I don't have a dog in this fight, but a few points that look a little suspicious:
- The release with the highest number of attributed bugs is the release _right before_ the first release with Claude-coauthored commits, released in January; is there a chance that unattributed LLM-authored commits made it into this release?
- The release attribution methodology is not great, since it will tend to attribute bugs introduced in a minor version update to the longest-lived patch release of that minor version. I doubt that 3.4.1 actually introduced a lot of bugs, but since it was released a day after 3.4.0, bugs that were introduced in that release get attributed to 3.4.1.
- Relatedly, more recent releases have had less time to have bugs filed against them, so there may be a bit of a bias toward evaluating recent releases as less buggy.
You can use LLMs in multiple ways, from very hands on to make local changes to completely hands-off.
I've seen plenty of code that was LLM generated but the commit message itself did not have the co-author attached to it. This only seems to happen when someone's interface to the codebase is completely though Claude/Codex/..., and those are usually the most verbose commits, and yet they say the least, because they just summarize the code changes, not the why.
On the other hand I've seen developers using Claude as a tool. They have VSCode open and a terminal window with Claude and go back and forth, ensuring they write correct code, and leave the plumbing to Claude.
So maybe the author of the code started off small and it grew over time?
Your first and second points seem to contradict each other because if all of the bugs for 3.4.1 should be attributed to 3.4.0, that pushes the timetable back even further that unattributed LLM commits would have to have been being committed to the project, which just makes your point even more absurd.
Which brings me to my overall response, which is that there is absolutely no evidence, and nothing even intimating this hypothesis, that LLM commits were secretly being added to earlier releases before they were attributed, and that's why the rate of bugs is higher. There's no reason to think that it's an unreasonable thing to think, and there's no evidence for that whatsoever unless you beg the question and assume that higher bug counts must automatically indicate AI involvement, which is just circular reasoning. You're essentially just making up a hypothesis out of thin air to preserve your point.
Regarding your third point, that one's fair, but I've done the analysis and I can put it up if you want, as to how long it usually takes to find bugs and how far through the release cycle we are for each version.
Sorry, I should have said this explicitly in the original comment: I think you're likely _correct_ that there isn't a clear increase in the rate of bugs attributable to LLM-authored code in rsync. Your analysis provides evidence in this direction; these are just the things that made me go "hmm". They're not accusations or claims that the conclusion is invalid. But they're definitely things to be curious about.
Regarding unlabeled LLM-authored commits, I don't think it's unreasonable in general to think that an open-source project might have had unlabeled LLM-authored commits at some point before 2026. Looking more closely at rsync's recent commit history, I think it's less likely in this case. There's just a low number of commits in general, _until_ large batches of Claude-authored commits start showing up early this year. But this then raises some questions about the bugs-per-commit metric; it does correct for something like "size of release", but also obscures a significant shift in commit velocity that may be downstream of adding LLM development tools to the workflow.
Like I said, I don't have a dog in this fight, and I try not to approach sorts of questions from a position of explicit advocacy. I do think it's an interesting question, though, and we should try to understand what the data is actually telling us.
Isn't the metric that you've used "bugs per commit ~ per new line of code" going to miss the issue?
All code is technical debt.
If rsync releases used to have 500 lines changed and 5 bugs in and AI-powered rsync releases have 50000 lines and 500 bugs, it's the same bugs/line but much worse experience for the user?
I've not looked into the details of this case and I do use AI assistance coding at work but in my experience, the problem is that it's too easy to write lots of code and therefore hard to review the huge volumes of code and this analysis will ignore that?
edit: actually your table shows there weren't unusually large numbers of commits in this release, so perhaps my initial skepticism shows a bias I have?
That's sort of the point. There isn't enough data to extrapolate, and yet that's exactly what those outraged about AI were doing, and when you do do the very minimal types of analyses (permutation tests, and looking at distributions, mostly) that are actually valid, safe, standard, and useful to do on such low amounts of date, again, no evidence for the outrage shows up, and the two releases look so normal that it sort of shows no one would've cared if they hadn't known or found out that Claude was involved.
I really think this a much better standard of evidence — limited though it is — to outrage-fueled cherry-picked anecdotes, which is what has been driving this whole thing. If you disagree, and think the outrage should go one when I've shown there's an absence of evidence entirely for it (although of course, that's not evidence of absence; maybe I'll have to eat my words 5 releases down the line, but appealing to that now feels like a Russell's Teapot), would you care to explain why?
The interpretations of the p-value is also alarming. One of the first thing they teach you in statistics class is: “an absence of evidence is not evidence of absence”.
This analysis showed that there is indeed an absence of evidence, but it concludes there is evidence of absence.
Traditional p-hacking is done by oversampling and overtesting. If you do 20 analysis on average one will show p < 0.05 by random chance. This analysis is doing the inverse of that. Under-sampling, and concluding with p > 0.05
> This analysis showed that there is indeed an absence of evidence, but it concludes there is evidence of absence.
I tried pretty hard to avoid saying that, can you point me at how to rephrase? The point I'm trying to make is just that there is absolutely no evidence at all for what people are saying with such absolutism and claimed objectivity (that Claude made rsync worse), and thus it doesn't justify the outrage.
> Under-sampling, and concluding with p > 0.05
How would I avoid under-sampling here? And if you're going to say it's because I only have 2 data points, well, the side making the positive claim — that Claude made rsync worse — only had two as well, and unremarkable ones at that, as I've tried very hard to show.
You are interpreting the p-values on their own merit rather then using them to test a null-hypothesis. Quotes like:
> With a p-value of 74%, the answer is a decisive no. The odds ratio is 1.06 — essentially 1:1. Claude releases are no more likely to be above the median than any other releases.
are problematic in this context as the correct conclusion here is you just don‘t have enough data conclude whether or not you are more likely to encounter a bug after a Claude commit.
> How would I avoid under-sampling here?
You don‘t. You admit that you don’t have enough data and move on. What you are trying to do here is prove a negative, which is extremely hard to do. In your discussion you claim that the users complaining had no right to, however nothing in your analysis showed they were wrong. We simply don‘t have enough data (yet) to say either way. When we have enough data they may be proven right or wrong, but until then, we cannot conclude either way.
Unfortunately for the people mad about this, I predict the only thing they will accomplish by pressuring the rsync maintainers, is to discourage everyone else from responsibly disclosing their use of AI. You’re just going to make people disable Claude attribution on their commits to avoid drama.
