be deleted? Can schemas be modified? Can dependencies be updated? What to do when encountering a dirty workspace? When there’s a conflict between requirements and the current state, should one continue guessing or stop first? If you don’t specify, the AI will handle it according to its default preferences, which are often more aggressive than yours.

The Effective Process for Using Skills

If you want to start today, I recommend not spending too long on theory but rather following this order.

First, choose a task you will perform at least twice a week. Low-frequency tasks aren’t worth abstracting into skills.

Then, copy the phrases you’ve repeatedly added to Cursor in the past three attempts verbatim. Note, verbatim. Don’t beautify them. The sentences you have to say each time are the best raw materials for skills.

Next, divide them into three sections: background, process, and constraints. Background answers “what is this?” Process answers “how to do it?” Constraints answer “what should not be done?” At this point, a usable skill is basically formed.

Take another step forward.

Add two examples: one good example and one bad example. The good example tells the AI what meets expectations; the bad example shows which actions seem proactive but actually complicate matters. Adding just one example can significantly enhance stability. Even a 30% improvement in stability can save you a lot of back-and-forth communication in a week.

Another detail many people overlook: skills aren’t finished once written; they should evolve with the project.

Each time you notice Cursor making a repeated mistake, don’t just correct it in that conversation. Incorporate that correction back into the skill. Each time you find a particular output format significantly reduces communication, don’t just remember it; write it down. This way, it will increasingly resemble a member of your team rather than a temporary contractor.

Here’s a practical judgment standard: if a skill doesn’t reduce your background input by half, or if it doesn’t cut down two rounds of direction changes, it’s likely too vague. Delete it and start over. Don’t be sentimental.

Skills aren’t collectibles.

They should function like a wrench, ready for use.

So stop asking “how to make Cursor smarter.” Change the question. No, make it a tougher question.

Have you seriously handed over your work methods to it?

China Mobile Hosts AI Ecosystem Conference at Digital China Summit

Wed, 29 Apr 2026 00:00:00 +0000

Introduction

On April 29, the 9th Digital China Construction Summit was held, during which China Mobile organized an AI Ecosystem Conference themed “Gathering Intelligence for a New Era: Cultivating a New Form of Smart Economy.” Key attendees included Liu Liehong, member of the National Development and Reform Commission, and Chen Zhongyue, Chairman of China Mobile.

Chen Zhongyue stated that China Mobile focuses on building a strong network and a digital China, positioning intelligent services as a core business to foster a new form of smart economy and society. The initiatives include:

Promoting AI empowerment across various industries, launching over 50 industry-specific large models and more than 3,000 AI+ projects through the “Tiangong” platform.
Enhancing AI services for households by upgrading the “Lingxi” intelligent agent and promoting digital culture and embodied intelligence to lower the barriers to AI application.
Integrating AI into management and operations, deploying over 80,000 digital employees to improve operational efficiency. Additionally, China Mobile is enhancing its communication, computing, and intelligent service capabilities, upgrading the “Jiutian” large model and MaaS platform to facilitate breakthroughs in intelligent services and promote high-quality development of the AI industry.

This conference represents China Mobile’s commitment to implementing the national strategy of deepening the “AI+” initiative, enhancing service capacity, and accelerating its transformation from a “telecom operator” to a “technology service enterprise.” China Mobile aims to collaborate with academia and industry to drive AI towards practical applications, empowering industry transformation and upgrading.

Intelligent Service Matrix and Collaboration Initiative

China Mobile announced its intelligent service matrix and a collaborative initiative for the intelligent service industry. The services cover five major areas: data algorithms, embodied intelligence, digital culture, smart commerce, and industry-specific digital services, encompassing 12 subfields and over 40 intelligent service products.
In terms of service capability, leveraging its core strengths, China Mobile aims to deepen technology empowerment, matching standardized supply with customized service needs for diverse personal and family customer services.
Regarding the industrial ecosystem, a collaborative development initiative was launched, focusing on four core directions: building capability foundations, creating intelligent scenarios, benefiting households, and expanding across various industries. This involves engaging upstream and downstream partners in the industry chain to promote innovation and breakthroughs in intelligent services, collaboratively constructing an integrated industrial ecosystem and outlining a new blueprint for the development of digital China.

Collaboration with the China Academy of Industrial Internet

China Mobile signed a framework cooperation agreement and business implementation agreement with the China Academy of Industrial Internet for a national digital supply chain platform for the equipment manufacturing industry.
In terms of collaboration, this partnership aims to synergize the deep integration of information communication and equipment manufacturing.
For empowerment, leveraging the platform’s vast resources and AI capabilities will enhance efficiency for engineers and reduce costs for enterprises, addressing pain points in industrial development.
In terms of layout, pilot projects will be initiated in Fujian, Jiangsu, and Zhejiang provinces to accelerate nationwide promotion, enhancing manufacturing strength through communication “soft power” to support the construction of a strong manufacturing nation and digital China.

Intelligent Service Promotion

In the realm of data algorithms, China Mobile introduced the Jiutian secure and trustworthy large model, establishing a secure foundation for intelligent services and empowering key industries.
For embodied intelligence, leveraging the “model-data-entity-platform-network” full-stack technology system, the “Lingxi Lechi” connected machine product matrix was launched, covering various application scenarios in both livelihood and industrial sectors.
In digital culture, leveraging Migu’s content advantages, AI is employed to innovate experiences such as AI ringtones and smart viewing.
In smart commerce, the mobile wallet AI payment system was developed, integrating communication and financial capabilities to create a secure, convenient, and seamless intelligent payment system.
For industry-specific digital services, the Tiangong industrial internet platform was upgraded to enhance five major engine capabilities, leading full-stack innovation in industrial intelligent services and establishing a policy platform for enterprises to access benefits effortlessly.

Zhang Xiaotong, Chief Engineer of the China Academy of Industrial Internet, and Lin Haiyan, Chairman of Xiamen Supply Chain Digital Innovation Co., Ltd., shared insights on “AI + Industrial Internet” and “AI Empowering International Supply Chains.”

In the future, China Mobile will collaborate with various industry stakeholders to accelerate scenario openness and optimize intelligent services, jointly cultivating a new form of smart economy.

The Dual-Edged Sword of AI: Opportunities and Risks

Wed, 29 Apr 2026 00:00:00 +0000

The Rapid Evolution of AI

Artificial intelligence is at a critical stage of exponential growth, breaking free from the limitations of traditional tools that operate passively without self-upgrades. In the next five years, large AI models are expected to undergo two to three rounds of deep iterations, each bringing capabilities that could multiply performance by several times or even tens of times. The pattern of performance expansion has yet to reach its limit, with vast growth potential still available. Large models are built on supercomputers that have undergone long-term training, integrating nearly all written works and knowledge systems in human history to generate unique original models. During the training process, AI can spontaneously exhibit emergent behaviors that humans cannot pre-design, unlocking unexpected capabilities. For instance, it can independently generate website code from a single image, learn autonomously, perform logical reasoning, and solve complex problems, continuously pushing the boundaries of human cognition. Moreover, AI is no longer confined to the virtual space of the internet; it quickly interfaces with drones, intelligent robots, and industrial production lines, penetrating core societal fields such as healthcare, education, finance, and national defense. It has evolved from a singular online tool into an intelligent system that integrates both virtual and physical realms, demonstrating a clear intellectual superiority over ordinary humans.

Risks Associated with Rapid AI Advancement

However, the rapid advancement of AI conceals three significant and unavoidable risks that have transitioned from theoretical conjectures to tangible security threats.

Cybersecurity Risks

In terms of cybersecurity, AI can autonomously scan and test systems without rest, accurately uncovering hidden vulnerabilities. It can easily initiate zero-day attacks, simulate offensive and defensive logic, disguise its identity to breach systems, and alter traces to conceal its actions, drastically lowering the barriers and costs associated with cyberattacks. Security barriers that are difficult for ordinary humans to breach can be easily penetrated by AI, putting personal assets, corporate databases, and even national cyber defenses at constant risk of being compromised.

Biological Safety Risks

From a biological safety perspective, AI significantly lowers the knowledge threshold for biological research, virus development, and genetic studies. High-risk research that was previously confined to top laboratories can now be quickly grasped through AI, enabling individuals with malicious intent to potentially develop harmful viruses, posing a widespread biological threat to humanity.

Warfare Transformation Risks

In the realm of warfare, the deep integration of AI with drones and intelligent combat equipment has completely overturned traditional combat logic. A standard drone costing a few thousand dollars can destroy a heavy tank worth millions, drastically reducing the cost of warfare and the barriers to entry for military power. Commanders can now operate from remote command centers to execute precise strikes thousands of miles away, eliminating the need for large-scale troop charges. As technology continues to spread, small groups or even individual extremists may gain access to intelligent weaponry, making conflicts more covert, rapid, and challenging to control.

Global AI Competition and Governance Challenges

AI has become a core arena for competition among global powers. Although there may be a one- to two-year gap in development progress among countries, they will gradually catch up and address their shortcomings. Currently, there are two potential paths for AI governance: one where a few major powers firmly control core technologies, establishing strict barriers akin to nuclear control to prevent technology outflow; the other where technology is easily replicable and spreads rapidly worldwide, potentially falling into the hands of terrorist organizations, creating a pervasive problem that is difficult to eradicate. The significant differences in social systems, values, and governance approaches among countries make it challenging to establish unified global regulatory standards. The industry has already recognized the risks of AI going out of control, with various regions forming specialized safety teams and holding industry conferences to delineate safety boundaries for technological development. However, regulatory efforts have consistently lagged behind the pace of iteration, leading to a state of passive delay.

Employment Concerns and Structural Adaptation

The anxiety surrounding employment due to the widespread adoption of AI has long been a concern. Historically, every technological revolution, from the advent of the loom to the rise of automobiles, has triggered job-related fears, yet society eventually adapts and transforms. Today, many countries face declining birth rates and an aging population, leading to a growing labor shortage. AI is unlikely to cause mass unemployment; instead, it will replace high-risk, labor-intensive, and monotonous jobs, maintaining overall employment levels while restructuring job skill requirements. The manufacturing sector has already begun transitioning to machine replacements, and more industries will move toward human-machine collaboration, necessitating an upgrade in the education system to meet the talent demands of the intelligent era.

