Notices by Dennis Schubert (denschub@mastodon.schub.social)

This entire thread is now also available as a single-page outside of social media on https://overengineer.dev/txt/2025-08-09-another-llm-rant/

Don't post it on the orange site.

A contact just told me that my old "LLMs generate nonsense code" blog post from 2 years ago is now very outdated with GPT5 because it's so awesome and so helpful. So I asked him to give it a test for me, and asked it my favorite test question based on a use-case I had myself recently:

Without adding third-party dependencies, how can I compress a Data stream with zstd in Swift on an iPhone?

and here is the answer from ChatGPT 5: https://chatgpt.com/share/68968506-1834-8004-8390-d27f4a00f480

Very confident, very bold, even claims "Works on iOS 16+".

Problem with that: Just like any other LLM I've tested that provided similar responses, it is - excuse my language but I need to use it - absolute horseshit. No version of any Apple SDK ever supported or supports ZSTD (see https://developer.apple.com/documentation/compression/compression_algorithm for a real piece of knowledge). It was never there. Not even in private code. Not even as a mention of "things we might do in the future" on some developer event. It fundamentally does not exist. It's completely made up nonsense.

This concludes all the testing for GPT5 I have to do. If a tool is able to actively mislead me this easy, which potentially results in me wasting significant amounts of time in trying to make something work that is guaranteed to never work, it's a useless tool. I don't like collaborating with chronic liars who aren't able to openly point out knowledge gaps, so I'm also not interested in burning resources for a LLM that does the same.

Because it already happened: if you read my post and you feel an urge to respond with something along the lines of "it's just a hallucination", "it's just a bug", "it will be better in ChatGPT 6", or anything even close into that direction, please stop. Read this post and the next post, think about them, and if you have a factual argument to respond to me, only then reply. Focus on my factual claims, not on some inaccuracies in my analogy because of course it's not 100% accurate, that's the nature of analogies.

ChatGPT making up ZSTD compression in the Compression framework is not a bug. It's not even a weird edge-case. ChatGPT is doing exactly what it is designed to do. Let me try to explain.

If we grossly oversimplify what an LLM is, it's "just a statistical model" that generates "language" based on a chain of "what is most likely to follow the previous phrase". "language" can be anything: it can be human language, a fictional language, but it also can be code or even genetic information. Any kind of textual thing that you can feed large amounts of into a model works. "Not having an answer" is not a possibility in this system - there's always "a most likely response", even if that makes no sense.

ChatGPT inventing ZSTD compression in the Compression framework isn't due to a lack of training data. If you request an overview over all compression algorithms supported, it answers correctly with a comprehensive list that does not include ZSTD. So, if you want to anthropomorphize ChatGPT, you could say "it knows that ZSTD isn't supported", but that doesn't matter. LLMs do not possess the ability of logical thinking, deductive reasoning, or anything else. "It knows" that there are a bunch of compression algorithms available, the constants are all called COMPRESSION_[method], so there's a high likelihood of COMPRESSION_ZSTD to be the answer to a user asking for ZSTD compression in Swift. And so it generates that.

The only way ChatGPT will stop spreading that nonsense is if there is a significant mass of humans talking online about the lack of ZSTD support. For example a bunch of StackOverflow questions asking "How do I do this?" and people responding "you don't, Apple doesn't support it, you have to use third-party libraries" - or if you have a bunch of white dudes working in tech complaining on social media about Apple not supporting ZSTD in the Compression Framework.

My next post will be an attempt at comparing human thinking and LLMs generating text. As mentioned earlier, it's an analogy - and it's not going to be 100% accurate. If you want to reply, focus on the factual claims. If you only want to nit-pick my analogy, I have to assume you're not interested in productive argumentation.

Let's imagine you're colorblind. The kind of colorblindness that only allows you to see grayscale - no colors at all - but everything else is fine.

You're stressed and need fidget toy - so a friend hands you a ball, roughly filling your hand. It's hard, but somewhat squishy, and has a weird fabric-like, furry texture. You now want to know what color that ball is. But, well, you're colorblind, and your friend already disappeared and isn't reachable - probably riding a Deutsche Bahn train or something.

So you take a picture and post it to a "what color is this?" subreddit. Seems reasonable. You get 200 responses - 198 of them say "it's yellow", two of them say "it's pink". A few people helpfully say it's a "tennis ball". That's helpful, because even the Wikipedia article states that only yellow and white tennis balls are officially approved colors. Sweet.

A few days later, a random person approaches you and says "wow, cool ball - what color is it?" and you say "yellow!". Alright, end of the chat. A LLM would do exactly the same - given the "yellow" responses far outnumbered the "pink" responses, your ball is probably yellow. Ball==yellow is something both you and the LLM "learned". A few weeks after that, another friend asks you "ALice has a ball, too! Do you know which color her ball is?" - and now it gets interesting.