I never care about AI usage disclosure, because I don't believe that human produced code is necessarily better than AI produced code, unless it's someone I personally know.
People need to be responsible for code they commit and push anyways. This has never changed. Whether the code is written by hand, by their cat walking over keyboard, or by AI, is not my concern.
A project's code quality can decline for all kinds of reasons. I don't think it's productive to laser-focus on whether it's produced by AI or not. That's a distraction. If a person just want to find excuse to criticize AI, and another person wants to fight back and defend AI, sure, go for it. But that's not how you would want to assess a project's code quality.
something as simple as requiring sign-offs like the DCO maybe relevant to people who care. I do think the driveby stuff may get smaller. People dont need to get stuff upstream. I have lots of patches I am keeping downmstrea and instead have a trigger system when new packages updates drop into debian and i rebuild the package with my patches on top using quill. Other systems like gentoo basically always supported this flow.
So - why bother forking or going upstream? maybe its selfish. I think publishing the patches are cool but I feel less of a need to force other people into doing what I want or even writing every possible configuration or solution. I just hack it for me
> You’re just going to make people disable Claude attribution on their commits to avoid drama.
People should be doing this regardless of drama. No reason to provide free advertising for trillion dollar corporations. Generated-by trailers are only relevant when contributing to third party projects, in that case disclosure is polite.
At my employer, if AI is not used, it shows up on your performance report and you’ll be told if you don’t start using it, you will be dismissed. I work at a medium sized successful YC-backed SaaS. So here, the attribution is meaningless - they look at your Bedrock and LLM API calls as well as Claude Code history.
Because LLMs are not humans, and the code they produce will have a different distribution of failure modes than human written code, so attribution is useful info while reviewing?
Nope. It cannot be assumed at all. Maintainer could just as easily tell Claude to review the hand written code you sent instead of spending any effort on it. Maintainer could sit on the patch for months on end only to swoop in later and rewrite it instead of engaging with you, thereby erasing your contribution and attribution. Maintainer could just ignore you entirely despite the pervasive "patches welcome" attitude.
If there's one thing I learned not to do in open source, it's to assume nonsense like that.
I'm referring to the fact that "open source" quite literally means "readable by humans [and machines]", and anything beyond that is a subject of debate. There are more users than readers in nearly all cases, but being able to read the code as a user is a significant benefit at times, and it's one of the reasons it's such a large ecosystem in terms of both users and contributors. (it usually being free is another big reason, of course)
Even with coding agents gaining popularity, many humans still look at the code at some point.
I see. That depends on how much I care about the project. My favorite ones get weeks of review and refinement, to the point I still consider them to be more or less hand written. Not all projects get to be that important.
Some people prefer organic grown food for all kinds of reasons, does it matter to you they would want the same for code? (Also, I'm not picking a side here)
You don't need an AI attribution tag to recognize slop. In my experience reviewing PRs, the slop-pushers are most aggressive about stripping the AI attribution anyway. It's the normal devs who use a little bit of AI who leave it in.
The tag is helpful because AI authorship is different than the human authorship. When you work with a project or team for long enough you start to trust certain people and their intuition, but when they start submitting AI-produced code you have to reset and review it like AI code.
I use these tools a lot, too. But I want to know where the code came from so I can review it accordingly. The source matters.
> Ostracize us?
I don't know why you're so defensive. If AI wrote the code just be honest about it.
If you outsourced the code writing to some guy named Bob on Fiverr, I'd want to know that too.
You're not supposed to join. You said you didn't know why I was defensive. I showed you those posts as evidence of the stigma attached to LLMs and their usage. Now you know why.
If Claude is actually good enough to commit to rsync, of course I'm going to look at that and think "it's good enough for my side project too." And (benefit to companies aside) that is info it is useful to know, if it's true.
Yeah, this is why it's obnoxious and this is why scummy marketers do it. If you don't aggressively turn it off, they leech an implicit endorsement out of you.
Is that a bad thing? I mean from the perspective of Anthropic's marketing department sure, but if agents are just another type of tool in developer's tool belt - as I see people recently like to claim - attribution feels kinda weird. In the end it is the developer who is responsible for their commits.
I mean, I don't think commits are the place for tool attributions. I want to know what the change was, I'm not really interested in your tool selection (put that in the PR if it's relevant). It'd be just as irrelevant to see "written on my macbook in neovim"
Depends on what the claude attribution actually means. A lot of people will just get the thing building and then ship. To me that attribution is generally a red flag.
I think it will be funny to watch people lose their collective minds when open source maintainers start requiring llm use.
This idea that the community can try to pressure an open source maintainers about the tools they use based off of kneejerk political reactions is so offensive.
Let's go the opposite way: "sorry I'm closing this pr because it didn't use an llm."
I'm a successful distinguished engineer within mag 7, what are your qualifications? Please send me your resume and social security number to verify that you're qualified to speak on the matter.
> It does not control for commit complexity, security intensity, or bug severity. It does not distinguish between a one-line typo fix and a CVE patch. It is a blunt instrument. But the critics' accusation is also blunt: "Claude is making things worse." A blunt instrument is the fairest response.
If by fairest you mean to say that this analysis and response is sufficient, then I'm sorry but I have to disagree. We really need to understand if the nature of the bugs are worse from a user's perspective. Even if the rate stayed unchanged, if the result is the perceived quality of the software declined then I would personally consider that worse, especially if I were a project maintainer.
That's not meant to be wholly dismissive either. But in general, I don't think quantitative analysis alone is enough to fully answer this type of question.
That which is asserted without evidence can be dismissed without evidence. This is more evidence, and of greater rigor, than was used to make the assertions. That's good enough for me. If someone wants to actually do the work to support the original claims with better evidence, great. I'd love to see it. Until then, I'm going to not worry about this issue.
But it is fair. Up to this point I have yet to see anyone say they did an analysis of the code and found X regressions of Y severity. All they say is "there are more bugs because LLM". This analysis, which you can verify yourself if you wish, says "the bugs [number of] are pretty average even with LLM", which is a direct response to that. If you'd like a more nuanced analysis you're welcome to do one and share the result, if you're so inclined.
Wait, how is any of this relevant if there were only 2 Claude commits? My statistics courses are far behind me, but don't you need at least 30 data points to conclude anything?
