@abdimoalim_ I'd love to see what type of PL an LLM would land on given enough compute and no human-knowledge pretraining in a sandbox where all they have is some set of ISAs to interact with.
@abdimoalim_ Given that they'd prefer whatever is most token efficient to represent, they'd settle on several abstraction levels depending on what stuff they're working on.
I think at the limit it's LLM-native garble output transpiled to machine code (or edit statements) by a program.
Takings bets on how many days after GTA6 release someone mods Minecraft in the same way as it was done here 90% autonomously by an LLM.
What follows is discourse as to why did GTA6 take 10 years to develop. (that part may happen regardless)
I do feel like I can guide LLMs quite well because I understand how the full tech stack works, but I don't have strong feelings on whether this will last.