
denschub@schub.social ("Dennis Schubert") wrote:
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 then, 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 ofCOMPRESSION_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.