The road to becoming an AI creator News: AI code assistants have evolved remarkably! But there are also challenges. What is the most powerful tool? Developers' true feelings and expectations for the future. #AIdevelopment #codeassistant #LLM
Video explanation
What is an AI programming assistant? AI will help you with your coding!
Hello, this is Jon! You may have heard the word "AI" a lot recently, but you may think it sounds difficult. But don't worry! Today, I would like to talk about "AI coding assistants," which are a hot topic among programmers, who say it's "like magic!", so that even beginners to AI can understand it. If you read this, you may be able to get a glimpse of the future of AI and programming.
Simply put, an AI coding assistant is a "smart AI companion that helps you with programming." When we ask it to "make a program like this...", it will suggest code ideas and point out mistakes just like a senior programmer with a lot of experience. It really does feel like the way we work in the future!
In the overseas article on which this article is based, the author actually used several popular AI coding assistants and summarized the points he felt were "amazing!" and "this could be better...". It's as exciting as trying out a new stationery item!
Popular AI coding assistants: their strengths and weaknesses
There are many AI coding assistants out there, each with their own unique features. Here we'll take a look at some of the most popular tools!
ChatGPT: An all-rounder, but maybe a little forgetful?
"ChatGPT" is a very smart AI developed by OpenAI. It is like a knowledgeable all-rounder, and can understand almost any request. For example, if you send it the contents of an open file and ask it to "fix this part like this!", it will return a suggested correction. It is a big step forward that you no longer have to copy and paste every time like in the old days.
However, if you ask for a lot of things at once or if the task is complicated and spans multiple files, it can get a bit confused. There is a limit to the amount of information that AI can process at once.Token LimitIf I go beyond that, I stop midway. It's like a new employee who loses concentration if the conversation goes on for too long.
- Specialties: Proposal of revisions in one file for a wide range of questions
- Dislikes: Understand the overall project context, execute external programs, and sometimes reach information limits
GitHub Copilot: Great at completing input, but limited perspective?
"GitHub Copilot" is a very smart assistant that proactively suggests "How should I write it next?" while you are writing code. It is compatible with development tools such as "Visual Studio" and "Visual Studio Code" that are often used by programmers.IDEThe appeal of this tool is that it runs smoothly within an IDI (Integrated Development Environment, which is like a convenient toolbox for creating programs). It's really convenient that you can write comments and it will quickly produce code that looks like the code you want!
However, while it is good at looking at a single file, it is not good at looking at the entire project or large changes that span multiple files (Cross-file refactoringI guess she's a bit of a narrow-minded, craftsman-like assistant.
- Specialties: Smooth input completion and integration with development tools
- Dislikes: Edit across multiple files, understand context beyond the open file
Cursor: The changes are easy to see, but a bit finicky?
"Cursor" clearly shows the changes made by the AI, line by line (Inline diff reviewThis tool is good at fixing files that have been edited. Sometimes, it can fix dozens of files at once. It's amazing!
However, this tool is based on the popular code editor "VS Code" (forkHowever, because of this, some of the original useful features can no longer be used (due to licensing issues, i.e. complicated rights issues), and there are strict limitations on the number of times it can be used (Rate Limit) and is a bit temperamental. Sometimes it seems to slow down or stop moving, so there seems to be an issue with stability.
- Specialties: Easy-to-understand display of changes, improved stability (still improving), and the ability to try out Claude 4, a high-performance AI, at a relatively low cost
- Dislikes: It is a unique modified version, there are strict usage restrictions, and the reaction time is unpredictable.
Windsurf (formerly Codeium): Fast and generous, but sometimes out of control?
"Windsurf" (formerly Codeium) is characterized by its fast response! The author of the original article describes it as being like a turbocharged Cursor. The wide range of free usage is also a nice point. It uses an AI model from OpenAI (the company that created ChatGPT), and sometimes returns answers 2 to 3 times faster than other tools.
However, in exchange for its speed, it can sometimes be clumsy. When editing multiple files, it may tamper with unrelated files or repeat the same task multiple times. When using it, try it in a different place first (sandboxI think of it as a safe environment like playing in a sandbox, and I take only the good parts.Cherry Pick) seems like a good option. I might hesitate to use it if it's no longer free.
- Specialties: Outstanding response speed, large free tier, and cheap use of OpenAI AI models
- Dislikes: Unpredictable editing, uncertain future (after acquisition by OpenAI)
RooCode: You're great at planning things, but is your working style extreme?
