[loudmax]

I think the next phase will be to train GPT-type AIs to forego the compiler entirely and just write binary rather than C or code that's designed to be understood by humans. You should be able to point the AI to a binary executable file and tell it to produce a version with a slightly different behavior. Point the AI to say, /usr/bin/chromium, and tell it you'd like a version that displays tabs at the bottom of the page instead of the top. Or port it to RISC-V, or optimize for speed rather than memory use.

Obviously, ChatGPT can't do this today, but I don't see any fundamental reason AI won't be able to do this in the decade or so.

Even in this context I don't know that there won't be demand for humans that have the mental rigor to program computers. I do think we will need to adapt.

[endigma]

In compiled code? Just moving assembly around to move tabs from one part of the screen to another? This seems nonsensical, you'd have to have your magic AI bullshit not only understand the abstractions at play in, say, chromium, i.e. skia and xorg or wayland or whatever, but also be able to surgically affect not the code from skia, but specific instructions that (dynamically) position elements.

This abstraction, compiled away, into machine code that does one thing. Compilers also tend to take all kinds of shortcuts to make programs a lot faster, and therefore changing them might not be as straightforward as you think.

If you think I'm wrong, I'd like to see what "fundamental reasons" you've considered and how you've reasoned that they aren't an issue for this sort of system.

[loudmax]

Not assembly. Machine code. The ones and zeroes.

Abstractions like assembly code or C or Python are for the benefit of us humans. We can't reason about even fairly trivial x86 binaries because our minds aren't designed for that much complexity. An AI will have a completely different set of limitations.

This isn't going to happen tomorrow, but you'd be foolish to bet against it happening within the next decade. This advanced, but it isn't any more "magic bullshit" than ChatGPT or Stable Diffusion.

[lasagna_coder]

What reasoning is this based on? I haven't seen any research into this area. My understanding would be that it would be more efficient to produce its own language-like abstractions and compilers to assembly/C, which would be faster for it to write with as well.