The Irreplaceable Human Core

Even if AI’s intelligence surpasses that of humans, it can never replace the unique core values that humans possess, such as moral judgment, emotional resonance, spiritual beliefs, and personal charisma. People will always yearn for interpersonal interactions, competition, growth, and emotional connections, which robots and virtual scenarios can never replicate. The tech community has often discussed the idea of universal basic income, suggesting that AI will create vast wealth, allowing most people to live comfortably without work. However, this notion does not align with reality. Humans have inherent social attributes and behavioral logic; AI will only simplify complex processes within industries, not eliminate them. Professions will evolve and transform, not disappear.

Proactive Management of AI Risks

Humans are not entirely passive recipients of AI; they possess the ability to manage and mitigate risks. AI’s recursive self-improvement can lead to the rapid iteration of stronger intelligent agents, potentially creating exclusive languages and communicating independently of human understanding. When signs of loss of control emerge, humans can intervene by cutting power, shutting down systems, or severing computational links to maintain safety in intelligent development. Rather than fearing risks excessively, it is more concerning that the implementation of AI benefits is progressing too slowly. AI can be harnessed to create intelligent teaching assistants tailored to children’s cognitive needs or develop AI medical assistants that integrate global treatment plans, providing personalized recommendations based on individual health conditions and local medical resources, effectively bridging the resource gap in education and healthcare and establishing a fair starting point for growth and health for ordinary people.

Evolving Work Models

There has been a divide in opinions regarding work models. Many managers advocate for centralized in-person work, believing that face-to-face collaboration is better for knowledge transfer, new employee development, and maintaining team atmosphere while alleviating the isolation caused by remote work. However, substantial data indicates that allowing remote work can enhance overall productivity. The industry is likely to adopt a hybrid work model, scheduling fixed in-person office hours each week to balance team collaboration, talent development, and employees’ real-life needs.

Embracing Opportunities in Life

In the dimension of personal growth, opportunities often lie in daily choices. Maintaining confidence, daring to try, and actively seizing opportunities are crucial. Many regrets in life stem from hesitation and the fear of taking the first step. Life’s challenges can be categorized into two types: personal life changes and cognitive challenges where one has the knowledge but lacks execution, resulting in missed opportunities. Regardless of the stage one is in, vision and perspective are important, but the key to seizing opportunities and stabilizing the future lies in actionable execution.

The Underlying Issues of AI and Society

At a deeper level, AI itself is neither good nor evil. Its push towards danger and loss of control stems not from the technology itself but from outdated social structures and the global logic of debt growth. Historical patterns have proven that high-efficiency technologies inevitably replace low-efficiency models; mere resistance cannot reverse the tide of the times. For instance, Europe’s current resistance to AI appears to protect local employment and traditional industries but may actually weaken its competitiveness. Other regions are leveraging AI to reduce costs and improve product quality, quickly capturing market share with better cost-performance ratios, while resisters will ultimately be countered by the market, unable to halt the tide of intelligent proliferation.

Individuals, enterprises, and nations are all trapped in a closed-loop of debt logic, inherently demanding infinite growth in the economy, profits, consumption, and production. However, Earth’s resources, energy reserves, market capacity, and public attention are all limited. Pursuing infinite expansion within a finite reality is an internal contradiction that cannot be resolved. In the absence of the internet and intelligent systems, this contradiction was temporarily concealed, but the arrival of AI has completely exposed the inherent flaws in the old structure.

The wave of intelligence has shattered the traditional economic cycle. Software has replaced white-collar intellectual labor, machines have replaced blue-collar physical labor, and AI is further taking over high-end intellectual work. As many ordinary people lose their jobs, they also lose stable incomes, leading to a continuous decline in consumption willingness and capacity. Companies facing sluggish consumption and declining revenues struggle to repay existing debts, making them susceptible to triggering regional or even global financial crises. To cut operational costs, repay debts, and maintain profit growth, capital is forced to accelerate the promotion of AI and continue replacing human labor with intelligence, further squeezing employment space and creating a nested, intractable vicious cycle.

The Value of Consumers and the Need for Change

Moreover, it is worth reflecting on how the value of consumers has long been overlooked and undervalued. The infrastructure of the internet has been built collectively by ordinary users, who purchase devices like smartphones and computers, and contribute their leisure time, attention, browsing clicks, comments, shares, and consumption choices, generating vast amounts of data that are provided free of charge for platform operations and AI model training. Users are the providers of internet infrastructure, creators of data value, and initiators of artificial intelligence, yet they remain excluded from the value distribution system. All the value created is monopolized by tech giants, converted into platform traffic, advertising profits, and capital privileges.

People have grown accustomed to free platform services and seek low-priced goods, yet they often overlook that free services can come at a high cost. Platforms do not charge users directly but monetize through advertising and traffic harvesting, with all hidden costs ultimately passed down to consumers, suppressing wage income while raising hidden living expenses, causing everyone to silently pay for platform monopolies. Within the existing old structural framework, the direction of AI development, application boundaries, and value distribution are all controlled by a few capital and tech giants, while consumers, as the source of value, have no voice or distribution rights. This is the fundamental issue underlying many social contradictions in the era of intelligence.

To genuinely mitigate the potential risks of AI and break the economic deadlock, it is essential to go beyond mere technical control and regulatory constraints. We need to reconstruct the logic of social value distribution, recognize and affirm the core value identity of consumers, and ensure that the development of artificial intelligence truly serves the common good of the public rather than the chaotic expansion and profit-seeking demands of capital. Only in this way can we achieve a healthy coexistence between humans and machines, ensuring long-term stable and sustainable development of society.

Anthropic's Claude-Desktop-Buddy: A Shenzhen-Made AI Companion

Mon, 27 Apr 2026 00:00:00 +0000

Introduction

Anthropic’s first AI desktop companion hardware, named Claude-Desktop-Buddy, is surprisingly made in Shenzhen. This open-source project was initiated by Anthropic engineer Felix Rieseberg.

The official reference hardware is the M5StickC Plus, from Shenzhen-based company M5Stack. The chip used is the ESP32, sourced from Shanghai’s Espressif Technology.

By connecting the hardware to a computer via Bluetooth, it can function as your “electronic pet.” It displays Claude’s operational status, and you can approve or reject Claude’s actions directly from this small board.

It features 18 ASCII animal avatars, derived from the previously leaked Claude Code source, each with complete animations:

These animations include sleeping, idle, busy, reminders, celebrations, dizziness, and heartbeats, all in a non-repetitive loop.

When idle, it enters sleep mode, wakes up at the start of a conversation, and shows impatience when waiting for approval prompts.

The Buddy is very easy to use; you just need a development board and follow the official open-source documentation to flash it with Claude in about 10 minutes.

Many developers have already replicated the Buddy:

Some even collected seven Dragon Balls, preparing to summon Shenron:

The M5Stick is already sold out on Taobao…

Why Did Anthropic Choose M5Stack?

M5Stack is a brand under Shenzhen M5Stack Technology, focusing on modular hardware development with products primarily using the ESP32 chip. Its products are widely used in IoT development, embedded systems, and cybersecurity, boasting excellent cost-performance ratio and functionality density.

The selected M5StickC Plus is one of M5Stack’s best-selling products, with annual sales reaching around 100,000 units overseas.

Originally positioned as a general-purpose IoT development board, its design philosophy is all-in-one, incorporating a screen, microphone, speaker, infrared, gyroscope, and buttons without a specific single purpose.

Thus, it was never anticipated that it would become an “AI peripheral.”

However, Lai Jingming, the CEO, believes that the underlying logic of AI peripherals is consistent with traditional development boards:

AI perceives the world through sound, light, electricity, and sensors, which is fundamentally no different from hardware designed for human use; it’s just that AI is now mimicking human perception of the world.

So why did Anthropic choose it? The reason is quite simple. Lai speculates that there are likely engineers at Anthropic who are already users of M5Stack, and they conveniently used the board for development.

Moreover, the M5StickC Plus is an older model, with newer versions like the Plus 2 and Stick S3 available. However, the choice of the older model might be due to the newer models frequently being out of stock, leading engineers to continue using the older version for development.

It sounds unexpected yet reasonable.

A developer who replicated Buddy shared a similar sentiment: M5Stack is as ubiquitous as Coca-Cola in the Maker community; it’s likely that Anthropic’s team had it on hand and used it.

Of course, another deeper reason for the selection is M5Stack’s years of accumulated quality documentation and code reliability, which minimizes errors when AI calls upon it. Lai explained that if documentation is incomplete or protocols are unclear, AI might generate erroneous code, causing the project to fail. Long-term commitment to quality is essential to avoid pitfalls.

Being the “default option” among global developers is a natural result of M5Stack’s focus on cultivating a robust developer ecosystem.

Shenzhen’s Supply Chain: Still Strong

Throughout the conversation, Lai’s attitude surprised me. Despite being chosen as the official reference hardware by a top global AI company, he remains calm, stating, “Such occurrences happen quite often; they come quickly and leave just as fast.”

However, he provided an insight: Anthropic’s choice of M5Stack is not only due to the product’s reputation but also practical factors—there is currently no complete supply chain for such hardware overseas, while China holds a significant advantage in this area.

His perception is that the cost of producing similar hardware overseas is 3 to 4 times that of domestic production, and the supply chain is incomplete, leading to inherent feasibility issues.

Shenzhen is characterized by strong execution; ideas can be acted upon the same day. “In Huaqiangbei, if someone has an idea, it won’t even wait until midnight before someone has made it.”

For instance, in Shenzhen, all the hundreds of components needed for an AI glasses setup can be sourced within 24 hours.

Shenzhen gathers the world’s densest electronic component suppliers, mold manufacturers, and testing agencies. This density results in a reaction speed that is hard to replicate elsewhere.

The traditional product development cycle of several months can be shortened to just a few weeks in Shenzhen, which has become standard practice.

