9月20日 星期日
日记内容
最近在用 AI 写小游戏,体验很有意思。
一开始用的是 DeepSeek,写起来很顺,但 token 消耗得飞快——一个简单的小游戏往往要烧掉不少额度。后来换了 MiniMax(M3),效果不错,性价比也高。再之后试了 Codex,最终做出来的效果最让人满意。
参考了 https://github.com/htdt/godogen 这个项目的开发思路,前后做出了三个小游戏:
每个都不复杂,但都是完整可玩的。AI 能从一句需求出发,自动拆解任务、写代码、调试、再迭代——这种”从想法到上线”的链路,以前要几天,现在几个小时就能跑通。
印象最深的是 AI 模型的进步速度。几个月前还经常需要反复纠正它的逻辑错误,现在大多数时候一次就能给出可用的方案。开发者要做的,反而是更准确地描述需求、把控节奏,而不是事无巨细地写实现。
这是一次非常有趣的尝试,也让我对未来开发的方式更有信心。
Diary Content (English Version)
I have been using AI to make small games recently. It was a very interesting try.
At first, I used DeepSeek. It was easy to work with, but my tokens went down very fast — a simple small game would use up a lot of them. Then I tried MiniMax (the M3 model). The result was good, and the price was fair. After that I tried Codex. In the end, the result I got from it was the most satisfying of the three.
I looked at the way the open project https://github.com/htdt/godogen was built, and made three small games one after another:
None of them is complex, but each one is a full game that you can play. AI can start from a single need, break it into small tasks, write the code, fix the bugs, and try again — this “from idea to live product” path used to take days, but now I can get it running in only a few hours.
What strikes me most is how fast AI models are getting better. A few months ago, I often had to correct their logic again and again. Now, most of the time, they give me something I can use on the first try. What we as developers need to do is not write every small step, but say clearly what we want and keep a good pace.
It was a very interesting try, and it makes me more sure about the way we will build software in the future.

