> LLMs are steeped in decades of C++, much less so for Rust
To me it looks like a disadvantage specifically due to the "unsafe from systems perspective" reason (if you meant the low-level memory bugs). I think it's still important.
I'm an ex-C++ dev btw, but I never truly enjoyed it. After decades of complications and mistakes, it becomes another dying COBOL. To me, C++ was also always one of the least inclusive as well, and I think it's less inclusive than Rust today.
I'm not a #Rust cult leader either; I don't enjoy the unsafe semantics in it in particular (C or possibly Zig are better choices for unsafe code today, IMO). The security issues in #uutils or whatever Rust reimplementations were expected for me as well.
And I think people make these reimplementations in too naive ways: AI agents need to take bug reports/mail lists (even commit messages, if they are still available from the ancient version control systems) from these old projects into account if they really want to build something secure.
Reimplementations are inevitable though, IMO. Coreutils, for example, works, yes. Now what if we want to add any tiny little feature there?
https://github.com/coreutils/coreutils/blob/6f36b64689796bf89b57eb86265e41dc212968c8/src/copy.c#L914
Huge functions with goto this, goto that. It's not the Linux way usage of goto; it's a classic legacy anti-structural mess. And a lot of C++ projects are messy too, in some other ways; while stuff written in new languages is obviously clearer.
I also think that LLMs trained on code written by humans are becoming less and less relevant; they can now be trained solely on experience (with AlphaZero-like techniques), and the results are better:
Are you aware of this paper, #devstr, #machinelearning, #ai, #asknostr?
https://arxiv.org/abs/2505.03335
If this one was not a quantum leap in vibe-coding in particular—I don't know what was.
I find it weird I still can't find any overview of this paper on YouTube in English (however, there's a good one in Russian; I believe some non-techies can understand it). Have you seen one? Are you possibly making such videos? I think this paper is worth considering.
See my slightly simplified explanation (just under the paper link in the attached post) what it is about and how the same principle is possibly applicable to broader areas than something as formal as programming and math if the GPUs were more accessible on our VPSes.
I'd appreciate any corrections by ML engineers on my explanation or the possible prospects I mentioned, whether they are far-fetched or close to truth already (yeah-yeah, I know ML guys, many of you are NDA-enslaved, but at least wink a reaction).
nevent1q…0jsx
So depending on how the LLMs are trained, programming language complexity becomes less relevant (which still doesn't protect us from those who create PRs with who knows what LLMs). And I think that training LLMs on experience gained from a stricter environment (that Rust provides) will lead to better results. The design choices and related security considerations are still important though (but they don't really relate to language).
I'm glad if have anything that may change my perspective.
