I just used AI to write a paper review of DeepSeek-V4.

The workflow was smooth. I dropped the PDF in, asked AI to extract the content, organize the structure, generate illustration ideas, and then turn everything into a technical blog post that looked fairly complete. The key innovations and the interpretation of evaluation results were both pretty decent. I put the outline below. It looks quite good, doesn’t it?

A Shutdown Notice

On April 20, GitHub published a blog post with a restrained title: “Changes to GitHub Copilot Individual Plans.” The content was much less restrained:

• New signups for Pro, Pro+, and Student plans were paused
• Usage limits were tightened through both session limits and weekly limits
• Claude Opus was removed from the Pro plan

The reason was direct: agentic workflows were consuming far more compute than expected, and “a small number of requests can now cost more than the subscription price itself.”

How I Stopped

Over the past few months, I had been using vibecoding almost obsessively. The term has become common enough by now: using AI to quickly build products, sometimes taking an idea from scratch to launch in just a few days, or even a few hours. I enjoyed that pace. I could have an idea at midnight, start coding the next day, and publish it over the weekend.

During that period I built two apps and a client somewhat like a combination of Lobehub and local Claude Code. The number was not huge, but the last one gave me the strongest sense of frustration.

Introduction

Have you had this experience too: you use an LLM to write code, produce thousands of lines in a day, and then look back only to realize that the truly valuable part may be no more than a few hundred lines?

It sounds absurd, but it is becoming normal. Bryan Cantrill, CTO of Oxide and co-creator of DTrace, recently wrote The Peril of Laziness Lost, aiming directly at the heart of the problem: LLMs are killing one of the programmer’s most important virtues: laziness.

Recently, I developed a new app called Creative Util, this time exclusively for the Mac platform. In both work and daily life, I often need various small utilities—things like converting Markdown to WeChat articles, uploading screenshots to OSS, color picking, document processing, and more. Usually, I’d either use tools like utools or find some online website for a quick fix. But I’ve always wanted a more complete local toolbox that could handle all my needs. After some deliberation, I finally built this app. It includes three categories of tools: Design & Creation, Development Assistance, and Document Processing. Currently, it already contains nearly 20 commonly used tools, and I’ll continue adding more over time.

In the rapidly evolving field of deep learning, innovative neural network architectures are constantly emerging. Keeping pace with these developments necessitates the study of these case studies. This blog is based on the content from the second week of the fourth course in Professor Andrew Ng’s deep learning specialization, focusing on some case studies of convolutional neural networks.

Significance of Case Studies

Firstly, consider why we need to study these cases.