Login
You're viewing the cleary.au public feed.
  • David Beazleydabeaz@mastodon.social
    Oct 18, 2026, 2:39 PM

    One thing I don't quite get about all of these people using LLMs to target Rust is the idea that these same people would be able to just look at any part of the generated Rust output and easily understand what's happening or why the code looks like it does.

    I don't mean that as a knock on Rust, but as an experienced programmer, I felt like it took me quite a while to build an intuition around what Rust was doing. Even now, if you showed me some Rust fragment, it'd take awhile to understand.

    💬 3🔄 0⭐ 0

Replies

  • David Beazleydabeaz@mastodon.social
    Oct 18, 2026, 2:47 PM

    I guess what I'm saying is that if someone told me they're using AI "responsibly" because they they checked the output and they're using Rust, I'm kind of inclined to call bullshit on the whole affair because they must be using a Rust that exists in an alternative universe than the one that I learned.

    Either that or I somehow missed the learn Rust in 24 hours tutorial.

    💬 3🔄 0⭐ 0
  • Patrick Laimbock 🇳🇱 🇪🇺plaimbock@fosstodon.org
    Oct 18, 2026, 3:04 PM

    @dabeaz I guess it depends on the Rust experience of that someone. If that was said by the likes of Jon Gjengset, Mara Bos and their peers then I'd accept it. But if that someone was me, definitely call BS on it 😁

    💬 0🔄 0⭐ 0
  • Martijn Faassenfaassen@hachyderm.io
    Oct 18, 2026, 4:51 PM

    @dabeaz
    I spent years using Rust to develop stuff. I grant that Rust has some deep magic, but a lot of Rust code is pretty straightforward? I guess my experience has biased me?

    💬 0🔄 0⭐ 0
  • David Beazleydabeaz@mastodon.social
    Oct 18, 2026, 5:23 PM

    @faassen The word "years" is doing a lot of work in that comment. Would it apply to most of the people targeting Rust with an LLM?

    Of course, maybe no one is actually doing anything interesting enough to warrant deeper knowledge. I dunno.

    💬 0🔄 0⭐ 0
  • Paul Zuradzkipaulzuradzki@mastodon.social
    Oct 19, 2026, 2:09 AM

    @dabeaz projects that come to mind are Python to Rust ports of data parsers where the code is black box and verification trust is put in known test data input-output. Or fuzz the data input using two implementations and test equivalence property. Open q: what happens when coding agent hits a wall and devs need to understand the code more deeply. Just-in-time learning perhaps 😬

    💬 0🔄 0⭐ 0
  • 💬 2🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Dustin Mitchelldjmitche@fosstodon.org
    Oct 18, 2026, 2:56 PM

    @Azuaron @dabeaz so much this. Collectively software engineers say they are looking at the code ("human in the loop" being the au courant phrase) but individually we are incentivized not to: it's voluminous, we need to ship faster, the eng next to us is making more PRs, it wrote tests and those passed so how bad can it be?

    💬 1🔄 0⭐ 0
  • Dustin Mitchelldjmitche@fosstodon.org
    Oct 18, 2026, 2:56 PM

    @Azuaron @dabeaz "How do we ship slop" is the question to ask. And I think the only answer is to increase tolerance for failure and treat software components as disposable. Your invoicing component messed up? Eat the cost and have some agents rewrite it with "make no mistakes" in all caps this time.

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0