Comment by sublinear
5 hours ago
It makes perfect sense that LLMs are good at this. We've always known that a binary is not a true moat, but that it's just a lot of toil to transpile.
Compiler output is highly predictable, but it doesn't make the inverse true. LLMs are not themselves compilers.
It's easy to train LLMs to decompile code. All you need to do is take the source code from existing training data and compile it. Then you swap the order and show it the compiled code, then the original, and judge it by how well it predicts the original.
I imagine this would go far, but also note that the variety of compilers used over the last 30 years of game dev would make this not so obviously generalizable.