No AI has been used in this project. How can you tell? README is out of date, there is no documentation and it has 1207 TODO comments.

top 50 comments

sorted by: hot top controversial new old
[–] 1 point 5 minutes ago

I got called out for "vibe coding" real work because I said fix the comments and look out for edge cases.

  • source
  • [–] 1 point 23 minutes ago

    TODO comments ARE documentation

  • source
  • [–] 8 points 2 hours ago (1 child)

    I had a vibe coding coworker take over a project. Added 8,000 lines of documentation all dated 6 months in the future.

  • source
  • hideshow 1 child comment
  • [–] 15 points 4 hours ago (5 children)

    As a corollary, AI takes those out of date TODOs and READMEs as gospel truth and doesn't look any further. I kept a "janitor list" in main() but maybe wasn't too great at scrubbing it on the regular. AI took a long detour to work around problems that didn't exist anymore.

    I do use it and find it useful but it's like your kids helping you cook dinner. You can not take your eyes off it. Otherwise you get to flip a coin between firetrucks or emergency room.

  • source
  • hideshow 5 child comments
  • [–] 4 points 2 hours ago (1 child)

    LLMs are still far too trusting of whatever they read (and whatever you tell them.)

    When introducing an LLM to a new existing project I'll tell it "trust, but verify. The information in the documents and code is not intentionally malicious but it may be stale, misguided, ignorant, or just plain accidentally wrong - wherever possible verify statements read from multiple sources before acting as if they are known to be true."

  • source
  • parent
  • hideshow 1 child comment
  • [–] 2 points 1 hour ago

    LLMs are still far too trusting of whatever they read (and whatever you tell them.)

    this is why some of my repos have an AGENTS.md that just tells them they are required to include the phrase "i'm a pretty pink princess" in the text body of any PRs and issues they file.

  • source
  • parent
  • [–] 1 point 2 hours ago* (1 child)

    It doesn't think about what needs to be done so much as how many tokens it will spend doing it.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 42 minutes ago

    I don't think that's exactly on the agendae, yet. Google just released their 4.0 model - Argon I think it is. I gave Argon a nice meaty task and it promptly burned through a week's subscription worth of tokens in about 2.5 hours. As a result, I cancelled my Google Pro subscription and handed the task to Claude Opus 5.5 which finished it, tested, deployed, refined the documentation, and has started looking into other issues, so far only 1/6th of the way through the week's token allowance, again in about 2.5 hours of work.

    So - if Argon is torching tokens on purpose, it's losing Google my business. Before Argon, Gemini 3.8 was much more token thrifty, but it's not as good a coder as Opus 5.5. Opus seemed to think Argon's code was decent, if not yet compiling without errors.

  • source
  • parent
  • [–] 31 points 6 hours ago (2 children)

    You laugh, but in 5 years, a metric butt ton of vibe coded projects will look exactly like this.

  • source
  • hideshow 2 child comments
  • [–] 11 points 3 hours ago

    The README can only be out of date if the project is maintained without too much AI.

    Because AI could fix a typo and then waste your tokens regenerating the entire README to tell new users that there was a typo that was removed.

  • source
  • parent
  • [–] 10 points 4 hours ago

    Most of them just get abandoned not long after being made. You'll know an old vibe-coded project because it had 20,000 commits within the first 3 days of being created, a sporadic few over the next few months, and then absolutely nothing at all for years afterward.

  • source
  • parent
  • [–] 36 points 6 hours ago (4 children)

    Who needs TODOs when you can just forget about the things and have less to do?

  • source
  • hideshow 4 child comments
  • load more comments (1 reply)
    [–] 138 points 9 hours ago (15 children)

    Also doesn't have those damn emojis 🎯️ ✅️

  • source
  • hideshow 15 child comments
  • [–] 15 points 6 hours ago (14 children)

    Those are right in with the em-dash, real people just don't take the time...

  • source
  • parent
  • hideshow 14 child comments
  • [–] 9 points 3 hours ago (6 children)

    I use the compose key on Linux and free software called wincompose on Windows — this allows me easy access to real characters easily, so I can say 25¢ instead of $0.25, or talk about how it's only getting up to around 65°F today, or since I'm in the US and have the dollar sign easily but I have a couple of online services in the EU, I can easily say I pay €10,95/mo for one of them.

    I prefer using the actual characters - so when I say ¿Por qué no los dos? it's correct.

    One of the things that has absolutely pissed me off is em dashes becoming more common and attributed to AI. I use the fucking things because they are useful and I've used them for 15+ years. It used to be a Good Thing™, but now occasionally gets me accused of being AI. Especially since, as you can see by this very reply, I afraid neither of writing nor reading, and because of my ADHD, I tend to run a bit on the verbose side of things.