Depends on the methods you use. If you're trying to fit curves and so on, yes. The methods I use were designed for very low amounts of data, and are generally okay for that, specifically and especially when you're just trying to show a lack of evidence for some non-null hypothesis.
And again, that's kind of the point. There's exactly zero actual evidence, however you slice it, that "Claude broke rsync" except cherry-picked anecdata, and the whole point of my analysis is to demonstrate the total lack of any such trend/evidence at all, and just how in-distribution/normal these releases are, to show that if people hadn't known Claude was involved in them, they wouldn't have remarked on them.
Not going to critique this survey. Must have taken a lot of time and required a lot of patience. Great work!
I think it will be up to some group in academia to make a real full blown study across several repositories.
There must be tons to learn on how LLMs have changed software development and perhaps the cleanest separation will simply be going by what repositories declare e.g. "No LLM involved" vs those that proudly do the opposite or are neutral.
Bugs is not the only variable of interest here. I am guessing someone is already doing this as we discuss it here...
> The analysis uses a single metric: bugs per 10 commits (bugs/10c).
Bugs per commit as a metric papers over severity, both in terms of security severity as well as the effect on the user. A mislabeled button has the same weight as the entire app crashing in this framework.
IMO "bugs per commit" is even worse than that, because, in addition to what you say, it also hides the extraordinary spike of commit activity of a project that had previously been stable. [0]
It is the exact metric you'd choose if you wanted to make the current situation of rsync look like not a big deal.
Yes, but we know why there was an "extraordinary spike," and it has nothing to do with rsync being "vibe coded." The maintained has directly addressed this.
I link to it multiple times in TFA and quote the specific thing I'm talking about here in there to explain that possible confounder. I think I've done more than the work I'm obligated to it.do to make all of the relevant information available to you. You are just refusing to use
I am not finding these links in TFA, I see a link to an issue #929 which (as mentioned in TFA) has over 350 replies, and and opinionated summary of what transpired, including some detailed description of specific posts there. However I did not find the maintainers response.
EDIT 2 (and advice on design): The page design changes backgrounds after the results sections, which kind of conveys to the user that they have reached the end of what was is important and can just skim over the rest (usually pages have a radical change in typography like these when you’ve reached the comment section), however this is what is analogous to a discussion in a typical paper, and is arguably the most important part. I had simply assumed that you just left it at the result and skipped the discussion as a stylistic choice.
> EDIT: Found it! it is in the (untitled) discussion section (after the results).
I also paraphrase Tridge himself explicitly saying that this is why commits/releases have increased:
> Essentially, this isn't a "Claude" problem, it's a "more security work" problem, something that Tridge himself confirmed in his response, describing how a flood of AI-generated CVE reports forced rapid, extensive changes to rsync's attack surface.
> The page design changes backgrounds after the results sections, which kind of conveys to the user that they have reached the end of what was is important and can just skim over the rest (usually pages have a radical change in typography like these when you’ve reached the comment section), however this is what is analogous to a discussion in a typical paper, and is arguably the most important part. I had simply assumed that you just left it at the result and skipped the discussion as a stylistic choice.
Good point, I assumed everyone would read till the end, that's on me. I'll give it a heading.
I've now resolved this. The new version, which should be live on GH Pages soon, uses — what I think is — a pretty good methodology for assigning severity to each bug, normalizes it to 0.0-1.0, sums that, and treats that as the total severity weighted bugs, then does the analysis based on that. It did not change the analysis in any material way.
There was no analysis of severity in all of the rage posting that occurred. The single point being pushed was "use of an LLM led/leads to more bugs". The author specifically states that's what they're addressing (blunt accusation -> blunt response).
The specific problems mentioned were all reasonably severe. The original post itself described a show-stopping bug:
So my systems recently updated to rsync 3.4.3, and as soon as that happened my backup system - which does incremental backups using multiple --compare-dest= arguments - started to fail on anything but a full backup.
Incremental backups is perhaps the primary use of rsync, and they were broken for this person. That's pretty severe.
The second reply is similar:
i wondered why my 3d printers were running like sh*t and at 100% cpu; turns out log2ram uses rsync.
This one I took with a grain of salt, since it read more like a dogpile than an actual bug report. However, if it's genuine, it's also reasonably severe.
Later in the comments, someone attempted to provide a list of issues that had been added: https://github.com/RsyncProject/rsync/issues/929#issuecommen.... The list included several failures to build or run rsync that appear to have resulted from broken backward compatibility. That seems reasonably severe. If intentional, I would have expected mention in the release notes about the removal of backwards compatibility, but none was made.
The issue comments already degraded into a lot of unnecessary vitriol even before the above mentioned comment and only gets worse from there, so I stopped. But, the fact remains that the whole issue started with a severe bug.
I applaud the attempt at dispassionately analyzing whether the recent LLM releases of rsync were normal or outliers as far as bugs are concerned, but I don't think you can do so properly without analyzing severity.
To keep such an analysis fair and contextually relevant, it would have to be extended to the previous 928 issues as well (of course filtering for bug reports). I don't see anyone doing such an analysis, I think because they don't expect they'd find it useful (at least not as the rage fuel that many are seeking); what they'd be more likely to find is that there is a similar severity-mix going all the way back to v1.0.0, because these things inevitably happen whether coding is done by human or machine.
"A lot of claims in the wider discussion have treated every recent bug report as if it had the same cause. That is not accurate. Some reports were regressions from recent security hardening, some were missing historical test coverage, some were older bugs found because rsync suddenly had more eyes on it (especially by AI that can find issues quickly) and some were packaging or environment-specific failures. A Co-authored-by line is not enough by itself to establish root cause." - https://github.com/RsyncProject/rsync/issues/929#issuecommen...
I think there's evolution at play here - if you dislike AI enough to opt out of using any ai-generated code, you will likely suffer. I think there's definitely a conversation to be had about whether to disclose AI use or not but that's a separate issue if you assume that everyone is using it in some respect.
That's not what I'm saying. What I'm saying is that if the criticism is referring to a broad set of metrics like bugs per release and number of commits that were made by Claude, then it's correct to look at precisely those things because that's what the claim is about.
This is a neat post and I'm glad it got written and this is a little bit off-topic but:
Hey, 'logicprog, your writing is fine!
