The Stronger the Model, the More You Need a Language of Your Own
There is a pitfall in vibe coding: it is easy to lose track of what your characters are actually doing. Developer @trq212 has an approach: readable specs, custom languages, a software renaissance. The sweet side of the Bitter Lesson.
When you vibe-code a game, the happiest moment is when AI writes the characters for you. The most panic-inducing moment is when you don't know what those characters will do.
Vibe coding, put plainly, is 'writing code with your mouth': describe the requirements to a large model and let it generate the code—great for all sorts of prototypes and ad hoc demos. It is especially popular in game development because you can turn an idea into visuals on the spot without waiting for artists or programmers.
But @trq212 points out something: as models get stronger, codebases get larger, and creative control becomes harder. You can see the characters, but you can't clearly explain the logic behind their behavior.
His approach has three parts: readable specs, custom languages, and a software renaissance.
Readable specs means writing requirements as clear, verifiable rules instead of relying only on conversation. Custom languages means designing a dedicated DSL for your game so the AI generates according to your grammar. The software renaissance means: AI writes the code, humans write the specs, and software engineering returns to its most essential form—design.
The video is titled 'The Sweet Side of the Bitter Lesson.' Here, the Bitter Lesson refers to Rich Sutton's classic essay, whose core argument is that AI progress often comes from scale, while hand-crafted prior knowledge can only win in the short term. The sweet side may be: when models are strong enough, you no longer need to worry about low-level implementation and can focus on high-level expression and control.
If you are also vibe coding, I suggest you look back at your characters: can you still clearly explain every step of their behavior?
发布时间: 2026-11-20 05:35