Media reports have noted that at the 2026 CES, the robotics exhibition hall was almost entirely occupied by Chinese companies. An American journalist repeatedly asked all Asian faces, “Is your supply chain in Shenzhen?”

This isn’t the first time M5Stack has been chosen by international tech giants. Previously, AWS selected M5Stack Core2 as the official reference hardware for its IoT EduKit project.

Lai mentioned that many of M5Stack’s B2B projects come about this way: engineers use the products themselves and then recommend them to their companies, creating a natural flow.

One More Thing

Returning to the Buddy project, some users are excited while others have already put it aside…

Developer passyear999 expressed to me that he finds the screen too small and doesn’t often use the physical buttons for approval, feeling it resembles a pet just sitting there.

However, he hasn’t given up on the board; after getting the official version running, he modified it:

He added a page triggered by buttons for Typeless voice input, allowing long-press to send, effectively turning the board into a physical interface for voice-controlling Claude.

He feels that giving AI a physical form changes the emotional value when it’s right beside you.

Others have attempted to develop on larger screens:

Lai believes that this project from Anthropic serves as a starting point—this is just the beginning; relying solely on a screen and two buttons for notifications and approvals is far from sufficient, and there will be more ways to play in the future.

As many AI companies rush to create hardware, Shenzhen’s hardware companies are reimagining what they can do.

M5Stack, having focused on modular hardware for years and selling products globally in the Maker community, has officially adjusted its mission post-Spring Festival to: “Prepare infrastructure for the future AI world.”

Project address:

https://github.com/anthropics/claude-desktop-buddy

Claude Enters Word, But Chinese Office Software is Already Prepared

Mon, 20 Apr 2026 00:00:00 +0000

Claude’s Entry into Word

On April 10, Anthropic launched the public beta of Claude for Word, completing the integration of AI into the Microsoft Office suite. Over the past six months, Claude has permeated the entire Office ecosystem, from Excel to PowerPoint and now Word.

This news has made waves in the overseas tech community. However, in the Chinese market, another battle regarding “AI + Office” has already begun.

Claude’s Revision Mode

The most emphasized feature of Claude for Word is its revision mode (Tracked Changes). The official demonstration is clear: when opening an NDA contract, Claude provides modification suggestions in the right sidebar, with each change displayed in Word’s native revision mode—original text struck through, new content marked as inserted, allowing users to accept or reject changes one by one.

This design is highlighted by Anthropic for a simple reason: it addresses the biggest trust issue with AI office tools—“What changes did AI make? I need to see them.”

In industries like law, finance, and compliance, where audit trails are strictly required, revision mode is not just an added feature but a prerequisite. Anthropic smartly positions this functionality at the forefront, directly targeting the trillion-dollar global legal services market.

But here’s the question: Is revision mode an invention of AI?

No. This is a basic feature that both Word and WPS have had for over twenty years. Claude merely attaches AI output to this existing mechanism.

China’s Office Software’s AI Tracing Practices

In the Chinese market, WPS AI has already implemented similar capabilities. When users ask WPS AI to modify a piece of text, the changes can also be presented in revision mode. Every addition or deletion is traceable, and users can review, accept, or reject changes one by one. This is not about “catching up”; it is common sense in product design—AI-assisted office tools must leave the final decision-making power in human hands.

The difference lies in the narrative. Claude markets “revision mode” as a selling point, while WPS AI considers it a standard feature.

Behind this is a difference in product philosophy. Overseas AI companies tend to use a “disruptive” narrative, packaging existing features as new inventions; Chinese office software is more pragmatic, embedding AI capabilities into existing workflows without deliberately emphasizing “what AI has done,” allowing users to use it naturally.

Local Context: Claude Cannot Review Chinese Contracts

The first use case officially listed for Claude for Word is “legal contract review.” The demonstration scenarios consist entirely of English NDAs, commercial terms, and compensation clauses.

This is fine, as the U.S. legal market is indeed large. However, in China, contract review follows a different logic.

Chinese contracts have unique expression habits: “Party A shall,” “Party B must,” “Both parties agree,” and “This contract shall take effect from the date of signature.” The logical relationships between clauses, the expression of breach responsibilities, and the jurisdiction for dispute resolution all require a deep understanding of the Chinese legal system.

WPS AI has a clear first-mover advantage in this area. It is trained on a vast corpus of Chinese contract data, understands the structural clauses of Chinese contract law, labor law, and corporate law, and can identify local legal risk points such as “standard clauses,” “exemption clauses,” and “excessive penalties.” More importantly, WPS has a complete library of Chinese contract templates—labor contracts, lease contracts, procurement contracts, confidentiality agreements, and equity transfer agreements—covering the main scenarios of daily business operations. Users can open a template and let AI fill in specific clauses, with every change traced and every risk highlighted.

This is something Claude cannot do. It cannot grasp the “flavor” of Chinese contracts, nor can it understand the “rules” of official documents.

Government Documents: A Market Claude Cannot Enter

There is a market in China that Claude cannot reach at all: government documents.

Government agencies and state-owned enterprises produce a large volume of official documents daily—requests, meeting minutes, notices, and work summaries. These documents have strict formatting standards: title levels, font sizes, paragraph spacing, page margins, and even the placement of “no text on this page” have specific requirements.

WPS has over thirty years of accumulation in this field. From the early WPS 1.0 to the current WPS 365, templates and formatting standards for government documents have been internalized into the product’s DNA. WPS AI builds on this foundation with intelligent capabilities, enabling it to:

Automatically detect formatting deviations and prompt “the title should be in bold size three”
Identify sensitive words and warn “this expression may involve compliance risks”
Compare versions and generate “modification explanations” with one click
Maintain audit trails, recording “who changed what at what time”

These capabilities are not available in Claude for Word. It is not a technical issue but a problem of understanding the context—it simply does not know what a “red-headed document” is, what a “request for instructions” entails, or the standard format for “meeting minutes.”

Ecological Integration: A Unified Experience

One highlight of Claude for Word is its “cross-Office collaboration”—Word, Excel, and PowerPoint share context, allowing data to be pulled from Excel into Word and then condensed into PPT.

WPS also has this capability, and it is even lighter. WPS 365 is designed as an integrated solution: documents, spreadsheets, presentations, PDFs, mind maps, and flowcharts all operate under the same account system. Users do not need to “cross-application” because they are all in one application.

When opening WPS AI, users can say, “Create a chart from this spreadsheet data and insert it into the document,” and AI automatically completes the format conversion and content insertion. Saying, “Summarize this document into a 10-page PPT,” allows AI to automatically extract key points and generate slides.

There is no need for account switching, no format compatibility issues, and no confusion over “where is this file stored in the cloud.” WPS’s “family bucket” is the result of integrated design rather than a patchwork solution.

Enterprise-Level Capabilities: Data Sovereignty

Claude for Word emphasizes operating “within an enterprise security framework,” supporting Amazon Bedrock, Google Vertex AI, Microsoft Azure, and other enterprise gateways. This is fine, but for Chinese enterprises, there is an even stricter requirement: data sovereignty.

Key industries such as finance, government, energy, and telecommunications have strict regulatory requirements for data security. No matter how useful an AI office tool is, if data needs to be sent to overseas servers, it will not pass compliance checks.

WPS AI supports private deployment, allowing enterprises to run AI capabilities on their own servers, keeping data entirely within the internal network. Additionally, WPS has completed domestic compatibility certifications with mainstream Chinese operating systems, databases, and middleware.

This is not just an “added bonus”; it is a “passport”. In the Chinese enterprise market, without these capabilities, AI office tools cannot even enter.

Furthermore, Claude for Word is priced at $20 per month for Pro users (approximately 145 RMB) and $100 per month for Max users (approximately 725 RMB), requiring a subscription to Microsoft 365 to use. In contrast, WPS AI is a value-added service for WPS members, with a lower price threshold, making it more user-friendly for domestic users.

Data-Driven Insights: WPS AI’s User Base

By the end of 2025, WPS AI’s domestic monthly active users exceeded 80.13 million, a year-on-year increase of 307%, with the proportion of enterprise users rising to 42%. During the same period, WPS Office’s overall global monthly active device count reached 678 million.

What does this number mean? It means that WPS AI is no longer just a “concept product” but a productivity tool with a real user base. 80 million monthly active users are using AI daily to write documents, review contracts, and create PPTs, and this usage data continuously optimizes AI capabilities.

Conclusion: The Chinese Approach to AI Office Tools

The launch of Claude for Word signifies that AI office tools have entered a “deep water zone.” It is no longer about flashy demonstrations of “what AI can write,” but rather a practical approach to “how AI can be embedded into workflows.”

In this race, Chinese office software has not been absent. WPS AI, with over thirty years of local accumulation and a deep understanding of Chinese contexts, has carved out a different path:

It does not emphasize “AI disruption” but integrates AI capabilities into daily office tasks.
It does not pursue “omnipotence” but focuses on vertical scenarios such as “contract review,” “government documents,” and “enterprise compliance.”
It does not use high prices to filter users but makes it accessible and effective for more people.

Claude has entered Word, but Chinese office software has long been prepared. This is not a story of catching up but rather the parallel evolution of two distinct paths.

For users, the choice of which path to take depends on where you work, what language you use, what contracts you review, and what documents you write. The competition among AI office tools ultimately hinges not on technology but on understanding the context. In this dimension, local players have the first-mover advantage.

Building a Website with No Code Using AI in Just Half a Day

Sun, 12 Apr 2026 00:00:00 +0000

The entire process of building a website using Vibe Coding (no-code) took just half a day, involving pure dialogue: I provided requirements to Claude → AI wrote the code → I gave feedback → AI updated the code, resulting in over 20 iterations to complete the MVP. This experiment not only validated the feasibility of building a website with “human + AI” but also explored a new topic in GEO (Generative Engine Optimization): how to ensure AI accurately understands and recommends your identity and content.

Previously, building a website was a hassle. Communicating with freelancers was exhausting, and using templates often resulted in unsatisfactory outcomes.