"RooCode" is a planning assistant that divides big jobs into small tasks and proceeds while asking "Is this okay?" one by one. With the "Boomerang Tasks" function, it's like having a work director (Orchestrator) and manages the entire process.
It also shows you the changes, but instead of directly in the code like Cursor or Windsurf, it shows them in a separate little window (Modal Panel) is displayed. And the work speed can only be chosen between "Super Express Mode (leave it to the AI, useful for prototyping)" and "Super Slow Mode (approving each step in detail)," so it is frustrating not to be able to proceed at a comfortable pace. There seems to be a small glitch in that even if you change the settings, they are not reflected unless you restart the software.
- Specialties: Powerful task management features, provided as an additional feature (plugin) for VS Code
- Dislikes: How to display changes, no balanced work speed, occasional settings glitches
Claude Code: For experts? For those who can do without an IDE?
"Claude Code" is one of the tools that people in the know are talking about on Discord (a chat tool often used by gamers and developers). It is a tool that you use by typing commands into a black screen.CLI Tools: A command line interface tool) and seems to be useful for creating simple scripts in a programming language called Python.
However, as the project grew, the number of errors in the program (debugWhen this becomes important, the author of the original article feels that it would be safer to have a dedicated tool like an IDE (Integrated Development Environment). It may be good for people who like the old-fashioned way of doing things, but when it comes to efficiency...
- Specialties: You can use the AI Claude stably
- Dislikes: Debugging on the command line is hard, especially on large projects.
What's missing from your current AI coding assistant? I wish it was more like this!
We've looked at a variety of tools so far, but the author of the original article seems to feel that there are still some issues with the current AI coding assistants. Here are some features he would like to see in order to make them easier to use.
- A plugin, not a fork: Forks (modifying the original software to create a new software) tend to make useful additional features unavailable and delay updates to the original software, so plugins (things that can be easily incorporated as additional features) are better.
- Smart memory and recall of necessary information: Most tools tend to forget old information when there is too much information, but the forgetting is random and can even cause important purposes to be forgotten. I want something smart enough to pin what I need, forget what doesn't matter, and quickly recall related information when I need it.
- If you want more detailed settings: I want users to have more freedom and control, such as being able to choose the type of AI to use and deciding how much information the AI will remember.
- I want to see the changes line by line (inline diff required): It's unsettling when you can't check line by line exactly what the AI has changed and how. I hope this becomes a standard feature!
- I want it to work stably (stability first): It's much more frustrating if the app stops midway or crashes due to an error than if it makes some strange suggestions. Trust is the most important thing!
- Let's all make it better (open source is important again): With open source AI (where the blueprints are publicly available and anyone can contribute to improving it), we can modify and improve it ourselves.
- Please follow the rules (deterministic guardrails): Even if the AI's answers are sometimes capricious, it would be a problem if the basic aspects such as settings, restrictions, and memory rules did not work as predicted.
- Help me find mistakes, not code them! (Improved debugging assistance): It's already amazing that AI can write code for us. But it can take a lot of time to find the errors (debug) afterwards. I'd rather have them help me more!
Well, it certainly would be much more convenient if such a feature existed!
Summary: What does the future hold for AI coding assistants?
The author of the original article says that AI coding assistants are like "talented, but slightly distracted interns" -- they have great potential, but still need human support.
But each tool has its merits, like GitHub Copilot showing the convenience of input completion, Cursor pursuing the clarity of changes, Windsurf showing the importance of response speed, and RooCode showing the possibility of task management. If these advantages can be combined well and become smarter, more stable, and more flexible to use, the day when a truly reliable AI assistant like a "mid-level engineer" is born may be near.
"Vibe Coding," which involves programming together with AI, has already significantly improved productivity. I'm really looking forward to seeing how it evolves in the future!
A word from the author
Wow, AI is really evolving at an incredible speed! I tried out a few AI tools while writing this article, and I was always amazed at how smart they were. Of course, sometimes they make ridiculous suggestions that make you go "Huh?", and there are times when you feel that humans still need to take the reins firmly (laughs). But things that were unthinkable a few years ago are already becoming a reality, so I can't help but get excited about what the future holds!
This article is based on the following original articles and is summarized from the author's perspective:
What the AI coding assistants get right, and where they go
Wrongs