in reply to Ariadne Conill 🐰

i honestly do not know how i feel entirely about vibe coding? i think it is cool that people can theoretically get any program they want at any time.

but that's theory.

in practice, the code the tools generate has a tendency to be unreliable and frequently also has security issues.

and rsync is being vibecoded by just tridge without any supervision.

in reply to ✰ Alice D. ✰

@AmyZenunim that wasn't the question.

let me break it down:

1. alpine is interested in a reliable rsync implementation.

2. we presently use rsync, which is GPL, and now vibe-coded.

3. openrsync is an alternative rsync implementation, which is maintained by the OpenBSD project, and thus ISC licensed.

4. if we repeat this cycle over and over, to avoid other regressions from other unreliable vibecoded software, then the pool of influential GPL software wanes over time.

in reply to Cassandrich

@dalias @AmyZenunim rsync going slop is wild, because: chaos.social/@dentangle/116657…


Indeed. It was a shock.

Few people have done so much to fight for Open Source as Tridge has.

Samba was a key part of the whole anti-trust suit in the EU against Microsoft, and Tridge was at the front of that.

Git exists because Tridge wanted an alternative to the proprietary BitKeeper for the Linux kernel.

It's as though Elrond said, "Hey, that's a shiny ring you have there! Do you mind if I try it on?"

😢


in reply to Ariadne Conill 🐰

anyway: mad respect for tridge.

the man has done far more for software freedom than most of us have.

but he is still a person, and people can easily be convinced by these LLMs that things check out when they actually don't.

they use very persuasive language. if you depend on them, you will inevitably commit mistakes that you should have caught, because nobody does a perfect job. nobody.

in reply to Ariadne Conill 🐰

what I will say is this. there are pieces of software that are frankly "mission critical".

for example, pkgconf, as a key component of most build toolchains, cannot have regressions because those regressions will reverberate throughout the entire "software supply chain" in the form of build errors. it is a mission critical piece of software.

this is why as lead maintainer of pkgconf I have implemented a number of policies and initiatives to reduce the likelihood of software errors and promote correctness in pkgconf as part of the pkgconf 3.0 work.

these initiatives include banning LLM contributions, requiring DCO signoffs on commits, refactoring the codebase to remove entire classes of vulnerability, improving the quality of the windows port so it is equivalent to its unix counterparts and reimplementing and expanding the test suite from scratch.

why? because every single thing I listed reduces the likelihood for regressions.

rsync, like pkgconf, is used at all times of the day, all around the world. I try to visualize the scope to which pkgconf is used and it is just not possible.

rsync is the same way: everyone is using it somehow, either to back up their data, or to mirror data from one machine to another. there are numerous utilities which make use of it somehow to provide functionality.

a regression in rsync is even less tolerable than a pkgconf regression: if you have errors in rsync, they can potentially cause data corruption or loss.

but rsync goes in basically the opposite direction from pkgconf: it embraces LLM contributions. it also has had several regressions since doing so.

reshared this

in reply to Ariadne Conill 🐰

People at my job have suggested having the coding agents review themselves, because since they are nondeterministic, they'll notice different things the second time. Or, similarly, have different LLMs review one another.

I'm just sitting over here in my devops corner 😑 😐 😬

in reply to Ariadne Conill 🐰

in theory i do think it's really cool that a nonprogrammer could make a program without going through a lengthy process of learning how to code. however, the history of low code and no code (and prior synonyms and related ideas) makes me wonder if that's something many people really want and whether it will be undermined by increasing complexity (or even existing complexity).
in reply to Ariadne Conill 🐰

My work experience has been interesting. I do catch the LLMs introducing some things that if a human did them I would describe as very poor judgement (I'm not ascribing judgement of any sort to the LLMs). On the other hand, I've had them catch subtle downstream impacts that I missed, including avoiding introducing bugs that would have been a pain to track down. On balance they are improving, I think.

But I also don't count myself as a second reviewer and the LLM as the author; I am the author using the tool, and I still want real third party human review. Confirmation bias is there — I asked the agent to build something, so I'm clearly predisposed, even when consciously trying to read its plan and its code skeptically, to accept it at a light reading.

I'm not sure this is much different in practice from trusting people who have learned how to project confidence in their writing. The machine is statistically likely to produce writing that is at first glance like a person confident in their own analysis. I think it's a difference in degree, since the machines create so much more of it. But I recognize the same temptation to accept specious confidence.

in reply to Ariadne Conill 🐰

Oh, I was actually considering packaging openrsync myself in response to the recent release's bug reports (as a fail-safe for pmOS), but nice to see you beat me to it. I don't know enough about openrsync other than it is an OpenBSD version to really make any solid statements other than thanks 😀

This website uses cookies. If you continue browsing this website, you agree to the usage of cookies.