The LLM would immediately say "yellow". Of course it would. It makes sense. Yellow is the most likely response to that question.

But you're not an LLM - you're a human, and your brain is cool. Instead of saying "yellow", you respond "huh I don't actually know that? My ball is yellow, maybe she has a similar ball. But it could also be that she has a completely different ball that might a different color! Also, lol, I'm colorblind, so I can't really answer that anyway - you should ask Alice." And now, your brain is already doing better than any LLM. Your logical thinking engine already realized that you don't actually know something, and you're honest enough to just say that. Your job isn't to be a ball color guesser, you're just a person.

Wait, it's gets more fun! A few weeks after that, you hang out with me. You hand me your ball, and say "hey look at my cool yellow ball!". Oddly enough, my reaction is "huh? this ball isn't yellow, it's a pink tennis ball..." and now things get funky. If you were an LLM, you would either insist that no, your ball is absolutely yellow - or you'd come up with some kind of "oh, sorry for the misunderstanding - it's pink, you're correct", almost implying that my definition of color is different - and the next time someone asks you about the color of your ball, you'd still say "Yellow!!" again. Because of course, there's still only three people claiming it's pink, and still 198 people saying it's yellow.

But you're not an LLM. You're human, and your sexy human brain immediately goes into a "uhhh we have a conflict of information! how exciting! let's figure things out!" You now have to conflicting hypotheses, and you're thinking about ways to experiment on your ball to learn more. And you have an idea! You know your additive color mixing theory, so you realize that your phone camera can take pictures and you can look at the RGB values. If it's yellow, you'd expect to see lots of red and green but no blue - but if it's pink, you'd see lots of red and blue, but no green! You can test that!

So you take a photo, and... rgb(255, 0, 255). Turns out your ball is actually pink! It's still a tennis ball, but a fun one not meant for official tournaments, so it's pink! Wow! You immediately learned something new - and from now on, if someone asks you about the color of your ball, you'll say "pink!" and you'll have a heck of a story to tell alongside. Also, after some self-reflection, you realize that the subreddit your posted your image to wasn't a real "what color is this?" subreddit - it was one of those "false answers only" shitposting subreddits. Whoops.

This process of having assumptions, but being able to question them, to come up with tests for it, and to immediately change your opinion on something when you have good evidence for it is what makes humans awesome. You don't rely on the majority of people screaming "pink!" at you. You don't need to rely on manual weights that give some sources more weight than other sources - you can independently process information and deduct things. Give your brain a pat on the.. uh.. cranium.

LLMs can be a useful tool, maybe. But don't anthropomorphize them. They don't know anything, they don't think, they don't learn, they don't deduct. They generate real-looking text based on what is most likely based on the information it has been trained on. If your prompt is about something that's common and the majority of online-text is right, you'll most likely get a right answer out of the LLM. But if you're asking something that not a lot of real people had interactions on, the LLM will still generate text for you - but it might be complete nonsense. You're just getting whatever text is "statistically most likely".

If you're a coder stuck on something, identify a colleague or friend who is more knowledgeable in that specific area. They'll happily help you out and provide all sorts of fun added context that'll allow you to learn. If you're a nerd on the internet who enjoys ranting on social media, just do it yourself instead of having an LLM generate it, because that'll allow you to insert some bad jokes and a bit of your own personality to it instead of just getting a "default-feeling" text. If you're a manager in charge of something and you need to come up with new directions to push your company towards, go take a walk outside and listen to some cool music and let your ideas roam free - don't ask an LLM to generate the statistically-most-likely direction for your project, because that's by definition the opposite of creative and innovative.

Use your brains.

if you're responsible for a FOSS project, it's now time to consider how you can run your project without relying on the US.

like, using GitHub is fine. Continue using it - I do, too, and there's a lot of value of having projects centralized in one place.

but you **need** to make sure you have a contingency plan. is your repo with all branches and metadata backed up somewhere else? do you own your project's website, discussion forums, or something else that allows you to point your users and contributors to a new place if you need to?

can you continue work on your project if US-big-tech decides your very existence is no longer allowed?

now is the time to take inventory and build backups.

LLM training bots are a plague. https://pod.geraspora.de/posts/17342163

okay, I finally found a good use for an LLM. no, really.

https://github-roast.pages.dev/

this thing is brutal

I love how conservative Germans call it "terrorism" if climate activists block an intersection for an hour, but when the entire city is shut down thanks to soccer fans walking on the streets and firing fireworks everywhere, that's fine and "fun".