Christine Lemmer-Webber
Faulty towers, vibe sickness, and the vibe bobsled https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/
Reflections on how and why the use of genAI tools pulls even those claiming they will be principled towards their use to withdrawing from their understanding and participation of the creative process
Christine Lemmer-Webber
Thanks to @glyph for some good quotes and also the term "vibe sickness", though I am not sure if that is @glyph who coined the term or not.
Rich Puchalsky ⩜⃝
Here's another way "LLMs are in the air and you can't not breathe them" manifests: if you are a librarian, you index texts so that people can find them. But LLMs keep generating texts that look like human ones to such a degree that they will fool an automated indexing process. No one wants to find the LLM texts since they are slop, but -- if you are a librarian, your index still fills up with them so the human ones can't be found, even though you've never used LLMs.
Evan Prodromou
@cwebber @glyph this is a good blog post, but I'm not sure I agree with your conclusions. In particular, I agree that code review is only moderately effective; there are a lot of studies that give similar results to Yee's. But there are other more effective ways to identify bugs, especially automated and manual testing. I don't think we can conclude from the ineffectiveness of static code review that we should never run code.
Glyph
@cwebber thanks! as far as I know this was indeed a glyph original, quite intentionally trying to popularize it https://mastodon.social/@glyph/116428330599900610
Christine Lemmer-Webber
@evan @glyph I wasn't saying reviewing code isn't effective, but rather that it isn't effective for finding *bugs*, but something more important: building understanding of the system, which is the thing that is also happening less
Evan Prodromou
@cwebber @glyph The other thing that struck me is that LLMs are pretty good at this kind of static code review.
I tried asking Claude and Meta.ai about the modified Navigator.py according to Yee's description. Claude found the first two errors quickly, but never found the third:
https://claude.ai/share/9aea8f17-b435-40d0-a092-ab2896702523
Meta.ai found two also:
https://meta.ai/share/c/G6ipFYEf65
Anyways, neat exercise.
Christine Lemmer-Webber
Before anyone else sends me a message saying "You spelled Fawlty Towers wrong" yes congrats, you got the joke
Christine Lemmer-Webber
@evan @glyph Notably in the article I say:
> The thing is that all of this tooling is useful for *some* things, but the term "genAI" points at exactly what it's *worst* at: generating things. If we want to talk about *finding problems*, it's a different story. But even leaving aside the quality issues of the growing and wavering tower, there comes the problem of lack of understanding of how it is built, constructed, and maintained.
aeva
@cwebber we direly need to flip the script: it's rude to submit slop. it's rude to not disclose your contribution is slop. it is not rude for the maintainer to ask. i am so done with people copying and pasting slop to answer questions instead of putting rigor to their work.
Glyph
@cwebber @evan Given that I was heavily quoted here I am surprised that you didn't cite this one! https://blog.glyph.im/2026/03/what-is-code-review-for.html
Christine Lemmer-Webber
@glyph @evan There was a lot I debated putting into the post, but thought it was long enough! Including that article! But eventually you gotta just press publish and move on.
Christine Lemmer-Webber
@bkuhn I don't think LLMs are that safe. Maybe they're good at helping people feel safe though.
Bradley M. Kühn
@cwebber
My biggest thought from your post is: disciplined use of dangerous materials & methods.
You finish with a driving example. I've noticed in the last 15 years, drivers became increasingly more careless. IMO It's because vehicles are safer than ever before. (A runaway dog's life was saved recently because my car hit the brake faster than I could.)
Yet, I still drive like I'm in the 1981 Corolla I learned in.
People need to use LLM-gen-AI like it's a '81 Corolla even if it's the latest EV.
aeva
@cwebber my personal policy right now is that i don't accept big "generous" contributions from people I've never discussed the project with in advance, that's a huge red flag and it's not on me if they wasted their time. of course, I don't maintain anything anyone cares about so it's kind of a non issue lol
Rich Felker
@aeva @cwebber I don't get how that's not an obvious condition to everyone.
Like, if that happens, I don't know you. I don't know you're not acting on behalf of a state or criminal enterprise trying to slip in backdoors. I don't know you didn't take the code without permission from your employer. I don't know you're not a nazi who's going to turn my project into a nazi bar nobody else will want to contribute to or use once your stuff is ingrained.
If you're going to be trusted to make big contributions, you either need to be trusted by the community already, or we need to see the whole process of your contribution being developed in public so it's clear you actually did it and how you work with others.
FoolishOwl
@cwebber There was a television adaptation of My Brilliant Friend that I watched. A subplot, late in the series, involved a factory in the 1970s, and a self-taught engineer who was working out how to automate it, who started by mapping the logic the factory already followed.
I feel like that was the point where the computer industry made sense, because it was grounded.
We've layered abstraction upon abstraction on top of that, and so we have systems that not even professional developers can comprehend.
I think the gravity that pulls the bobsled is poorly understood, and I think there's the problem Ivan Illich tried to explain in Tools for Conviviality, with unfortunately problematic examples.
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.