Vibe Coding Was a Blast—Until the Character Went Rogue, So He Wrote a Language
Thariq Shihipar used vibe coding to build a fighting game, until the character behavior became completely uncontrollable. At CODE Summit 2026, he shared how to retain creative control in the AI era with readable specs and a custom scripting language.
Vibe coding feels great—until you no longer know what your character is doing.
Thariq Shihipar used vibe coding to make a fighting game. Halfway through, he realized he had no clear idea how the character worked. The behavior logic was like a black box: the code ran, but it was unreadable, uncontrollable, and impossible to debug. This got him thinking: as AI generates more and more code, how can developers retain creative control?
He brought this experience to CODE Summit 2026. The talk was titled 'The Sweet Side of the Bitter Lesson', 19 minutes, responding to Rich Sutton's 2019 classic article 'The Bitter Lesson'. That article argued that in AI research, general methods and compute will eventually beat human knowledge. Shihipar's version is: there is a sweet side to that lesson.
Full talk (from the AI.Engineer channel):
If the video won't play, click the YouTube link to watch.
He described the distribution of AI capabilities as a 'jagged frontier' and a 'smooth interior'. Where the model is strong, it's like an expert; its capability boundary is uneven, and one step over can be a cliff. This unevenness makes it hard for developers to predict what tasks can safely be handed to AI.
When visualizing model knowledge, this feature becomes concrete: the model knows some domains inside out, yet neighboring domains are as unfamiliar as another system. Stay within its strong range and it will surprise you; step slightly out of bounds and it will give you a fright.
This is where his fighting game failed. The character logic generated by vibe coding looked reasonable, but you couldn't see what lay underneath. The character sometimes did things he had never defined, and there was no way to trace the cause from the code. The problem wasn't that AI wrote badly. The developer's understanding of the system had been hollowed out.
His solution was to separate engine code from readable specs. The engine part goes to AI, but it must be constrained by a spec document humans can understand. AI handles implementation; humans define intent.
In the live demo, he wrote a custom scripting language for the game. This language isn't for AI; it's a behavioral contract for humans to read. AI generates code based on the spec, then maps character behavior onto the script. Developers don't read engine details; they just look at the script and know what the character will do next.
He also talked about Bash, React, and MCP. These technologies come from different eras, but they all point to the same question: how do humans communicate with systems. When new abstraction layers appear, people need new languages to maintain control.
Finally, the Industrial Revolution and the software renaissance. Technological breakthroughs are the starting point; what truly changes the world is the engineering discipline that follows. AI will bring a new era of software development, but only if we establish the discipline first.
The talk didn't peddle anxiety, nor was it blindly optimistic. It offered a concrete angle: the stronger AI becomes, the more we need clear human specs to anchor direction. Vibe coding is fun, but don't hand over control entirely.
The speaker is Thariq Shihipar, X (formerly Twitter) account @trq212, personal website thariq.io.
发布时间: 2026-11-20 22:55