數位小火爐，以及聊聊這陣子密集使用 AI 的感想 Digital Fireplace, and Thoughts After Six Months of Intensive AI Usage

It has already been six months since my last blog update, which is honestly wild. It’s about time to properly share some of the odd, useful, or totally frivolous little things I’ve built with AI lately.

👉 Digital Fireplace (Fireplace)

The idea behind this project was very simple: I wanted to create a web-based virtual fireplace that could run on screen as an ambient background, even if the season isn’t quite right. The initial version was quite basic. As I iterated and discussed with AI, I asked for more random variations in the animation to give it a pixel art style reminiscent of Minecraft.

As for whether it will receive further improvements, that depends on whether I have time or if inspiration strikes again. If a spark of insight hits, it’ll keep evolving; if it sits untouched, it means I’m busy with other priorities or working on bigger projects. Large-scale simulation games still can’t just be generated out of thin air—my conceptual grasp of large-scale games is still something I’m developing.

Is AI actually useful? I think lately, whether you write code or never touch programming at all, you’ve probably felt its impact deeply. Although code generated with AI assistance today is a mixed bag, it has genuinely become deeply woven into current workflows. It helps quickly churn out small utilities, bridging the gap for niche personal or small-group needs that neither solo developers nor large software vendors could fully cover in the past.

Amidst my busy schedule recently, I switched over to Codex (which voice input often misheard as ChatGPT or “cold” 😅). I found that starting from Codex 5.5, its coding performance became exceptionally strong, so my primary workflow shifted over to ChatGPT (and when quota runs out, I switch back to test Gemini / Antigravity).

Looking back at other tools, Gemini went through an awkward phase of being somewhat underwhelming after 2.5 Pro, but recent releases like Gemini 3.7 Flash have once again impressed people. Google clearly realized they cannot afford to fall behind in coding models. Meanwhile, those who were previously skeptical or negative about AI are likely re-evaluating where AI can unlock real productivity, given the surge of AI agents and workflow tools.

From everything I’ve used so far, I feel large models are essentially accelerators. To put it more vividly, they’re like the Dash Mushroom in Mario Kart—for better or worse, you hit it and you rocket forward. But choosing which direction to steer and maintaining precise control over the wheel remains entirely the user’s responsibility.

After listening to many presentations and reading other people’s experiences over these past few months, I’ve come to a deeper realization: in the long run, AI still holds an immense advantage when it comes to “creating software.” This advantage doesn’t mean it is so clever that it can conjure a flawless final product out of thin air for someone with zero technical knowledge; on the contrary, its true advantage lies in dramatically accelerating your research and development process.

This doesn’t mean what you create with AI will be flawless. In the real world, software development invariably demands continuous iteration and refinement. This is essentially the MVP (Minimum Viable Product) concept—except now we can assemble an MVP at breakneck speed. Some might fantasize that a single click can complete all large-scale software engineering tasks, which remains impossible today. Yet that doesn’t make it useless; it simply shows that the way we absorb and apply knowledge has fundamentally changed.

How things will evolve in the future is hard to tell—it depends on whether hardware and software achieve the next quantum leap. But it is precisely this uncertainty and rapid change that makes the world of technology so exciting.