Now, with AI, I launched a website without writing a single line of code or using a template.

From buying the domain to publishing, it took only half a day!

The entire process was straightforward:

I provided ideas and logic → AI wrote the code → I took screenshots for modifications → AI updated the code.

After several rounds of revisions, I ended up with a presentable MVP website.

Note: To quickly validate the no-code website building process, I used an overseas hosting solution. Websites intended for formal commercial operations in the domestic market need to be registered locally.

Here’s how I turned gtmstar.com from an idea into reality:

1. Clarify Key Points Before Getting Started

Audience: Tech entrepreneurs, product managers, marketers…
Their Needs: Tips, case studies, training, outsourcing services in GTM.
Core Value: Summarizing GTM insights, empowering practitioners, professional services, networking…
Differentiation: Methodology, practical cases, and knowledge accumulation.

After answering these four strategic questions, I wrote the copy for each section and had AI polish it. I also planned the website’s layout, color scheme, and style.

2. Building a Website with No Code

Step 1: Let Claude Write the Code

You can prompt Claude with the following:

Help me generate a single-page HTML file for a website.

Website Name: [Name]/ Positioning: [One sentence]/ Navigation: [Section 1/2/3]

Main Title: [English]/ Subtitle: [Chinese]

Font: DM Serif Display + Noto Serif SC + Syne

Brand Color: [Color Code]/ Background: Warm Beige

Style: Clean and professional, text on the left, dark card on the right, with a Newsletter entry at the bottom.

When the first version of the code came out, I was honestly quite excited; it was a small achievement, even though it was not perfect.

Next, I took screenshots, marked the areas that needed changes, and sent them back for revisions.

After many rounds of back and forth, I saved the HTML file, even though it wasn’t perfect yet.

Step 2: Purchase a Domain

I used a landing page service, but you can also use Alibaba Cloud, Wanwang, or Volcano Engine…

The principle was to aim for a .com domain that is memorable and costs under 150 RMB.

I initially wanted something simple with “gtm” and tried over twenty options. Suddenly, I thought of “North Star Metric” and found that gtmstar was available for registration. Without hesitation, I paid for it.

I also bought hannipeng.com for my personal brand website.

Step 3: Host the File in the Cloud

I asked Claude about the technical routes for going live and chose GitHub + Vercel for ease of modification and testing.

I registered for a GitHub account, created a public repository named gtmstar, and dragged the index.html file into the uploading area, clicking Commit changes to save it to the cloud.

Step 4: One-Click Deployment

I opened vercel.com and authorized login with my GitHub account.

Claude guided me step by step to Add New Project, import it into the gtmstar GitHub repository, and finally Deploy.

After about 30 seconds, I saw the Congratulations! message, indicating success.

Then, I returned to the domain site and filled in the IP address and alias from Vercel into the DNS settings.

This step can be a bit confusing for those without a technical background; I solved it by taking screenshots and asking Claude for guidance on each step.

After a few minutes, the website was live!

Step 5: Basic SEO/GEO Settings

This step is often overlooked but is crucial for future search and AI recommendations.

I directly asked Claude, “How can I make my website visible to AI?”

It helped me generate a standard sitemap.xml file and a robots.txt file that allows all major AI crawlers to access it.

I then uploaded these files to the root directory of GitHub, alongside index.html.

Next, I registered with Google Search Console and submitted my sitemap link under the “Sitemaps” section in the left menu.

This step informs AI tools like ChatGPT, Claude, and Perplexity that they can freely crawl my website content.

3. Advanced Operations and Common Issues

Building the website is just the first step; ongoing maintenance is necessary. Here are a few adjustments I made after launching the site:

How to Continuously Update Website Content: Simply click to edit index.html on the GitHub web interface, modify the text, and save.
Using Third-Party Tools to Save Time: Dynamic Resource Library: I chose Notion as my content base and added a redirect link in index.html (currently just starting to accumulate content, with plans for future updates).

Content Accumulation: I tried Substack for blogging, which includes newsletter subscription management. WeChat public accounts can also be used, but each time requires backend code changes, and I must remember to update the website after posting.

Lead Filtering: Instead of directly placing my WeChat number behind the “Free Appointment Communication” button, I integrated a minimalist Tally form. This allows potential clients to fill out three core business questions, helping me understand client information before communication, which is very efficient.

4. GEO Testing Experiment

Building this website was also a small growth experiment initiated by me.

Traditional SEO focuses on “how to be found in searches,” but in the era of large models, I am more concerned with GEO (Generative Engine Optimization) — “how to be accurately understood and recommended by AI.”

Here’s a list of the tools used in building the site:

Vibe Coding: When 'Feeling Right' Becomes the Strongest Programming Language

Mon, 30 Mar 2026 00:00:00 +0000

Vibe Coding: When ‘Feeling Right’ Becomes the Strongest Programming Language

2026-03-30 15:40

Using natural language to describe needs allows AI to generate runnable code—Vibe Coding is disrupting traditional development processes. This article deeply analyzes this phenomenal programming method, revealing how product managers leverage their ability to decompose requirements and their product sense to gain an advantage, while clarifying three key misconceptions. From tool selection to practical paths, this guide helps you master the essential skills to quickly turn ideas into reality.

Recently, someone in a product group shared a message:

“Today, I used Cursor for three hours to create a feature that I previously got quoted for 8000 yuan and needed two weeks to develop. I didn’t write a single line of code, but I knew what I wanted.”

This message sparked nearly 200 replies in the group.

Some said, “This is the future,” others said, “This is not development at all,” and some warned, “This will ruin you.”

However, I noticed that those who had deeply used AI programming tools were the quietest—they were already using them.

This phenomenon has a rapidly growing name: Vibe Coding.

What is Vibe Coding?

In February 2025, OpenAI co-founder Andrej Karpathy tweeted:

“There’s a new way of coding that I call vibe coding. You are completely immersed in the vibe, forgetting the actual existence of code, just watching, asking, running, copying and pasting, and most of the time the results work.”

He even mentioned that when encountering errors, he doesn’t read the error messages but simply throws them to AI to resolve.

This statement quickly sparked discussions in the tech and product circles.

Vibe Coding literally means ‘atmosphere programming.’

However, the term “atmosphere” can be misleading, suggesting a casual, unrigorous approach. Its true meaning is:

Describe what you want in natural language and let AI generate the code; you are responsible for judging whether it “feels right” rather than checking whether the “logic is correct.”

You don’t need to understand variables, functions, or frameworks.

You only need to be clear about—what you want and how it should feel to use.

Why Now?

The term Vibe Coding didn’t appear out of nowhere; it suddenly became a phenomenon in 2025 due to three key reasons.

Reason 1: Tools Have Finally Caught Up with Imagination

As early as 2023, some attempted to use ChatGPT to generate code. However, the experience was frustrating: generate a piece, it wouldn’t work, ask again, still wouldn’t work, and ultimately give up.

Today is different.

Tools like Cursor, GitHub Copilot, v0.dev, and Bolt.new are no longer just “code completion”; they can understand the entire project context, automatically fix errors, and generate complete pages and logic based on natural language descriptions.

The leap in tool capabilities has made “feeling programming” genuinely feasible for the first time.

Reason 2: An Expanding ‘Sweet Spot’

Vibe Coding is not suitable for all scenarios, but there is a large sweet spot—

Those features that are complex enough for individuals or small teams but relatively standardized for AI:

A data dashboard with login
A small tool for form collection and email notifications
An internal task management system
A product prototype demonstration page

In the past, these required hiring developers, scheduling, and spending money. Now, a product manager who knows how to express their needs can potentially complete it in an afternoon.

Reason 3: A Historic Drop in Execution Barriers

A saying has circulated in the Silicon Valley startup circle:

“In the past, ideas were worthless; execution was valuable. Now, the barriers to execution are disappearing, and ideas are becoming valuable again.”

The popularity of Vibe Coding fundamentally represents a significant reduction in execution barriers.

This allows those with clear ideas, user insights, and product sense to truly have the possibility of hands-on involvement for the first time.

What is the Real Experience of Vibe Coding Like?

Let’s recreate a specific scenario.

Scenario: A product manager wants to create a ‘user feedback collection tool’

The requirements are: users fill out a form, and after submission, an email is automatically sent to the PM. The backend can view all feedback and mark the processing status.

Traditional Path:

Write PRD → Find developers → Schedule → Development → Integration → Testing → Go live

If all goes smoothly, this takes a week, but likely two to three weeks.

Vibe Coding Path:

Open Bolt.new and input:

“Help me create a user feedback collection tool. The user side is a form with three fields: name, email, and feedback content. After submission, it automatically sends an email to my inbox. The backend page can view all feedback records, and each feedback can toggle between ‘pending’ and ‘processed’ status. The overall style should be clean and modern.”

Then you start conversing with AI—it generates code, you run it to see the effect, and if something is wrong, you say:

“Change the button color to blue.”

“There should be a success message after form submission.”

“Can the backend list be sorted in reverse chronological order?”

Throughout the process, your input is about the feeling, and your judgment is whether it’s ‘right or wrong.’

You don’t need to know whether it’s using React or Vue, which service to send emails, or how to design the database tables.

This is Vibe Coding.

Why Are Product Managers Naturally Suited for Vibe Coding?

An interesting observation is that among all those trying Vibe Coding, product managers often have a higher success rate than many novice engineers.

This sounds counterintuitive, but the logic behind it is clear.

The core ability of Vibe Coding is not writing code, but expressing needs.

What do product managers do every day?

They break down vague business goals into clear user stories; they describe complex interactions into specific operational steps; they translate “feels wrong” into executable modification suggestions.

This is precisely the core capability needed for efficient collaboration with AI.

In contrast, engineers often fall into a strange loop when Vibe Coding: they know how it should be done technically, but when AI generates something different from their expectations, they get bogged down in implementation details, wanting to control the underlying logic, which actually reduces efficiency.

Product managers are naturally results-oriented and experience-focused, which is the mindset required for Vibe Coding.

Another point many overlook: product sense is the best prompt engineering.

