Comment by drnick1
14 hours ago
Certainly, with Opus 5 and GPT-6, modernizing or even completely rewriting self-contained, non-critical tools seems like a Friday afternoon job now. I wonder however how important this is, in the grand scheme of things. I use GrapheneOS and the bundled messaging app is serviceable. It isn't pretty, but it gets the job done.
We're definitely not churning out code with AI models. There's no vibe coding going on in GrapheneOS. We're carefully writing and reviewing code as we've always done. We now have additional tools to assist with it. The main way we're using AI models is for code review and helping to write a lot more tests than we used to. It's helping to improve our code quality, not reducing it. We're not trying to get things done substantially faster.
The code generated by even the bleeding edge publicly available frontier models is rarely good enough to meet our standards. AI models are creating a lot of additional work for us because of how many more security issues are being discovered in upstream projects. It's also helping us raise our standards for our own work by pointing it at that to get extremely pedantic criticism catching many things we would miss.
We've hired multiple experienced app developers who have spent months working on modernizing the Messaging app and carefully reviewing it. There's no vibe coding going on for our apps. Why not look at the actual process of porting Messaging to Compose and our code review for each incremental part of the process?
https://github.com/GrapheneOS/Messaging/pulls?q=is%3Apr+is%3...
I'm really glad to read this. Thanks for putting in the effort to make quality software, far too few people are willing to do that these days.
[flagged]
Our port of Messaging to Compose was clearly not vibe coded as both the parent comment and yours are wrongly portraying it. Experienced developers have been working on overhauling the app for months with a lot of back and forth code review to get the pull requests into shape. We don't have a policy against using AI for assistance as long as the resulting code meets our high standards.
We definitely use frontier AI models with cybersecurity access unlocked for code review. Those often find problems we didn't catch via multiple rounds of human review. The output is filled with hallucinations and incorrect details but an experienced developer can sift through it, identify the real issues it uncovered and get those fixed.
How did my comment portray your port as you are vibecoding it? I just said that AI is being used.
It seems obvious to me that you guys would use it responsibly.
3 replies →
You need to stop the flagging ring operation of any mild questioning or criticism Daniel – it's a very bad look for the project, and also against the acceptable behaviour on this site
Flagging is not for what you perceive to be incorrect information or things you disagree with or to shut down conversation. You always reply to a comment to clear it up – that's enough. And of course organised flagging is totally unacceptable.
There are enough instances now I will compile a list and email the moderators who can hopefully correct your behaviour
I dunno. If they're using AI, it's well controlled. I don't see a lot of artifacts you usually get from full-on vibe-coding.
Our code quality standards are very high. Vibe coding is not capable of meeting those standards. The details of how things get written and improved to meet our standards is up to individual developers. We don't have a policy against developers using AI for assistance. They need to understand all of the code and it needs to pass our code review. Our code review is now usually a lot more thorough than before. In addition to human review, we regularly pass code through a couple frontier AI models and then review the results to figure out which claims it makes are valid. Even the frontier models hallucinate many non-existent problems and often get a lot of details wrong when they find real problems, but they do find a lot of real problems.
https://news.ycombinator.com/item?id=49592770