One of my key strategies for tackling increasingly complex software system development over the years has been test-driven development. Although it requires significant time and effort invested in testing upfront, the continuous returns in the long run are well worth it.

I've noticed many developers misunderstand Test-Driven Development (TDD), thinking they need to create a perfect test bench and test suite right from the start. This is completely wrong. Test suites and test benches can evolve dynamically - they don't need to be perfect from the

I found that many engineers online have a strong aversion to writing automated test cases. The main reasons often stem from using the wrong testing tools or having serious misunderstandings about testing methods, which leads to very high costs for introducing automated testing

Some programmers argue with me that software needs to be specifically designed to be "testable," claiming this increases development costs - which is completely wrong. I've always believed in testing software the same way humans would: if humans can use it and test it, machines

Many users are interested in the AI tool I wrote at Taobao over a decade ago - a system that could automatically build relatively complete data analysis products. Let me share more about it. At that time, this AI system, consisting of about 4,000 lines of code, outputs several

Many developers have asked about test-driven toolchains and implementation details, so I'll elaborate a bit more. I personally favor Perl's testing toolchain, especially the mini-language in Test::Base, which is extensible, concise, and convenient for testing any software system

For many years, my goal has been to quickly understand how complex software works. It's as exciting as exploring uncharted territories. I plan to share various open-source software's inner workings in a way that beginners can understand, which should be very interesting topics.

Many programmers are often afraid to modify code they've already written; they try to avoid changes whenever possible. However, I particularly enjoy refactoring or polishing old code. Programs that aren't modified much are essentially "dead programs." The refactoring process is

Large language models still have noticeable stability issues with prompt inputs. Sometimes even seemingly harmless or irrelevant elements can have unexpected significant impacts on these models. Whether it's the smartest SOTA model or a smaller model, such issues will arise. It

A good AI Agent program framework should meet these requirements: 1. Lightweight and efficient, 2. Easy to automatically maximize request concurrency, 3. Transparent switching between different LLMs (vendor APIs or locally deployed), 4. Transparent mixing of models of different

Looking at the current state of large language models in the market, I can summarize it in one sentence: "Sufficient for basic literacy, but lacking in professional expertise." These large models have inherited human flaws - they tend to be careless, prone to speculation,

Even the best SOTA large language models show clear signs of being influenced by mediocre code when writing slightly complex programs. Perhaps during pre-training, these models should only be exposed to truly high-quality code. #AI #LLM #CodeQuality #MachineLearning #Programming

I used to think that through error feedback, large language models could continuously correct their outputs. However, for truly complex problems, even the best current models aren't necessarily convergent under feedback prompts. They might get stuck in infinite loops, go down

I've noticed that the new form of programming in the AI era is writing "prompts." However, traditional programming is still very necessary right now. Both need to be organically combined to achieve good results. After all, we can't rely solely on manual back-and-forth

I typically set the temperature parameter of large language models to 0, except for occasions when I want the model to enter a more creative mode or generate more varied responses. #LLM #AIParameters #MachineLearning #Temperature #PromptEngineering

It appears there are noticeable differences in TPU computing power across Google's various data centers. We regularly use large language models in data centers across America, Europe, and Asia, and have found that for Gemini batch processing tasks, us-east1 is the slowest -

Our AI system discovered that SQLite 3 open-source database software uses Tcl scripts during the build process to dynamically generate final massive C files. For example, it generates corresponding definitions and interpreter implementations for different DB operation codes. In

We've just released OpenResty XRay 25.4.15, our real-time application analysis product! Key highlights: • Added NodeJS app detection with C++ level analysis (JavaScript tracing coming soon) • Enhanced target filtering by IP/hostname • Optimized data compression for faster

The open-source tool Pandoc is still not very reliable for generating PDF files from Markdown. When dealing with large Markdown files generated by AI or programs, it simply throws errors like "dimensions too large." The official team doesn't seem interested in fixing such issues,

We've recently released the new OpenResty open source version 1.27.1.2, which includes numerous recent bug fixes and new features. The complete changelog is available here: openresty.org/en/ann-1027001… We recommend all open source OpenResty users upgrade to the latest version.