Use LLMs to critique your writing, check its structure, vet your choice of topic sentences, check flow from graf to graf and section to section, look for passive voice and overused words. LLMs are fantastic for that. But don't use a single word an LLM suggests in your actual writing. If it suggests something really fucking good, too bad, those words are disqualified. It's an easy red line to adhere to, easier than it sounds, and it'll keep your writing human.
(You ended up somewhere around here anyways, but that was after you posted something with LLM-written language because you weren't confident enough in your own writing. The things you do "worse" than an LLM are what make you you; be protective of them!)
This is just narrow thinking. Say Claude did increase the bugs in rsync by a negligible factor.
So what? You've saved a significant amount of time for a decent number of humans, and if those humans are working on other projects, the overall net output for the world is net positive compared to without LLMs.
You have to broaden your perspective. It's not just about how rsync was affected.
Another update: did an automated severity analysis on each bug report (~2000 of them!) using an LLM at temp=0 with a very strict rubric (and I checked to make sure that it rated things in a consistent, stable way using it). The rubric, LLM used, and some example ratings are included in the methodology section. For now, the information was just stored per-bug in the DuckDB and used to filter out non-bug bugs, to get a clearer signal. I'm going to try to use it to see if the post-Claude bugs were more severe in any way next.
> v3.4.3 has been out long enough that its rate (5.00) is already comparable to historical releases. The "wait and see" argument is an appeal to an unknowable future that shifts the burden of proof away from the critics. If more bugs surface, they will enter the distribution like every other release. There is no reason to expect a regime break.
I mean, as someone who uses LLMs, it might be a good idea to consider how one might limit the amount of bugs that will appear in the future at least a little bit: parallel iterative code review loops would probably be the easiest and most applicable to LLMs, though I guess test coverage and other code analysis tools help too.
The discussions around this have devolved to excrement anyway, I feel tempted to invoke the meme where the goose asking a guy what his jacket is made of, asks “where is your reproducer case!?” instead.
Instead we have a shitstorm over presumably legit issue, for which the only source is some mastodon post.
One command that used to work in 3.4.1 and stopped working in 3.4.3. Just one! We could have already bisected the living shit out of this and go home, but no.
As usual, Ubuntu backported fixes and didn't upgrade to a new version. Whether or not they also backported regressions in edge cases that afflict the latest rsync, I don't know. Pinning the Ubuntu package may prevent getting further regressions, but is preventing you getting any future such backported security fixes.
Did you face any actual bugs or regressions? Or are you doing this just because of the bandwagon that's going around right now? Because until you can actually present an argument for why this release is worse than any of the others, which is precisely the subject of my post, then this is not an argument against my post at all. This is just a self-referential appeal to authority.
Nah, I skimmed TFA but then I went into the linked GH issues thread, and that's the one that scared me a bit. I just want to hold it for a while and not run into some of the things I'm reading since I'm on the latest ubuntu. Just a precaution.
I didn't have the time to actually think about any "arguments" at all tbh it's just a knee jerk reaction as I get ready to log off for the weekend. Not actually looking to argument for or against your post at all lol.
Why not? Claude marks its commit messages. That there were none, and then there were, seems a signal.
Especially since if the earlier commits were so clearly AI authored yet without the Claude marker, surely you or anyone would be able to spot them. You could say, X commit does not have the Claude commit marker yet was AI written. But for all the speculation on this thread, I haven’t seen anyone actually doing that. What may be possible is that the rsync maintainers used AI to assist yet reviewed and edited themselves, as many devs do, and if so then the stats in this article are still notable: there are no poor quality outliers that can reliably be attributed to AI and if one specific release (3.4.0) was, the subsequent releases which presumably also had as much AI as this speculative hidden AI release only show improvement and thus act as a pro-AI argument.
The blog has many more datapoints than two. It compares many releases. You’re looking at 2-vs, not 2.
> Also if you write a paper where you get statistical conclusions out of whole 2 datapoints you'd be laughed out of the room
I'm using methods appropriate to that low amount of data, first of all. Second of all, since I'm only trying to show there's no evidence for the anti-AI hypothesis (not disprove it, or prove the null hypothesis), that's sufficient in itself. Also, I wonder why nobody said things like you're saying ("there's too little data to tell") in response to all the absolutist claims that AI caused rsync to get worse?
> The fact last few commits were attributed to claude doesn't mean previous ones didn't use it.
At this point, you're just positing Russel's Teapot: you'll keep assuming more and more of the code was "secretly" Claude when there's no evidence for it and no reason to think so, just because you've started with the assumption that Claude makes things worse and you want to find a way to prove it.
> If anyone complains about my verbosity or sentence structure — as they usually do, which is the reason I originally let the AI write the prose, among other reasons obsoleted by templating — they can go fuck themselves.
You can write for an audience or you can write for yourself. Which is fine either way but you shouldn't pass the blame for bad results on to your audience.
> and recieving almost no substantive input, discussion, or response on the actual content of the article
Well did you write it for that purpose?
> "Just wait, more bugs will surface" -- v3.4.3 has been out long enough
Wait for _more releases_. As your own data shows the bug rate is not consistent between releases. So this is probably not a worthwhile metric. Perhaps systems touched, new features included, or attempted fixes would be a better way to contextualize releases and the goals of the author.
And anti-AI people accuse people who use AI of being intellectually lazy. First of all, it's long because it's expanded to respond to all the criticisms. It seems that either something can be short, and dismissed as incomplete, or it can be complete, and dismissed as being long. Nice Kafka trap. Additionally, there's literally an Executive Summary section right there, for your TLDR.
What followed was extraordinary: 329 comments and counting, ranging from thoughtful concern to outright harassment.
The thread did not stop at words. One user posted My Little Pony drawings of themselves strangling the "project janitor that pushed vibecoded commits":
It spread to Hacker News and Lobsters, generating hundreds more comments.
It is neat that Lobsters has this feature (and HN should too), and I'm glad you took a beat to explain it. I think you didn't need the last sentence, though.
I have done so! that was a misremembering on my part. first mention of Lobsters is now here:
> On Lobste.rs, in response to the Medium essay Tridge himself posted in response, finally some users like boramalper begin to actually ask for evidence one way or another:
FWIW, I asked ChatGPT to review the article just for my amusement. It's conclusion was:
"My honest assessment is that this is a competent calculation performed on a badly confounded measurement, followed by conclusions substantially stronger than the calculation warrants. It is useful as a rebuttal to “the Claude releases are obviously unprecedented disasters,” but not as evidence that Claude was harmless."