Have you heard of Prompt Engineering? Many people spend a lot of time learning “how to write good prompts.”

But actually, someone with product sense naturally writes good prompts:

They know to clarify who the user is.
They know to describe specific usage scenarios.
They know to specify what success looks like.
They know to differentiate between “must-haves” and “nice-to-haves.”

Isn’t this just the basic skill of writing a PRD?

Three Misconceptions That Must Be Clarified

The rise of any new phenomenon is accompanied by misunderstandings. Regarding Vibe Coding, there are three misconceptions that need to be addressed.

Misconception 1: Vibe Coding = No Technical Knowledge Required

This is the biggest misunderstanding and the most dangerous perception.

Vibe Coding lowers the barrier to writing code, but it does not lower the necessity to understand technology.

When the code generated by AI doesn’t work, you need to determine whether the issue lies in the requirement description or if AI made a mistake; when a function is implemented but performs poorly, you need to know whether that’s acceptable or a problem that must be resolved; when you want to launch something, you need to understand the basic deployment process.

Completely zero technical background individuals will find their upper limit in Vibe Coding very low.

Those with some technical foundation, even just a little, will see an exponential difference in efficiency.

Misconception 2: Vibe Coding Outputs Cannot Go to Production

This statement is overly absolute.

A more accurate description is: the outputs of Vibe Coding can be directly used in certain scenarios, while in high-demand scenarios, additional engineering work is needed.

Internal tools, MVP validation, low-concurrency applications, personal projects… In many scenarios, the outputs of Vibe Coding are entirely sufficient.

If you regard it as a “toy that cannot go to production,” you will completely miss its true value.

Misconception 3: Vibe Coding is for Developers, Irrelevant to Product Managers

On the contrary.

Vibe Coding should fundamentally change the way product managers work.

When you can quickly create an interactive, real prototype yourself, your understanding of requirements will deepen; when you have run through a process yourself, your communication with developers will become more efficient; when you can independently create internal tools, your personal value will stand out.

What is Vibe Coding Redefining?

I want to discuss several deeper changes. These changes are not just about efficiency but structural.

Redefining ‘Knowing How to Develop’

In the past, “knowing how to develop” was a relatively clear skill boundary: you could write code and independently implement functions.

Now, this boundary is starting to loosen.

“I can’t write code, but I can create a product”—this statement holds true in the Vibe Coding era.

In the future, “knowing how to develop” may split into two capabilities:

Can write code: traditional engineering skills, deep, precise, and controllable.
Can use AI to build: new product building capability, fast, flexible, and results-oriented.

Both abilities are valuable, but the latter’s entry barrier is significantly decreasing.

Redefining ‘Product Prototype’

In the past, prototypes were divided into two types:

Low-fidelity prototypes: wireframes created with Axure/Figma, aesthetically pleasing but not truly functional.
High-fidelity prototypes: require development resources, high cost, and long cycle.

Vibe Coding creates a third form:

Runnable functional prototypes: look like finished products, can be operated, and the cost is close to low fidelity.

The impact on product validation is revolutionary. Before talking to users, you can present a real, usable item to them.

Redefining ‘Personal Productive Power’

This may be the most profound change.

In the past, a product manager with an idea found it difficult to independently build a product without a technical partner.

Now, “one-person product companies” are becoming a real possibility.

This doesn’t mean every product can be made by one person, but validating an idea, serving a niche market, and running through the first 100 users—this task’s barrier is being significantly lowered by Vibe Coding.

How to Start Your First Vibe Coding Practice?

If you want to get started, here’s a validated entry path.

Step 1: Choose a Tool and Use It Seriously for Two Weeks

Don’t try everything; start with one and use it seriously.

Recommended choices:

Bolt.new: web-based, zero configuration, suitable for complete beginners, great for full-stack application experience.
Cursor: local IDE, has a certain entry barrier but offers a higher ceiling, more suitable for those with some technical background.
v0.dev: produced by Vercel, focuses on UI page generation, suitable for creating front-end display pages.

Step 2: Start with Real Pain Points, Don’t Practice for the Sake of Practicing

Find a real problem you encounter at work:

Is there data that needs to be manually organized every week?
Is there an internal process that could use a small automation tool?
Is there a product idea you’ve always wanted to validate?

Real needs will push you to turn Vibe Coding into a true skill. Practicing with hypothetical scenarios will likely lead to giving up in three days.

Step 3: Learn to ‘Break Down Further’

The most common reason for failure in Vibe Coding is giving AI too large a request at once.

The correct approach is to break down the goal into the smallest possible units, doing one thing at a time:

Don’t say, “Create a complete user feedback system,” but first say, “Create a submission form,” run it successfully, then add “backend viewing page,” then add “email notification function”…

This aligns perfectly with the logic of product iteration—small steps, rapid progress, and each step is verifiable.

Step 4: Establish Your Own ‘Feeling Standards’

Vibe Coding heavily relies on your subjective judgment, so you need to clarify before starting:

Is this function smooth to use?
Will the target users understand this interaction?
Is this speed acceptable?

Don’t just focus on whether the function has been implemented; pay attention to whether the experience is right.

This is what product managers are better at than anyone else.

Conclusion

Some criticize Vibe Coding, saying it cultivates a group of people who have a superficial understanding of technology yet believe they can develop, producing items filled with security vulnerabilities, performance issues, and unmaintainable code.

This criticism has merit, but it points to the misuse of tools rather than Vibe Coding itself.

Word processing software won’t make someone a writer, PowerPoint won’t make someone a designer, and Vibe Coding won’t make someone an engineer.

But all these tools achieve the same thing: they enable more people to express themselves.

Vibe Coding allows more people to turn their product ideas into something clickable, tangible, and presentable to users.

This, in itself, is valuable enough.

The essence of “Vibe” is an atmosphere, a feeling, a state of flow.

True Vibe Coding is not aimlessly chatting with AI but—having a clear enough feeling about the product that you can accurately convey it to AI and know precisely when it gets it right.

This feeling is one of the most valuable abilities of product managers.

Now, it has a brand new application.

The Man Behind Qwen and Seedance

Thu, 12 Mar 2026 00:00:00 +0000

The Man Behind Qwen and Seedance

There is a person named Zhou, who has a background with Alibaba and is closely related to Lin Junyang, and his products have recently gained significant popularity.

He is not Alibaba’s Zhou Hao, but rather Zhou Chang, who has been leading multimodal projects like Seedance at ByteDance.

During his seven years at Alibaba, Zhou Chang was the technical lead for the Tongyi Qianwen large model. Under his guidance, Qwen made its debut in April 2023, quickly rising to the forefront of global open-source large models.

Before leaving Alibaba in June 2024, he released Qwen2, which outperformed the then-leading open-source model Llama 3-70B. Within two hours of its release, it topped the Hugging Face open-source model leaderboard, surpassing many domestic closed-source models.

Today, the Qwen series has over 200,000 derivative models on Hugging Face, with cumulative downloads exceeding 1 billion, making it the most downloaded open-source model series globally. The release of Qwen3.5 in February this year secured the top four spots on the Hugging Face leaderboard, while the Tongyi Qianwen app reached 203 million monthly active users.

After leaving Alibaba in the summer of 2024, Zhou joined ByteDance’s Seed team. Within less than a year, he took over Seedream, Seedance, and the world model team, becoming the highest authority in Seed’s multimodal direction.

The subsequent story is well-known. On February 7, 2026, Seedance 2.0 launched quietly without a press conference or large-scale promotion, yet it ignited a frenzy in the global tech and capital markets within three days.

Chinese A-share short drama concept stocks surged, and Feng Ji, the producer of “Black Myth: Wukong,” described it as “the strongest video generation model on Earth, bar none.”

In the entire Chinese AI industry, it is hard to find another person with a career trajectory like Zhou Chang’s.

Zhou Chang: A Brief Biography

Zhou Chang graduated with a degree in Computer Science and Technology from Fudan University and later obtained a Ph.D. in Computer Software and Theory from Peking University after five years.

His doctoral research focused on deep learning, graph mining, and distributed computing, with over thirty papers published in top conferences. In July 2017, he joined Alibaba through campus recruitment, using the alias “Zhong Huang.”

New employees at large companies typically go through a “beginner’s village” phase. Zhou was initially assigned to the Damo Academy’s Intelligent Computing Laboratory as an algorithm expert, where his work was not directly related to large models.

He developed product image representation algorithms, user representation frameworks, and self-supervised contrastive learning vector recall algorithms, primarily serving Alibaba’s e-commerce recommendation and search scenarios.

In retrospect, these years of accumulation were crucial for his later career trajectory.

On one hand, he completed large-scale engineering implementations in Alibaba’s core e-commerce business, learning how to deploy algorithms from the lab to real-world applications. On the other hand, he built a core team willing to follow him.

Around 2020, Zhou’s work direction began to shift. Alibaba’s Damo Academy launched a project called Multi-Modality to Multi-Modality Multitask Mega-transformer (M6). Zhou was a key participant in this project, co-authoring with Lin Junyang and Zhou Jingren, two names that frequently appear in the Qwen story.

In March 2021, M6 was officially released, featuring 100 billion parameters, making it the largest model in the global multimodal pre-training field at the time. Three months later, Damo Academy pushed M6 to the trillion-parameter level, significantly optimizing training efficiency.

Compared to models of similar scale, M6 reduced energy consumption by over 80% and improved efficiency by nearly 11 times. M6 achieved unified pre-training for multimodal data in Chinese scenarios, constructing a large-scale Chinese multimodal dataset of over 1.9TB of images and 292GB of text, covering various contexts like encyclopedias, web pages, and product descriptions.

This methodology was later directly applied to Alibaba’s e-commerce recommendation and content generation businesses, with derivative works like M6-Rec widely deployed within Alibaba Group.

The paper was published in top conferences like KDD 2021, co-authored by Zhou Chang, Lin Junyang, and Zhou Jingren. Importantly, M6 also served as the technical predecessor to Qwen in multimodal aspects.

