>[X] is not something an LLM will ever be very competent at
I'm no longer convinced that this is a particularly safe bet for any value of X. LLMs have absolutely been demonstrated as capable of doing things that they're theoretically the wrong type of model for, like generating images by writing SVGs or simulating brush strokes.
And then they wonder why the their ugly-ass game with no consistent art style and no coherent design doesn't take off since it was coded with a superior artificial intelligence, in less time, and must therefore be objectively better.
They don't understand that there's some knowledge that exists within the heads of human beings and their experience that you can't reverse engineer from the human-generated artifacts in the training data. You can't derive Leonardo Da Vinci from a PNG of the Mona Lisa. Am I claiming that all designers are Leonardo Da Vinci? No, so hands off the keyboard. I am saying there are aspects of art that are bound to the human experience in a way that are beyond current LLMs, that can't be replicated through the simulation of mechanics alone.
Say that and they'll just think you believe in ghosts, though.
It is fantastic at the code. Monobehaviors are completely solved. I ~never have to look at c# code or shaders again if I choose not to. However, this was like 5% of the problem.
Composing a three-dimensional scene graph in a way that is compelling to a human observer is not something an LLM will ever be very competent at. I worry about those who are in the midst of conflating the capabilities of diffusion models with those of "world" models. The former make it seem as if we are "so close". The latter are more or less non-existent.