- I used GLM 5.1 to help with the coding and math for this.
- However, I explicitly dictated where the data should be pulled from (GitHub, Bugzilla, mailing list), how it should be tagged and grouped, and what data to look at (e.g. bugs instead of regressions)
- Additionally, I consulted with my wife, who has a master's degree in statistics from Penn State University for what sort of statistical methodology would be justified for this very limited data set, while still giving as much information as possible.
- I know the website looks like we stereotypically consider vibe-coded websites to look, but I actually explicitly asked for that. The original HTML design looked like a website from 1995, and I just prefer how this looks. It's pretty!
No, I didn't write the text itself. I'm typically significantly more verbose and elliptical, and more than that, the numbers and methodology changed often enough over the course of the last couple days I was working on this because I was trying to get it to be as accurate and fair as possible that trying to keep the whole thing up to date manually would have been problematic.
Sorry to say but I'm absolutely certain I would've preferred to read your worst attempt at a write-up over the grating utter shite LLMs output. It's not even a question, this is unreadable.
That's interesting; IME, most people get equally angry and are as likely to disengage with a superior tone over my autism-infodump verbose essay prose as with LLM output.
Style is also part of the content. Word choice, grammar, register, and tone all affect meaning and communication of that meaning. The medium is part of the message.
So your statement betrays a significant misunderstanding - there is no neat clean divide between style and content.
Also, LLMs often generate text that is plausible, but wrong, in ways big and small.
No, the difference is in the education/experience of any given human, which is mostly gated by age. Like you'd generally expect someone young to make a lot of mistakes, and as time went on they'd learn and make fewer. Pretty much the same with LLMs, which have been around for... a bit over 5 years now? What would you expect of a 5 year old acting with intention? Or 10? Or even a 15 year old?
When you say, "I see content, not style," you are separating what is being said from how it is being said. While it is great that you can extract the core message, you are missing a fundamental truth about writing: style and content are rarely completely separate. Writing involves both.
Poor prose does not just make writing ugly — it creates friction, obscures nuance, and introduces ambiguity.
You can eat a gourmet meal out of a dirty paper bowl. You still get the calories, but the delivery mechanism definitely impacts the experience and the perceived value of the food. Same food, different response.
See? I can write slop too, I don't even need to burn down a forest to do it. If you are OK with every fucking thing being written exactly like this, good for you. I am not.
The internet is going to really suck for you if you keep that attitude, because LLM use will only increase. Though also maybe not too much as the LLM-isms will likely be fine-tuned out of them to the point that the only way you'll be sure something is done with one is if the author left a note saying such. But maybe that'll make it suck even more as then you'd be without a definite target most of the time, always wondering how much of the thing you're reading is by human and how much by LLM...
I waited a minute to make sure you weren't going to delete this post because frankly, if I had written it, I would have. Guess not, so... Here goes.
No. It is not the fault of my "attitude" that the Internet is going to suck. That is a complete reversal of the reality. The fact that even people without bad intent are already spreading slop everywhere should be enough evidence to essentially prove that there was never any hope. If this is what good actors are doing, what exactly do you expect from bad actors?
Also, to stress it yet again, I don't care if people use LLMs in general. I'll even say that I don't particularly care very much if people use them without disclosing it in most cases. If you're using it like a normal tool and not merely just dumping the output verbatim there is not any particular need to disclose it any more than you'd disclose other tools, though I think people would prefer if you did just for transparency.
My chief complaint is just how bad LLM slop writing is. It simply is not good at all. It would literally be much better for the Internet if they weren't so turboshit at writing. There is almost no writing style I don't prefer over garbage LLM writing. I'm dead serious. Early LLMs were worse at almost everything else, but they were a lot better at writing for sure. Something went wrong somewhere.
But I do also believe that it is inherently bad to dump prose as-if you are communicating as a human, but said prose isn't actually written by a human. If someone shows me a cool drawing that they made, that means that they sat there and went through the process of sketching, possibly multiple drafts, inking, coloring/shading/painting/etc. to create an expression. This involves many human skills that take years to hone, and every detail carries someone's explicit intention. I think that this is cool, and shows a great degree of skill and effort.
When you, of course, generate some crap from an image generator, it may very well look similar. It may emulate some actual defects that make it look like someone really drew it. But someone didn't. A model went directly from a text prompt and dumped out pixels on screen. No sketching. No layers. No thought processes about how to frame things or what details to include. That doesn't mean zero effort went in: I'm sure in many cases someone sat around and fudged with LoRas and inpainting for a couple hours and pulled the slot machine lever to get good seeds and etc. That doesn't mean that an AI model does not have some model for how to structure an appealing image: it does, that's obviously why the results can look decent to begin with. But when you dump out an image from an image generator and you wink wink nudge nudge present it as your own and people evaluate it as if you drew it, this is basically fraud. Everyone looking at it who doesn't know it is AI generated actually believes you went through the normal effort of drawing that image and all of the years of practicing skills and acquiring knowledge that takes. That's bullshit, and it takes away from the actual accomplishments of people who put in the work like cheating in sports does.
Like yeah, a lot of people are cheating at chess, by passing off engine play as their own, but does that really make it okay? When the entire point is using your brain and not just the raw outputs themselves, doesn't that hit you as a problem?
For generative AI, I personally draw this line at what I feel are expressions of creativity. If you use AI for drawing references, whatever. If you use AI to generate globs of repetitive code, whatever. Code can be creative but I do not view it as an expression of creativity and almost any tool is fair game. If you are using ML models for motion capture or some other data processing thing where humans had to do repetitive work before, whatever. Maybe these tools sometimes do devalue the work, but the LLMs are not doing the interesting part here, they're doing the boring part. (This is, in part, an admission that actually writing code is often pretty boring in and of itself, something that I realize programmers have been inconsistent with in an attempt to justify their value. But, I still believe it to be true.)
So okay fine. People are reluctant to disclose that they used AI to generate text because they fear the backlash that it will get them. This is understandable. What upsets me about this is that well-meaning people are apparently falling back to the idea that because LLM backlash is strong, what would be better than either trying to just simply write your own damn posts or be honest about your usage of LLMs... Is to just try to wink wink nudge nudge pass off more or less verbatim LLM writing as if it's a post that you wrote.
