Comment by hammyhavoc
3 years ago
I mention this all the time.
I wanted GPT to give me an NGINX config for Active Collab as we're previously using Apache, and thus an htaccess file. I fed it all the documentation from Active Collab, I couldn't get anything valid out of it. Hallucinated all kinds of things that weren't there. I then gave it the URL rewrites that would be required, line-by-line, spent a long time trying to correct it. No bueno, even worse hallucination. I spent days on trying to get it to output a valid NGINX config that incorporated these URL rewrites. It can't reason, it's doing exactly what LLMs do, which is next word prediction.
I can't imagine what people are using it for in terms of a valuable addition to their workflow with how much it hallucinates. If it can't even do an NGINX config, what use is it for anything else? People saying it's helping them learn programming languages. Fuck me, they don't know when it's wrong, and it will be wrong at some point, it's an LLM.
There are flaws but also consider PEBCAK error.
For one next time when it starts hallucinating and a gentle course correction doesn't do it, just start a new chat with a different prompt approach. Having the error in its context reinforces the same mistake and sometimes it can't get out of this loop.
PEBKAC in what context? The prompts themselves or the documentation? Because I got Active Collab running on NGINX myself.
I already did this in terms of starting new chats, I spent days on it, and consulted with half a dozen devs supposedly using it in their workflows. It's very easy to make it hallucinate.
In terms of using gpt optimally. But fair enough. If you tried it in multiple sessions trying to convert Apache rewrites to nginx and it wasn't able to do it, I guess this is another failure mode. I just found that curious because chatgpt is usually very very good at regex.
Side note is Google extra terrible lately or is there really no docs on this almost anywhere?
All I could find about it is this and the rules looked very simple, from my experience chatgpt should have got this https://activecollab.com/help/books/self-hosted-activecollab...
1 reply →
> but also consider PEBCAK error.
No. It's an AI error, the person between the chair and keyboard just hit enter. If this "error" goes away when I hit the enter key a few more times and get lucky, it's not my fault.
Well, the error doesn't go away, it's a "fail state", but that still doesn't mean PEBCAK either.
It’s true that their training goal sometimes rewards them hallucinations, encouraging them do that, but that doesn’t mean it’s not reasoning. If anything hallucinations involve more reasoning because they have to make up something new that wasn’t there
It's doing what LLMs do. Next word prediction. It didn't "make it up", it incorrectly predicted the relationship of words.
It isn't reasoning about the solution to a problem either, it's running as expected in relating words to each other, that doesn't mean it has any form of understanding of the words or even what it's rendering as an output.
It does just predict the next word, yes, but it turns out that predicting the next word is a complex problem that involves solving many types of subproblems that involve relatively complex reasoning, or at least something that looks very, very much like reasoning
10 replies →
>If it can't even do an NGINX config, what use is it for anything else?
Turn it around. I'm sure there are lots of people around who couldn't do it either, even with a ridiculously high amount of time to do it
Yes, but we are discussing the claim that LLMs might exhibit superhuman smartness.
They are fallible, but they quite clearly exhibit superhuman smartness when compared to the average human.
As a thought experiment, assume the average human may be able to translate text between two human languages, or write code in two-three programming languages. GPT4 can perform those tasks on a much more diverse set of human _and_ programming languages. Is that not superhuman?
Yes, it makes mistakes. But take a hundred humans off the street and ask them to write an NGINX configuration or translate between Indian and French - how many would be able to do that? How many would be able to do that without any mistakes?
The vast majority of humans would fail your reasoning test.
And? We are talking about AI, not humans.
People on HN claim they're using it for XYZ in development, yet it can't even generate the necessary NGINX config, despite being given the URL rewrites it'll need to incorporate.
The point is that it hallucinates. It isn't that it failed, it's that despite giving it everything it needs to know, it hallucinated all kinds of things not in the documentation, not in my prompts et al.
Why? Because it's an LLM. It isn't fit for purpose in this context. A next word prediction AI (an LLM) isn't appropriate for these kinds of problems.
Is your argument that AI needs to be held to a different standard than humans? Because humans confidently hallucinate answers all the time.
To me, it feels quite intuitive that an AI trained on human knowledge would automatically learn to do the same.
1 reply →