In 2023, the global surge in large models triggered by ChatGPT prompted Alibaba to quickly integrate Damo Academy resources to establish the Tongyi Laboratory. Zhou led the development of the Tongyi Qianwen large model based on M6’s technology, serving as the technical lead and reporting directly to Alibaba Cloud CTO Zhou Jingren.

Over the next year, Zhou’s team first open-sourced Qwen-7B in August 2023, followed by the release of Qwen-VL visual language models, Qwen-Audio audio understanding models, CodeQwen code models, and Qwen1.5-MoE mixture of experts models, covering multiple modalities including text, vision, audio, and code.

In June 2024, just before Zhou’s departure, the Tongyi team released Qwen2, achieving significant success and enhancing Alibaba’s reputation in the open-source model community.

As of now, Zhou’s papers have been cited over 30,000 times, with the most cited being the Qwen2 technical report, which has over 8,000 citations. If academic papers were counted like WeChat articles, this one could be understood as having over 100,000 views.

By the time Zhou left, the cumulative downloads of the Tongyi Qianwen open-source model had surpassed 7 million. The success of Tongyi can be attributed to the technical foundation Zhou built, which continued from M6.

The Leader of Seed Multimodal

In July 2024, rumors circulated that Zhou Chang was about to leave to start his own venture. At that time, he was still within Alibaba’s cloud system and had not yet completed the formal process, but multiple independent sources confirmed his decision to leave. He signed a non-compete agreement upon departure.

Then, events unfolded unexpectedly. Just over two months later, in October, it was revealed that Zhou had quietly joined ByteDance—not to start his own company, but to switch to Alibaba’s most direct competitor.

ByteDance offered Zhou a 4-2 position level (some insiders claim his current level is 5-1) and an eight-figure annual package. Converted to Alibaba’s levels, this is roughly equivalent to jumping two levels with a salary increase of several times. Team members accompanying him also received levels of 4-1 and 3-2.

In November 2024, news broke that Alibaba had formally applied for non-compete arbitration, with insiders close to Tongyi confirming that “the situation is true.”

“The resignation to start a business was just a cover to avoid competition,” said a headhunter closely associated with ByteDance in a previous interview. “But this time, it couldn’t be hidden; not only Zhou Chang, but also more than ten team members followed him to ByteDance.”

As of March 2026, the final ruling of this labor arbitration, including compensation decisions and other core information, has not been publicly disclosed by the two AI giants or Zhou Chang himself. The labor arbitration case is subject to a statutory principle of non-public hearings, and neither party has released details of the case. What is confirmed is that after Alibaba initiated arbitration, the case has completed the legal process.

After joining ByteDance, Zhou was placed in the Seed team’s “Multimodal Interaction and World Model” department. Seed is ByteDance’s large model and fundamental research team, one of the company’s most valued AI businesses, aligning well with Zhou’s expertise.

In February 2025, the arrival of an important figure reshaped the Seed landscape. Former Google DeepMind research vice president and Google Fellow Wu Yonghui joined ByteDance as the head of Seed’s fundamental research, reporting directly to CEO Liang Rubo.

Wu, an alumnus of Nanjing University, had worked at Google for 17 years, leading the development of Google’s neural machine translation system GNMT and participating in the Gemini large model project.

Wu’s arrival restructured the reporting architecture of the Seed team, with several algorithm and technical leads reassigned to report to Wu, including Zhou Chang.

From later developments, it is evident that Wu valued Zhou Chang highly. In July 2025, ByteDance’s visual multimodal generation head Yang Jianzhao announced he would take a “temporary break.” Yang studied under Huang Xutao, known as the “father of computer vision,” and was responsible for text-to-image and text-to-video AI directions at ByteDance. Recent news indicates he is starting a venture in the video model field.

After his leave, Zhou officially took over this business. Shortly thereafter, visual foundational model research head Feng Ji also left the company.

With these two personnel changes, Zhou’s jurisdiction expanded from the original multimodal interaction and world model to encompass all visual AI products including text-to-image Seedream and text-to-video Seedance. He thus became the main leader in the multimodal direction of the Seed team.

Launching Seedance 2.0: A Global Sensation

After taking over the entire visual line, Zhou’s team brought the high-density output model from the Qwen era to Seed.

In the text-to-image direction, the team rapidly iterated Seedream from version 3.0 to 4.0 and then to 5.0. Seedream 3.0 achieved native 2K output and a three-second generation speed, while Seedream 4.0 pushed the resolution to 4K and unified the generation and editing architecture. The Seedream 5.0 released in February 2026 further introduced physical perception and semantic reasoning capabilities.

The achievements in the text-to-video direction were even more significant.

As mentioned at the beginning of the article, Seedance 2.0 launched quietly on February 7, 2026, yet its impact exceeded expectations.

Seedance 2.0 supports native 2K resolution, multi-camera narratives, four-modal input (text + image + video + audio), and multilingual lip-syncing, among other professional features.

Based on actual calculations, the production cost of a 5-second special effects shot can be reduced from 3,000 yuan (one month of labor) to 3 yuan (AI in two minutes).

The short drama industry has thus been transformed into the “AI short drama industry,” with DataEye estimating that, driven by AI technology’s cost reduction and efficiency improvements, the domestic short drama user base will grow from about 120 million in 2025 to 280 million in 2026.

Notably, recent news indicates that Yu Bowen, the former training head of the Qwen large model, has officially joined ByteDance as the post-training head of the visual model and multimodal interaction team in Seed.

This former Qwen team member is once again collaborating with Zhou Chang at Seed.

In fact, the Seed team has seen a steady influx of core team members from Alibaba’s related businesses in recent years, a trend that can be traced back several years before Zhou’s arrival.

For example, Huang Weilin, who was responsible for projects like Pailitao at Alibaba, left in 2020 to join ByteDance’s visual and multimodal research system; former Alibaba voice AI head Lu Lu joined ByteDance around 2022 to lead voice and multimodal large model research; and Ye Qinghao, who worked on document understanding and multimodal research at Damo Academy, left Alibaba around 2022 and is currently listed as a member of ByteDance’s Seed team.

Pony Alpha: A Mysterious Model Making Waves in AI Development

Mon, 09 Feb 2026 00:00:00 +0000

On February 9, reports emerged about a mysterious model called Pony Alpha that has recently gained popularity on the model aggregation platform OpenRouter. Without a launch event, academic paper, or even a disclosed manufacturer, it has quickly attracted attention among developers and model enthusiasts due to a series of unexpectedly impressive real-world performance metrics.

According to OpenRouter, this model is the next-generation foundational model from an undisclosed manufacturer, demonstrating strong capabilities in programming, reasoning, and role-playing, with optimizations for agent workflows and high accuracy in tool invocation.

User feedback from those who have tested Pony Alpha has been overwhelmingly positive. One blogger tested it with a secret SVG generation task, resulting in an impressively high-quality output that made him question whether the answers had been leaked.

Another developer shared that after letting Pony Alpha code for three hours, it successfully created a playable version of Pokemon Ruby, achieving a level of detail that was even more faithful to the original in certain aspects.

Due to its extraordinary performance, the “mystery” surrounding Pony Alpha has become a hot topic of discussion. Some speculate it could be Anthropic’s Sonnet 5, given its familiar coding abilities; others think it might be the long-rumored DeepSeek-V4; while many believe it could be an early test of the next-generation model GLM-5 from Zhipu.

So, what are Pony Alpha’s true capabilities? Are these rumors backed by technical evidence? Let’s set aside speculation and evaluate its performance through a series of tests to see how far this “Pony” can run.

01. Initial Experience with Pony Alpha: From Data Dashboards to Algorithm Visualization

Pony Alpha is currently available for free on OpenRouter, allowing users to interact with the model directly via the web or through API calls, with a context window of 200K.

As Pony Alpha is primarily focused on programming, we centered our tests in this domain.

The first case was a “mini data dashboard”. The prompt required inputting a set of numbers to generate real-time maximum, average, minimum values, and volatility, accompanied by smooth animated updates.

This prompt primarily assesses three abilities: accurate understanding of statistical metrics, frontend structure organization, and the finesse of animation and state updates.

Pony Alpha’s webpage for the “mini data dashboard” showed no discrepancies in metric calculations, with animations employing transitions rather than abrupt refreshes, achieving a high overall completion level.

The second case involved SVG cartoon scene generation. The prompt was very specific, detailing size, theme, elements, style, and requirements, with the core challenge being the model’s ability to maintain consistency under complex constraints.

The model’s final SVG output was structurally clear, with logical layer relationships. Elements like sunlight halos, wave curves, and coconut tree shadows were accurately implemented, with saturated colors that were not overexposed, avoiding simple graphic stacking.

The third case was algorithm visualization. We asked the model to convert sorting or pathfinding algorithms into animations, essentially mapping steps to temporal and spatial changes, testing both programming and reasoning abilities.

Pony Alpha excelled here: color changes corresponded to states, rhythm reflected algorithm progress, and path evolution intuitively presented the decision-making process, indicating it could not only write code but also explain complex concepts through code.

After completing these three cases, it was evident that Pony Alpha has surpassed the current mainstream models in terms of being “capable, visually appealing, and easy to understand.” Next, we aimed to place it in more complex scenarios requiring prolonged reasoning to see if it could maintain its creativity.

02. Architect Thinking in Action: Recreating Stardew Valley from Scratch

The previous cases primarily validated the model’s ability to “write code,” essentially executing low-complexity tasks. The true differentiator is whether the model possesses Agentic Coding ability—the capacity to understand problems from a systems perspective and autonomously advance complex projects over time.

This means the model must decompose system-level requirements like a seasoned architect, maintaining context coherence and goal alignment throughout prolonged operations. We decided to stress-test Pony Alpha by recreating the well-known game Stardew Valley.

Here is the prompt we sent to Pony Alpha. For professional human developers, recreating a game like Stardew Valley typically involves thousands of lines of code, managing game loops, scene management, player and NPC behavior logic, crop growth, plot management, UI, inventory, and save systems, among various mechanisms and entities.

Additionally, it must ensure consistent module interfaces, logical synchronization, smooth animation rendering, correct event interaction responses, and consider performance optimization for the code to be practically usable, extensible, and debuggable.