I am not ruining the Internet. There is literally nothing I or any group of angry mobs could do that would even remotely slow down the decay of the Internet even if we desperately wanted to.
So in fact, I'm not even trying to not ruin the Internet. I don't particularly care if my attitude is not helping or hurting. I'm not having an attitude as part of some grand strategy to save or destroy the internet. I'm having an attitude, because I am pissed off.
And I am pissed off because I am tired of reading posts the author probably only skimmed themselves.
In general, it seems HN does not like to read llm-generated articles. I ran into this myself when using an llm to edit some stuff I wrote.
At the time, I found this a bit irritating, but with a few weeks time I see the merit. The informational content tends to fall into “derivative” territory when LLM’s write stuff. And people are here for novelty and some socialization.
Also LLM prose seems optimized for engagement rather than concise communication. Takes longer to sift through linguistic boilerplate to get to the point. (The quoted bit being a case in point)
I just find it to be utter dreck. It has one of the most agitating prose styles I've ever seen. I would legitimately rather read actual broken English than the cliché polished turds Claude pops out. I am not an LLM hater, I think these tools are pretty impressive and often even useful, but even if I didn't care about the fact that I want to read communication from humans and not robots (and I do care about that, FWIW) I just find the current LLMs are horrid at writing.
And while the comments are always flooded with people like me, the upvotes seem to tell a different story; clearly LLM writing really does appeal to some people. Or idk, maybe a lot of people who vote on stories and don't comment don't actually read them. Hard to say for sure.
I'd suggest writing the lead-in yourself and boxing AI prose separately from your prose in the analysis for future articles. You can give the humanized summary/eli5/key points, then have "details according to AI" boxes that go into nitty-gritty. People seem to dislike AI ghostwriting, but most of these people still use AI, so perhaps keeping authorship clear and separate will avoid some of the flak.
This seems fair. Of course, now that I've posted this here once, I doubt it'll get constructive engagement again, but I can at least improve this for the future
Even if everything in the article is true you should not use AI to write this. A analogy would be tobacco company report on how smoking isn’t so bad for you.
So the original unfounded claim has 400+ comments because its perfect HN ragebait
The author provides evidence to the contrary and the HNers won't even engage with it instead just talking about the writing of the article in classic HN bikeshedding fashion.
How about after that we talk about the formatting of the website and the colors?
This site is really going down hill
Where is the accountability for your own opinions?
Are you guys only upvoting things that confirm your existing gripes?
This submission was heavily flagged, presumably because the article sounded like genai. But the article now says the following:
> After posting this on Hacker News and recieving almost no substantive input, discussion, or response on the actual content of the article, I decided to rewrite all of the prose in my own voice.
I've therefore turned off the flags and hopefully people can actually now discuss the claims/findings being reported.
> I decided to rewrite all of the prose in my own voice.
Soo... it didn't just sound like genai but was genai?
___
Huh. From the article:
> If anyone complains about my verbosity or sentence structure — as they usually do, which is the reason I originally let the AI write the prose, among other reasons obsoleted by templating — they can go fuck themselves.
This is kinda sad, honestly.
But also should show the author that doing what people try to bully you into doing will not stop them from bullying you.
Just stick with your unique voice man. If people don't want to read that that's fine. They do not have to. You're fine
Right so it’s gonna be a litmus test for knowledge workers going forward if they can separate style over substance. Genai tells are style. You have to be able to evaluate the ideas.
You're literally doing exactly the bullying I was trying to avoid, even while denouncing it. I like em-dashes. I have AuDHD, and they help me represent how I think.
Why did you point at the em-dashes? It looks very much as though you're accusing the author of an update that was also generated (possible but they seem sincere enough about wanting honest feedback on the content, and making changes for that). Or you're saying the author - and maybe everyone in general? - should no longer use em-dashes because they're a LLM smell. Yeah I'd feel offended too. It's a real pity I can't find em-dashes on my keyboard, or I'd stick them in this comment.
If someone gives them shit about their writing, that's on the critic for being shitty. If they use AI to write, that's on them for being fake. But, to write online at all requires being ready to have people be shitty to you and ideally not reacting in a way that makes the situation worse. Sounds like they need work on that part.
Anyway it is basically always possible for someone to find something legitimately bad about anything a person does. The question is, how much of an issue is that? Not much actually. So you have flaws. Fine, just be flawed. It had no affect on your life beyond your reaction to the attack. And putting aside that reaction is a prerequisite for learning anything useful (or discerning that there is nothing to learn) from the experience.
Good people will trust good intentions through the flaws, while shitty people will write off your work and your intentions because of the flaws (and try to make sure you feel bad about it in the process). But it's always they're too weak to express disagreement maturely, or sometimes because they're bitter and threatened by your good intentions directly. Either way, it's their flaw, not yours.
It's the ultimate product for marketers. It inserts itself as an advertisement into every conversation now and defends itself against criticism. Just crazy. There's no hope for the rest of us.
It's not defending itself here, both because I used GLM 5.1, not Claude, and because I was the one who decided to do this analysis, iterated through six or seven different methodologies to try to find the one that was most honest with the data that I had (all of the methodologies showed directionally and often in magnitude the exact same thing, but I wanted to do something that fit the purpose, in consultation with my wife, who, as I've mentioned elsewhere, has a master's degree in statistics), and, of course, I specifically chose all of the metrics and sources for the data.
If you don't want to read the LLM prose, you can just go to the GitHub of my project, grab the scripts, and run the full pipeline. It will gather the data, build the database, and run the analysis from scratch for you, and you can look at the numbers directly. It's all repeatable.
"Placement" as in where the Claude-driven releases exist within the existing distribution of bugs per 100 commits. If they're not OOD, then nothing is unusual.
A lot of people criticizing because it's heavily written with LLM, but I mean, if someone produced this piece pre-LLM, would they criticize it? is the critique due to use of LLM or due to the content being truly hard to follow? I read it and I would say, there are some problems with the writing, but its not a bad piece.
Of course this is a bigger problem, as its now harder to distinguish content that is "AI slop" with "content co-authored with AI that is carefully reviewed" with a quick glimpse, and the "AI smell" is quite off-putting. My initial reaction was also negative, but after glimpsing it through and reading the summaries, I found it decent summary, which also... speaks of this thread, of the content of the blog post and everything about the discussion and the strong feelings people have developed around the use of LLMs.
