[🧑‍💻] The Current Level of AI Code Autocomplete (feat. VSCode Copilot)

✨ Gemini 2.5 Pro’s Summary 　

Seeing once is better than hearing a hundred times. First, let us look at current VSCode Copilot AI code suggestions through a very shallow example.

• Original: http://blog.naver.com/hyeogikarp/223875141105
• Naver publish time: 2025/05/23 11:53 KST
• Original category: Development

Original

Seeing once is better than hearing a hundred times.

First, let us look at current VSCode Copilot AI code suggestions through a very shallow example.

Typing once is better than seeing a hundred times.

If you are interested in development, try it once first. You will experience a whole new world.

Securing a communication route between AI and backend through MCP, passing tasks and context to AI Agents equipped with MCP to achieve work automation… Looking at technologies like this these days, there were moments when I suddenly worried that the development paradigm had already gone far beyond simple “code suggestions.” “If I do not learn MCP, is my developer life ruined?! Tremble tremble…”

https://www.youtube.com/shorts/2gGxaiS9BOM

But if you dig into the truth, realistically, “for now,” even those cool tools often cause trouble or look like they have a screw loose when high-level core logic is needed or the project scale grows. Why? Because AI performance is “still” lacking. Unless AGI has appeared, there are clear limits to context understanding. (Of course, looking at the frightening growth speed of current AI, I do feel like AGI may be less than five years away.)

So at least for me right now, while I am in “development brain-muscle rehabilitation,” the most appropriate and realistically applicable development paradigm is probably the level shown in the photos above: clearly convey the planning intent to AI through comments, carefully read and understand the structure of the generated code, then “allow” it to be applied. Exactly the “knows-what-I-mean and neatly handles it” autocomplete level. Well, that is what I mean for now.

Of course, for a simple MVP-level app or webpage, the era has already arrived where an AI Agent covered in MCPs can produce something in under 10 minutes by skilled-person standards if you say, “do this.” (Leaving aside the completeness of the output… apparently even that can be handled with Context7 MCP, but I do not know well either, so pass.)

Anyway, since that is the situation, it is clear that I should experience and get used to prompt-centered development, so-called “vibe coding,” in advance to prepare for the arrival of the new paradigm.

Looking at Gemini’s overwhelming performance shown at the recent Google I/O, and Google’s confidence toward AGI, I intuitively feel that beyond this “vibe coding,” even the era of natural-language conversation-centered “mouth coding” is not far away. It is time to wake up hard and fasten the seatbelt tightly.

But that is that, and what matters is the “present.” If I get swept up in impatience about the future, I will end up accomplishing nothing properly.

The task given to me right now is to focus on studying Flutter and developing the main app. “Vibe coding” or “mouth coding” is the next stage.

Before discussing grand future technology, let me properly use the tools already in my hands.