They're achieving nothing, really.
The ability to do what AI does for them—copy/paste/adjust—has always existed. The Internet is full of samples, engines other than Three.js, and Three.js content.
It’s nothing new. Most are just grifters or having fun with a new tool.
Even technically, it is much harder to decide, design, and implement a material system and pipeline, or an asset system in general, than to make a basic renderer.
FFT waves, physical atmosphere, foliage, and terrain have existed for a while, and a student can pull them from CryEngine or Unreal. There are thousands of engine and game demo projects all over.
Engines are not about your terrain showing up; they’re about working for artists and designers and deploying to all platforms.
Games are not only about loading a level. They’re about smoothness, gameplay, story, and feel.
You can’t vibe-code a game any more than you can vibe-code a proper engine.
The moment you try something off the rails, these models fall apart. Fable or Opus output code they were trained on, not what an intern would write after seeing the existing codebase.
In my own tests, my engine uses an unorthodox, cutting-edge approach to rendering with render graphs, bindless resources, vertex pulling, and mesh shaders. Yet the models (including almighty Fable) poisoned my codebase with average Internet code, ignoring both the existing code and guidelines given right before the “go” command.
Like all rapid copy/paste/modify tools, AI is fantastic. It sped up my work by tens of times in some areas. But no, they can’t make quality stuff they haven’t seen.
Unless a carbon copy of a game from some repo is desired, they’re not going to work.