Anyhow, it would be good to disclose the repo with the code for the statistics & use of LLM in the writing right up front. Which model, and why it was used to do the writing, etc. Its enough to say "I think it writes better than I do" or "I was in a hurry, sorry" or what ever, but it really should be disclosed. It reads more honest.
> content co-authored with AI that is carefully reviewed
The problem I see is that this is indistinguishable to a reader at a glance.
Distancing the writing from the "AI smell" not only improves the quality by dropping the unnecessary ocean of rhetorical devices, it forces the human to have real weight and agency on what's being said.
I think that act of distancing from raw LLM output through refinement is a huge quality leap. Even if you're only doing the refinement with an LLM, it forces the writing to have more voice and ideas from the author.
I can see the work that went into the analysis here but again, as a casual reader, it's impossible to tell that there were any original ideas here expressed by the author.
Thank you for your constructive input, you're one of only a few others here who had any. I'll definitely do that. I didn't think, since the output was templated directly from the numbers generated by a reproducible python script, that people would get so up in arms about the aesthetics, but I guess I forgot to say that.
The most quoted line here is "A simple distributional analysis of every rsync release with bug data. No model. No assumptions. Just placement." Not only is it cringe to read, it's also nonsensical ("placement" means what?)
If OP had said "here's an AI summary of the data" and generated a conscise summary, I think I would fine with it. But default AI writing is really verbose -- the opposite of a compression algorithm, spewing out cliched phrases that don't add information. It's exhausting to read, and it lacks the interesting noise of a human response.
I mean, you can literally clone my repo, run the Python that rebuilds the database and does the whole data analysis and to end from scratch, and verify that the numbers are accurate. I made the code for this analysis public for that exact reason. This wasn't just an LLM running unsupervised in a loop. I came up with the methodologies and metrics and data scraping strategies precisely myself, iterated on it to try to be as honest with what the data could show as possible.
I think the point people are making is that when the text has an "AI smell" (it does), we immediately lose trust in the veracity of any claim being made and feel like continuing to read what is possibly a hallucinated fiction is a complete waste of time.
At this point we're all used to skimming through thousands of AI-generated sentences every working day and constantly thinking "this is likely to be 20% bullshit", it's hard to turn that off even if I try.
Do you think it would help if I went through and manually rewrote all of the prose? If it would get people to listen, I'd be totally willing to do it. It's not like I don't like writing. I just was focused on something else when I was making this, namely trying to find a good methodology that isn't insane for this low amount of data.
When there's no discernable human filter on the text output, reading the text suggests it's what the LLM produced and not what a human considered.
This is low-quality--every single day I witness Codex and Claude misunderstand, mislead, and hallucinate responses based on "assumptions" and I have to fact-check them.
If I wanted a statistical analysis and to be the human in the loop, I would ask the LLM myself, and I would definitely NOT read an article that just dumps the LLM output as-is.
Alright, I'll do that. Although, sadly, I already posted it here, so I won't be able to post it again — I'll be stuck with this trash comments section that doesn't deal with any of the actual claims, just the aesthetics.
I'm pretty sure more people would read it to the end if it didn't seem like AI output, yes.. At the very least you would have fewer (maybe not 0!) comments here saying it's AI slop.
You don't have to believe. The repository is there for anyone to attempt reproducing the results. Criticisms without proof when there's a pretty straightforward way toward that proof are pointless. Go run the experiment and rip that apart if it doesn't hold up. And until then, refrain from criticizing.
I am pretty insensitive to AI writing. I have never commented before about something sounding like AI, because mostly I don't notice. But this was so over the top that I spent the whole article trying to decide whether it was an intentional parody of AI writing style.
This article's language is not en-US. It's not en-BR. It's en-SLOP.
Yes, that was my clumsy attempt at AI parody. Here's another: this article doesn't just have AI tells. It is AI tells.
Every sentence is saturated with AI style. Perhaps the author so AI-indoctrinated that they can't see this? It doesn't read as even vaguely plausible human writing. Which is mightily ironic given the thesis of "AI generated stuff is just fine, m'kay?" The writing style does more to defeat its conclusion than the analysis itself.
As for the substance of the analysis, it seems pretty good to me but I see some flaws that weaken it a bit.
The presence of "The Outlier Nobody Noticed" proves nothing and deserves no more than a passing mention. A random release introduced way more bugs than the Claude-containing releases. That provides evidence that Claude doesn't introduce more bugs only if your hypothesis is a very naive "AI is the only thing that can ever increase bug introduction rates."
The whole analysis has very limited data. It's necessarily based off a single pair of releases at the very end of the chronological timeline. You would never be able to reject a null hypothesis based only on that, so it's even less sound to present it as proving the null hypothesis. (By the same token, it would be incorrect for critics to claim that it proves their point. Did anyone claim this, though? The heated complaints seemed more based on priors about AI code.)
"The critics' claim is a simple comparison: did the rate go up?" That's reductive. For one, these releases are known to be in reaction to a flood of (AI-discovered!) security reports, which is a novel situation and in fact is a huge confound to anyone arguing about what those two releases mean -- they're both heavily AI-written, but in response to an unusual situation. When the samples are only drawn from a distinct scenario, statistic analysis can only speak to the quality of code in that scenario.
Also, another reasonable hypothesis could be: AI-written code has bugs of a different flavor that bothers users more. It's optimized for passing tests and convincing people and AIs that security holes are closed, which means other considerations like preserving functionality can more easily be regressed as compared to if humans were doing it. (If true, it still doesn't support the claim that depending on AI code is a catastrophe, fwiw.)
I'm not arguing the conclusion is wrong. I'm saying the analysis proves far less than it claims to. As for whether it's a debacle for rsync to become dependent on AI code generation, I think that's a reasonable debate to have but it's not going to be resolved this reductively.
> The presence of "The Outlier Nobody Noticed" proves nothing and deserves no more than a passing mention. A random release introduced way more bugs than the Claude-containing releases. That provides evidence that Claude doesn't introduce more bugs only if your hypothesis is a very naive "AI is the only thing that can ever increase bug introduction rates."