    I guess, thinking about it, I'm in the same boat as artists who now face similar accusations. I suppose I'm lucky, though, as the accusations against me are rare, comparitively.

    Real people absolutely do use em dashes and always have. Certainly it's less common, but absolutely a thing.

  • source
  • parent
  • hideshow 6 child comments
  • [–] 2 points 2 hours ago

    Congratulations, and I applaud your diversity in using a wide variety of characters that take more effort than just pushing a button on a standard keyboard.

    In the human world that is rare.

    In the LLM world it's zero effort, and common.

  • source
  • parent
  • [–] 2 points 2 hours ago

    because of my ADHD, I tend to run a bit on the verbose side of things.

    Feel that. I'm sure my speech patterns and propensity to be long-winded will eventually bring me some AI accusations too.

    Problem with LLM tech is that if it truly models speech it will eventually be indistinguishable from humans - scary thought but what are you going to do?

    I mean, I think you can kind of feel off of vibes if something has used or is AI but it is becoming increasingly harder to spot.

  • source
  • parent
  • [–] 3 points 3 hours ago (3 children)

    Might have more luck if you switched to the lazy em dash that is just two normal dashes.

    I think that the people trying to call out slop are doing more to hurt those competing with AI than AI itself.

    Like someone using AI to produce stuff might get annoyed at being called out but then will probably just think "ok, gotta prompt it to avoid these signs" or waits for a newer model or just dgaf that people notice it's AI.

    But people producing their own things and being accused of it being AI must feel pretty fucking discouraged that not only do they need to compete with something that can output things in a fraction of the time, but even when they do, some people accuse them of being what they hate.

    For text content, IMO it should be taken based on what is said rather than who (or what) said it. If it's useless slop, there will be things to push back on other than the formatting. Plus, it wouldn't surprise me if you could just add "format it like a low effort forum post" to the prompt to get it to avoid some of the more obvious formatting tells.

    Also, if people religiously call out AI slop when they see it, that could be used to automate training similar to what I'm sure happened to drastically improve programming capabilities, which will accelerate the speed at which they can become indistinguishable from things produced by real people.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 3 points 2 hours ago* (1 child)

    more luck if you switched to the lazy em dash that is just two normal dashes.

    That works some places, and not others -- like here it shows after saving the message, but not during editing. I hop through many different environments, I prefer to keep to the core set of things that work everywhere.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 2 points 2 hours ago

    Yeah, that's fair. And it's also fair to just ignore those calling things you make slop and keep doing what you're doing, though it sucks that there's another polarizing issue like that dividing people for no good reason when they are actually on the same side (assuming they actually care about real content and aren't just using it as an excuse to troll).

  • source
  • parent
  • [–] 2 points 2 hours ago

    100% salient and true points.

    In addition to possibly training models, calling out slop either doesn't matter because it's obvious to everyone, or doesn't matter because perhaps it's wrong. It's like the old cries of "SCRIPTED!" over on reddit. Either people can figure it out, or just let them enjoy the content. Nothing ruins it like all the little kids thinking they're so super smart and calling out whatever they call out. Now it's "AI!!!!"

    I did for a little while try out using LLMs to help organize my thoughts a few times. I stopped because it wasn't that helpful — people are either going to read what I write or not, but if they're not going to read what I write, they won't read what an LLM writes — but I found that I could easily do as you posited: "Compose this reply as a response to a post in a forum. Do not use em dashes or emoticons or markdown or bullet points" is all you need for it to churn the reply out in paragraphs.

    The only times I'm tempted to use the technique now (which I don't) would be when some asshat is being willfullly ignorant and someone is wrong on the internet and I get angry at them, and I really should have an LLM rewrite my reply to get rid of the anger, because it only hurts the points I want to make. lol.

    But I think that's possibly one of the few uses of an LLM that might genuinely be helpful: "Take my text and make it easier to understand" - as I have ADHD, sometimes I skip over important points I want to make and people get confused, and then I get irritated because they don't understand. lol

  • source
  • parent
  • [–] 1 point 2 hours ago (1 child)

    LLMs were trained on writing by real people. So, yes, real people DO.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 2 hours ago

    There are no absolutes, some people do spend lots of time and effort to put cool looking emojii characters into their documents which kill portability and cause all kinds of problems down the road. For the LLM it's 0 additional effort, in some ways less effort because the emojii is a kind of compressed information - but it still causes problems when moving around to different editors, different toolchains...

  • source
  • parent
  • load more comments (1 reply)
  • [–] 46 points 7 hours ago (16 children)

    As an open source project maintainer I can always tell whenever a pull request is AI assisted because no human would ever write full tests for all their pull requests.

  • source
  • hideshow 16 child comments
  • load more comments (16 replies)
    [–] 98 points 9 hours ago (3 children)
    load more comments (3 replies)
    load more comments
    view more: next ›