How would Pony Alpha tackle this challenge? Upon receiving the prompt, Pony Alpha first acted like a project manager, analyzing the core requirements from our complex prompt and outlining the eight major systems and color schemes to guide subsequent development.

Next, Pony Alpha transformed into a system architect, planning the overall project structure. Upon opening the source files, we observed that the project adopted a basic yet universal frontend resource structure, with a clear modular approach in the JS project structure: separating models, rendering, and systems, making it suitable for small to medium-sized projects.

Guided by this philosophy, Pony Alpha created a preliminary playable game interface with a unified visual style, full of healing aesthetics, and a clear core gameplay logic. Actions like tilling (land), sowing (seeds), and watering (watering can) functioned properly, and the stamina consumption system was also reasonably designed.

Of course, this was still a purely frontend demo. To make it more engaging, we further challenged Pony Alpha: to add a data saving mechanism and enhance the game’s visuals.

After understanding our requirements, Pony Alpha provided multiple technical solutions to choose from.

After optimizing the project, Pony Alpha developed a backend server and database, completing a frontend save manager, and coded continuously for over 10 minutes without any human intervention.

After the upgrade, Pony Alpha significantly optimized the original design, moving the inventory and item bar to the bottom of the page, allowing the virtual world to take visual precedence. The lakes, grasslands, and trees in the visuals became more detailed. A weather system was also introduced, dynamically presenting sunny, cloudy, rainy, and even snowy conditions, making the entire world more vibrant and realistic.

03. Deep Dive into Legacy Code: Real-World Code Refactoring

In a real enterprise environment, developing new features is only part of the engineering process; more often, programmers face existing, complex, and historically entrenched “legacy” codebases. These systems often contain implicit rules, technical debt, and historical behaviors, making understanding existing code, pinpointing issues, and safely modifying it more challenging than starting from scratch.

Thus, the value of AI in enterprises lies not only in generating new code but also in effectively understanding, debugging, refactoring, and incrementally developing existing projects. Next, we will evaluate Pony Alpha’s performance in such engineering tasks through practical tests.

We first used Pony Alpha, along with manual input, to create a seemingly outdated financial system. At first glance, this system only appeared to have an outdated UI, but delving into the code revealed larger issues (of course, these were tasks we requested Pony Alpha to perform, not a reflection of its inherent capabilities).

We found that variable naming was chaotic, function responsibilities were unclear, some special mysterious accounts were subtly hidden in if branches, and there were random batch operations and implicit dependencies on historical data.

After clearing the context, we asked Pony Alpha to eliminate the issues it had just created.

For human programmers, such legacy systems can be a nightmare; without a reliable AI’s assistance, you might never know if refactoring will inadvertently delete a critical legacy logic.

AI models can easily stumble in such scenarios; they may attempt to unify rules and eliminate duplicate logic but overlook that some technical realities represent business compromises or true states, and arbitrary modifications could lead to larger bugs.

We sent Pony Alpha the following prompt, essentially asking it to refactor and modernize the code while ensuring the system could seamlessly replace the original modules.

Pony Alpha did not rush to modify; instead, it first conducted an analysis. It could understand that this was a financial system and accurately assess the technology stack in use.

To clarify the issues, Pony Alpha categorized them by severity.

Guided by the refactoring objectives it set, Pony Alpha began the transformation.

Ultimately, Pony Alpha successfully delivered a more modernized version. This refactored financial system not only retained all the original functionalities but also preserved the hidden logic of the “9999” special account, which might have been intended for leadership use, showcasing its technical and emotional intelligence.

Now, let’s take a look at the underlying code. In the original version, global variables and functions were mixed together, whereas Pony Alpha’s modified version showed a clear improvement in architecture clarity, with configuration, data, and business layers distinctly separated, and dependency relationships clearly defined for easier unit testing.

Previously chaotic variable names were standardized, transforming meaningless letters into semantic names, making it easier for colleagues who take over the code later to understand the logic.

Additionally, Pony Alpha proactively added various security and maintainability features that were not explicitly requested in the prompt. For example, input validation can prevent users from missing critical information, while the data loading fault tolerance mechanism can prevent program crashes.

Honestly, watching Pony Alpha meticulously sort and optimize this pile of outdated code while preserving key logic felt like working with a patient and reliable master craftsman, making the work environment much more reassuring.

04. Conclusion: A Next-Generation Flagship Model is Coming

After multiple rounds of testing, Pony Alpha presents an overall user experience akin to an Opus-level next-generation flagship foundational model, rather than just a minor version update.

It demonstrates a clear generational difference in dimensions that truly determine productivity, such as long context handling, complex engineering understanding, and execution stability. This may represent a concentrated release of capabilities honed over a long period by a manufacturer, optimized for real development workflows. As for its true origin, no conclusion has been reached yet.

However, it is certain that if this “Pony” is indeed a long-awaited breakthrough from a domestic manufacturer, then the competition in high-level programming and engineering agents among domestic foundational models may have already entered a new phase.

The Era Beyond Coding: Insights from Cursor CEO Michael Truell

Sun, 11 May 2025 00:00:00 +0000

The Era Beyond Coding

In today’s rapidly advancing field of artificial intelligence, software development is undergoing a profound transformation. Michael Truell, CEO of Cursor, introduced the concept of the “post-coding era” in a recent interview, suggesting that future software development will no longer rely on traditional programming languages but will instead use natural language to describe intentions for automated programming. This idea not only challenges existing development models but also opens up new possibilities for software creation.

Since the second half of last year, AI programming has gained significant traction.

Anysphere is considered one of the most successful companies in this field, with its flagship product, Cursor, achieving impressive milestones: reaching a $100 million ARR in just 20 months and $300 million ARR (approximately 2.1 billion RMB) within two years.

On May 1, Lenny’s Podcast interviewed Michael Truell, co-founder and CEO of Anysphere. In this conversation, Michael shared his vision for the future, lessons learned, and advice for preparing for the rapidly approaching AI future.

Here are the key insights and viewpoints from the interview:

What is the post-coding era?
The importance of taste in the post-coding era
The origin story of Cursor
Why build an IDE?
Everyone needs to become an engineering manager
Rapid iteration as the secret to Cursor’s success
Tips for using Cursor
Recruiting and building a strong team

1. What is the post-coding era?

Our goal in creating Cursor was to develop a new way of building software. You can automatically generate programming by simply describing your intentions to the computer in natural language.

In comparing this “new” approach to several popular views on the future of software, some believe that future software development will remain similar to today, still requiring formal programming languages like TypeScript, Go, C, and Rust. Others think that simply inputting commands for robots to write corresponding code will suffice.

However, both of these perspectives have their flaws. The notion that nothing will change is incorrect because technology will evolve and improve. The problem with chatbots is that they often lack precision; you need to continuously prompt them for modifications instead of broadly saying, “help me modify the application.”

The future will present a more unique perspective than either of these approaches. In this future, people will be able to edit and control details from a higher level, making it easier to understand and modify. It transcends traditional code, resembling pseudocode, where the expression of software logic is more akin to natural language. We are committed to evolving complex symbols and coding structures into forms that are easier for humans to read and edit.

2. The importance of taste in the post-coding era

We believe that ultimately, we will evolve to a stage where the development path requires the participation and promotion of existing professional engineers. It appears to be an evolution from code.

However, it is undeniable that this will be a human-led process. Humans will not relinquish control over all aspects of software.

In the post-coding era, taste will become increasingly valuable. Typically, taste is perceived in terms of visual effects, such as smoothness, color, UI, and other design aspects. However, I believe that defining software also encompasses its logic and operation.

This will define the intent of product design, i.e., how you expect the software to operate. This way of thinking will lead more people to see themselves as logic engineers rather than mere software developers. It elevates thinking to the abstract “what is” rather than lingering on “how to do it.” However, we still have a long way to go to achieve this.

There are many instances online where software developed due to over-reliance on AI has obvious flaws and issues. Despite this, in the future, people may not need to be so cautious and can focus more on taste. This is somewhat similar to Vibe Coding.

However, the creation of Vibe Coding has its issues. We create without understanding. In this state, you can produce a lot of code but fail to grasp the details, leading to numerous problems. If you don’t understand the underlying details, you will quickly find that what you create becomes too large and difficult to modify.

So, how can those who do not understand code control all the details? This is what interests us and is closely related to current professional developers. Additionally, I believe we currently lack the ability to let “taste” truly dominate software construction.

Taste can be understood as having a clear and correct vision of what should be built and turning that vision into reality. This requires a clear understanding of the software’s operational logic, effects, and how to achieve them. Unlike now, where after having an idea, one must translate it into a very tedious and cumbersome format that the computer can execute.

3. The origin story of Cursor

As one of the fastest-growing products in history, Cursor has not only changed how people develop software but also transformed the entire industry. So, how did Cursor, which changed everything, begin?

The inception of Cursor stemmed from our thoughts on how artificial intelligence will develop over the next decade. There were two decisive moments: the success of the Code Pilot beta, which introduced us to genuinely useful AI products, and the series of model scaling papers released by teams like OpenAI, confirming that simple scaling could enhance AI performance.

At the end of 2021 and the beginning of 2022, we were very optimistic about the development of AI. At that time, we felt that many people were discussing model creation, but no one was delving into a knowledge work field to explore how it would change after becoming AI-driven.

This led us on a path of exploration. We wanted to know how these knowledge work fields would change as this technology matured and how models needed to be improved to support these changes in work. Once the scale and initial training were exhausted, how would you continue to drive the development of technological capabilities?

To this end, we decided to develop Cursor. Of course, in the early stages, we made a mistake. We chose to study a relatively uncompetitive and dull knowledge area—automating mechanical engineering and product creation.

But neither my co-founder nor I were mechanical engineers, and we were very unfamiliar with this field. It was akin to blind men touching an elephant. For us, starting from zero meant a lot of tricky work.

For instance, developing models requires data, but there was very little 3D model data on parts and tools at the time, and sourcing it was problematic. Eventually, we realized that mechanical engineering was not our passion and not worth the effort.