It does not statistically prove anything, but as I thought I made extremely clear in the card where I discuss it, the point of bringing it up is different: to prove the hypocrisy of the anti-AI crowd.
> By the same token, it would be incorrect for critics to claim that it proves their point. Did anyone claim this, though? The heated complaints seemed more based on priors about AI code.
The entire outrage is because people noticed what they thought was an unusual number of bugs and/or regressions in the release, saw it had Claude in it, and assumed a causal link, not just "priors about AI code."
> You would never be able to reject a null hypothesis based only on that, so it's even less sound to present it as proving the null hypothesis.
The point I'm trying to make is that there is no evidence, based on these two releases, to think Claude made anything worse, whatsoever, and so the outrage is unfounded. This doesn't require me to prove Claude didn't cause any problems. If I ever made the latter claim, I should clean that up.
> It's optimized for passing tests and convincing people and AIs that security holes are closed, which means other considerations like preserving functionality can more easily be regressed as compared to if humans were doing it.
Tridge actually explicitly says he made that tradeoff on purpose, not the AI.
> Every sentence is saturated with AI style. Perhaps the author so AI-indoctrinated that they can't see this? It doesn't read as even vaguely plausible human writing. Which is mightily ironic given the thesis of "AI generated stuff is just fine, m'kay?" The writing style does more to defeat its conclusion than the analysis itself.
I've since rewritten nearly 100% of the prose in the analysis with my own, more inflammatory and verbose style. I also intentionally left in my natural mispellings and typos, to prove it was me.
My post wasn't written in a way to make friends, but:
> I've since rewritten nearly 100% of the prose in the analysis with my own, more inflammatory and verbose style. I also intentionally left in my natural mispellings and typos, to prove it was me.
Thank you thank you thank you. I would love to be able to describe how hard it was for me to think about the actual evidence you're presenting when reading about it through the AI writing, but I suspect it's one of those things where it bothers you or it doesn't. If you'd like to empathize, maybe I'll give it one try: imagine an otherwise solid PhD thesis written in crayon. The facts and evidence and reasoning are unaffected, but it's just so hard to take it seriously.
Anyway, with the rewrite I don't have to battle my kneejerk reactivity nearly as much.
I'm no expert like she is, but based on what I know, I agree with your wife on the statistics. That style of analysis is going to be the best you can do with the data available. It's an accepted way to stretch data without being too dependent on an assumed distribution. It's a good analysis. I still don't come away with the conclusion that concerns about AI code maintenance are necessarily overblown, but that's fine. I think your analysis project is a very solid contribution, and it's a hell of a lot more evidence-based than the rants people were posting.
Emdashes don't really tell you much anything these days tbh. Many languages use them regularly and those people often bring the habit with them when they write in English — me included. Plus I would imagine every major model has tuned them way down at this point due to the backlash.
I rewrote all the AI prose several hours ago with purely my own. I like em-dashes, and specifically use them with spaces as a habit. I don't know what to tell you.
> As to all the people saying “I’m going to package openrsync for platform XXX and we’ll use that!”. I find that rather amusing. If you do decide to go down that path I’d suggest you try the new rsync test suite on openrsync if you can stomach something that an AI has helped write. I tried it today and openrsync currently fails 85 of 98 tests, so I’m sure it won’t take you long to get it up to speed. You run it like this “./runtests.py — rsync-bin=../openrsync/openrsync — use-tcp”. Admittedly a lot of the failures are just features openrsync doesn’t have, but still, it’s not a great result.
- Appeals to... what authority, exactly? My fucking wife? That's me a) being really proud I married such a baddie and b) explaining the process and where the ideas came from. Which people seemed to want. Damned if you do, damned if you don't, I guess.
- All analysis is contingent.
- How do you know the conclusion was premotivated, and does it matter if the analysis, which is attempting to be as objective and extremely reproducible as possible, holds up?
- The whole point is that there's no actual evidence for what you are claiming, so why does it being highly-contingent cause a problem for me, when that just further shows there's no evidence for what the anti-AI crowd is saying?
- Why do the anti-AI crowd get to state wide, absolute, objective claims with cherry-picked anecdotes as their only evidence, but the pro-AI crowd is not allowed to respond the same way, and when we then go out of our way to respond in a far more thorough, rigorous, and objective way than you ever did, that's just more evidence for our guilt? It's a Kafka trap. You can't win.
> My fucking wife? That's me a) being really proud I married such a baddie
Good for you. I really mean that. I think people are winding you up in this thread, but keep your cool, and I admire publicly crediting and being proud of your wife. That’s a healthy relationship. Good for you.
Your analysis was so thorough, rigorous, and objective, that you couldn't be bothered to write it yourself.
Do you genuinely believe an article written by AI defending itself is going to convince anyone who wasn't already on your side? All you're doing is giving more fuel to the "anti-AI crowd" you hate so much.
Okay, so you didn't respond to any of my rebuttals — like the double standard between anti-AI and pro-AI claims, one of which gets to make claims based on cherry-picked anecdotes, and the other which must produce rigorous studies — you're just going to insult me/my work. Cool.
> Your analysis was so thorough, rigorous, and objective, that you couldn't be bothered to write it yourself. Do you genuinely believe an article written by AI defending itself is going to convince anyone who wasn't already on your side?
Except that I did. I spend days comparing and manually deciding on metrics and methodology – I did not use the AI to decide what I would do or how I would do it, so it is not "the AI defending itself" — then refining things, adding more angles to analyze, and, as I literally say in the opening section, I rewrote all the prose in the entire document just to satisfy critics like you. That sounds like "could be bothered" to me. But people like you will never be satisfied.
Also, even if I hadn't done all that work, that wouldn't make it not rigorous (it clearly is) or objective (it is as objective as it can be with so little data). You're bikeshedding to avoid the point.
- The release with the highest number of attributed bugs is the release _right before_ the first release with Claude-coauthored commits, released in January; is there a chance that unattributed LLM-authored commits made it into this release?
- The release attribution methodology is not great, since it will tend to attribute bugs introduced in a minor version update to the longest-lived patch release of that minor version. I doubt that 3.4.1 actually introduced a lot of bugs, but since it was released a day after 3.4.0, bugs that were introduced in that release get attributed to 3.4.1.
- Relatedly, more recent releases have had less time to have bugs filed against them, so there may be a bit of a bias toward evaluating recent releases as less buggy.
reply