> > I'd also not accept big infrastructure designed by LLMs. I still hope for > > sensible use of LLMs for GCC, like finishing long neglected partial > > transitions. I do not think LLMs are ready to implement new compiler > > features. > > Yes, that seems to be a safe and sensible middle ground, in other words to > limit changes to the local "scope" so that any mistakes made are also local. > Something (slightly) above mechanical, but certainly nowhere near global.
I'd want to clarify this a little bit. I wouldn't want to accept big infrastructure changes *designed* by LLMs, but have no problems accepting big infrastructure changes *implemented* by LLMs. As an example, if I told an LLM "take this compiler and change the internal representation so that instead of using integers to represent constants, use the XYZ multi-precision package", I'd be comfortable accepting that result (assuming it was properly reviewed, of course). Indeed, I'd almost be *more* comfortable accepting that change if done by an LLM than if it were done by a human since a human is far more likely to miss a case than an LLM is!