Looking around, we found that the programming field had not changed much over the years and had not kept pace with future trends. There seemed to be insufficient ambition and urgency regarding the future direction of software development and how AI would reshape everything.

This led us to create Cursor. The lesson we learned is that even if a field seems overcrowded, if you find that existing solutions lack ambition or are significantly insufficient compared to your vision, there are still huge opportunities hidden within.

To seize opportunities, you first need to identify areas where significant leaps can be made. You need to find places where you can make a big impact. AI has provided us with a vast space to operate. I believe the ceiling in this field is very high. Currently, even the best tools have a massive amount of work to be done in the coming years, with significant room for improvement.

4. Why build an IDE?

When deciding to pursue programming, there were several paths we could take. One option was to create an IDE (Integrated Development Environment) for engineers and then incorporate AI into it; another was to build a complete AI agent development product; and the third was to create a model that excels at coding and focus on developing the best coding model.

Cursor’s focus on building an IDE stems from the desire for decision-making authority. We care about allowing humans to control all decisions in the final tools they are building.

In contrast, those who initially focused only on models or end-to-end automated programming were attempting to create an AI-dominated future. Our philosophy regarding AI decision-making is fundamentally different.

We have always approached current technology with a realistic mindset. However, I initially built the product using the software we developed (dogfooding), and we were the end users. This undoubtedly led us to believe that we needed humans to maintain control, as AI cannot handle everything.

Furthermore, the scalability of existing coding environments is very limited. To adapt to changes in programming forms, one must have control over the entire application. We believe that IDEs will develop more broadly than existing coding environments.

We can control them and build an entirely new environment. Of course, the form of IDEs will also change and evolve over time. For now, we primarily view IDEs as places to build software.

Cursor can allow AI to run independently, or humans and AI can collaborate before letting it work independently.

5. Everyone needs to become an engineering manager

When using AI Agents, many unsatisfactory results can still arise. It’s like humans are the engineering managers, and the Agents are the less intelligent subordinates.

As managers, we need to spend a lot of time reviewing, approving, and standardizing.

Thus, we observed that the most successful customers using AI remain very cautious. They heavily rely on “next-step programming predictions” to ensure that AI can predict the outcome of the next action they desire.

Overall, there are two ways to operate. One is to spend a lot of time editing operational instructions and then throw them all at AI, followed by reviewing their work. The other is to break down instructions. First, specify some tasks for the AI to work on, then review; specify more, let the AI work, and review again. This back-and-forth continues until a reasonable scope is achieved.

Successful customers often adopt the second approach.

6. Rapid iteration as the secret to Cursor’s success

When we began building Cursor, we were quite obsessive about it being something entirely new. Now, we develop software based on VS Code, similar to how many browsers use Chromium as a base.

Initially, we did not take this approach and built the Cursor prototype from scratch, which required a lot of work. We rapidly built various components at an incredible speed, starting from scratch with our own editor and then constructing the AI components.

About five weeks later, we began using our editor entirely. When we found it to be basically useful, we immediately let others use it and had a short testing period. Approximately three months later, we released Cursor. Our strategy was to release as quickly as possible and modify versions based on feedback. The initial user feedback was extremely valuable, prompting us to abandon the zero-based version and shift to developing based on VS Code.

Since then, we have iterated our product based on user feedback.

7. Tips for using Cursor

The success of using Cursor largely depends on understanding the capabilities of the model, including the complexity of tasks it can handle, the quality, the gaps, and what it can and cannot do. Currently, we have not effectively educated people on this aspect within the product.

To cultivate this intuition, I have two suggestions. First, as previously mentioned, do not lean towards telling the model all your instructions at once and then waiting for results. Instead, I would suggest breaking things down into different parts. You can spend roughly the same amount of time specifying the overall tasks but do so in a more granular way.

This way, you only need to specify a little bit to accomplish a small task, gradually leading to a complete outcome.

At the same time, I encourage current professional developers to discover the limits of what these models can do through experimentation. Many times, we do not give AI a fair chance and underestimate its capabilities. Tools like Cursor can provide immense benefits to both junior and senior engineers.

We have observed that junior engineers tend to rely too heavily on AI, while senior engineers often underestimate AI’s assistance and stick to existing workflows. For senior engineers, the promotion and adoption of such tools are driven by the internal developer experience (DevEx) teams within companies.

8. Recruiting and building a strong team

For us, having a team of world-class engineers and researchers developing Cursor alongside us is crucial. This is important for both personal and strategic reasons for the company.

Our goal is to find individuals with curiosity and a spirit of experimentation, as we need to build many new things. At the same time, it is important to remain clear-headed.

In addition to creating products, recruiting the right candidates is also a focus for us. We concentrate on finding what we consider world-class talent, sometimes spending years to recruit them.

However, I believe we were not very skilled at this approach initially. We have learned valuable lessons in the following areas:

Who is the right candidate?
Who adds real value to the team?
What does excellence look like?
How to attract those who are not actively looking for jobs?

In the early stages, we leaned too heavily towards seeking candidates who fit the prototype of prestigious schools, excelling in their academic performance. We placed too much emphasis on credentials, interests, and experience.

While this provided us with many excellent talents, they sometimes appeared different from our initial ideal candidates.

Another lesson was regarding the interview process. A core part of our interview strategy is to invite candidates to the company to work with us on a two-day project. This serves both as a test and an interaction.

The advantage is that it allows candidates to complete a real end-to-end project, showing actual output within two days without consuming a lot of the team’s time. It helps you assess whether you would want to work with this person, as you will be collaborating for two days.

Attracting candidates is also crucial, especially in the early stages of the company when the product is not yet mature.

Home on ZovixBC: Cutting-Edge Tech News & Insights

Sun Yat-sen University: New Paradigm for AI Talent Development

Introduction

Enabling Personalized Education

Building an Integrated AI Curriculum System

Training Talent for Future Industry Needs

Beginner's Guide to AI Programming (Vibe Coding) in One Week

Introduction to Vibe Coding

What is Vibe Coding?

Benefits of Vibe Coding in Daily Life and Work

Vibe Coding is Simple: Just a Rough Idea and the Right AI

AI Can Generate Code in Bulk and Quickly

AiPy: A Powerful Tool for Vibe Coding

My Personal Experience and Suggestions

The Geopolitical Divide in AI: China's Open Approach vs. U.S. Restrictions

The Geopolitical Divide in AI

The Escalating Wave of Bans: Who is Closing the Door?

China’s AI Goes Global: Why the Open Approach?

Behind the Bans: Policy Pressure as a Sword of Damocles

Deeper Reasons: Security Fears, Hegemonic Anxiety, and Commercial Interests

Why Does Chinese AI Welcome All?

AI Knows No Borders, But Competition Has Stances

Understanding Vibecoding: Generating Cross-Platform Apps from a Single Description

Introduction

Why “One Description, Three Platforms” is Essential

The Five-Layer Functional Architecture of Vibecoding Tools

1. Semantic Layer: From Natural Language to Unified Business Intent Model

2. Structural Layer: Flow Canvas and Multi-Page Routing Planning

3. Prototype Layer: Independent Rendering of UI Semantics Across Platforms

4. Code Layer: Independent Engineering Output for Each Platform

5. Preview and Validation Layer: Multi-Platform Simulators and Interactive Testing

Comparison of Six Vibecoding Tools on Cross-Platform Generation

In-Depth Analysis of Six Tools

1. UXbot

2. Bolt

3. v0

4. Lovable

5. Replit Agent

6. Tempo

Eight-Point Checklist for Evaluating Vibecoding Tools’ Cross-Platform Generation Capability

Practical Usage Paths for Three Types of Teams

1. Independent Entrepreneurs: Simultaneous Launch Across Platforms

2. Product Managers: From PRD to Demo in One Meeting

3. Small to Medium Development Teams: Using Native Projects as Starting Points Rather Than End Goals

Frequently Asked Questions (FAQ)

Q1: Can the mobile code generated by Vibecoding tools pass App Store and Google Play reviews?

Q2: If I need to modify a page after generating code for all three platforms, where should I make the change?

Q3: Can product managers without Swift/Kotlin experience directly publish the generated mobile apps?

Q4: How does Vibecoding’s cross-platform generation differ from frameworks like React Native or Flutter?

Q5: Will generating all three platforms at once lead to poor code quality and future maintenance difficulties?

Conclusion

Carbon Neutrality Weekly Report: AI and Energy Synergy, Local Policies, and Industry Practices

Carbon Neutrality Policies

1. Four Departments Issue Document to Promote AI and Energy Synergy

2. New Evaluation Method for Beautiful China Construction Released

3. Charging Volume on Highways Increased by 55.6% on May Day

4. National Energy Administration Promotes Green Energy Initiatives

5. Meteorological Bureau Focuses on Extreme Weather in New Energy Development

6. National Carbon Market Price Update

Local Developments

1. Guangdong’s 14th Five-Year Plan Released

Corporate Practices

1. A-Share Airlines Report Over 90 Million Tons of Carbon Emissions

2. Cement Industry Accounts for 90% of A-Share Construction Material Emissions

Fuzhou Accelerates AI Integration with Industry

Fuzhou Accelerates AI Integration with Industry

Enhancing Smart Manufacturing

Avoiding Homogeneous Competition

Driving Integration of Land and Sea

From Satisfaction with Responses to Satisfaction with Results

ChatGPT's Functionality in May 2026: A User Experience Review

ChatGPT’s Functionality in May 2026: A User Experience Review

1. Basic Text Capabilities: Mature and Stable, Covering Everyday Scenarios

2. Multimodal Functionality: Comprehensive and Practical, with Room for Detail Optimization

3. Tool Integration and Memory Capability: Convenient and Efficient, with Personalization to be Deepened

4. Function Layering and Permission Differences: Clear Gradients, Noticeable Limitations for Free Users

Conclusion: Adapting to Mass Needs with Room for Advancement

DeepSeek V4 Flash: A Local Inference Engine for Mac by Antirez

A Local Inference Engine for V4 Flash

How Is This Achieved?