Hacker News Reader

Comment by Aicy | Hacker News Reader

You are making a mistake. ChatGPT does not care about truth, but just giving an answer that sounds like what you'd like to hear the most.

So replacing SEO blog spam with ChatGPT is not an improvement. Blog spam optimises for SEO rules, ChatGPT optimises to sound like legitimate (but wrong) blog spam.

View on HN

Honestly, I agree with what you're saying, but there are alarm bells in my head going off saying that this sounds suspiciously like what the school librarians and teachers used to say against wikipedia. And despite its flaws, Wikipedia is usually the first place I stop for knowledge.

The reason why Wikipedia is such a great first stop for knowledge is because it has citations you can follow.

With ChatGTP you have to take the information on pure trust, and the information it gives is not at a level you can trust yet.

>The reason why Wikipedia is such a great first stop for knowledge is because it has citations you can follow.

I don't know the actual statistics, but it seems like in my experience, at least 50% of the time, the linked sources are dead links, or don't have the information they're purported to have.

Why couldn’t it also offer citations then. That sounds ideal actually.

That sounds like https://news.ycombinator.com/item?id=33910863 which was posted here on HN a few weeks ago. It's a ChatGPT-esque model that provides URLs for where it synthesized each individual sentence from, but also feels a lot more summary-heavy rather than strictly new text a la ChatGPT. The results leave quite a bit to be desired, but it's nice to know there are people working on exactly what you're suggesting. :)

OpenAI, specifically, has also been working on allowing GPT models to browse the internet and include citations with e.g. https://openai.com/blog/webgpt/

Because neural network frameworks still offer no implements to trace output back to training data, or for that matter, means to compile an AI without brute forcing through parameters, and also because ChatGPT is not interpreting reference materials in text forms but talking from its "memory".

Are we sure the citations were not written by Chatgpt too?

I could be mistaken but I think another ChatGPT discussion on HN had people saying it would sometimes cite sources, and other times just make them up.

>> Explain briefly what the Fourier Transform is. Provide sources with more information about the topic.

>> The Fourier transform is a mathematical technique that decomposes a function or signal into its constituent frequencies. It allows us to represent a function in terms of the frequencies that make it up, rather than in terms of its time or spatial behavior.

The Fourier transform has many applications in a variety of fields, including signal and image processing, engineering, and physics. It is a fundamental tool in the analysis of periodic phenomena and is widely used in the study of waves and oscillations.

For more information about the Fourier transform, you may find the following resources helpful:

Wikipedia: https://en.wikipedia.org/wiki/Fourier_transform

Khan Academy: https://www.khanacademy.org/math/differential-equations/lapl...

Wolfram MathWorld: http://mathworld.wolfram.com/FourierTransform.html

If you ask for an arxiv or pubmed link it also happily makes a link up. But the link just leads to a random paper.

woah! No need for the "explain briefly" part but when I asked for 50 references it wanted money.

No https for wolframalpha?

Indeed, the current iteration of ChatGPT can't access the internet, but a future iteration should be able to look up references from reliable sources and fact check itself.

I don't know, I've had it get a lot of really basic things wrong. Where do the pilots sit in a Cessna Skymaster? What are their responsibilities? Answers with basic math mistakes too.

Its amazing at some things, but there are foundational mistakes everywhere. Wikipedia has always had much higher quality overall.

Right, this is a huge problem. Wikipedia does not stick to purely scientific topics this pages become hotly debated and political. This isn’t a partisan thing, you can find examples going both ways, but since Wikipedia is not a telling of sourced facts[1] and can not be expected to cover all topics elaborately it has inadvertently controlled the flow of peoples understanding and id say even interests.

1. it actually has policy against this to make understanding sources simpler to the layman for any topic.

from what I understand it’s trained to give the good anwser but can be misguided (like people)

I seem it lack the capability to say it’s level of confidence in the anwers. I gave me very good result when I give him lots of context. You can put pages of code about a project and then ask a question to complete. In this case it’s way more accurate (more context)

At the speed it’s progressing I am sure it will be very good in a year or two.

> […] giving an answer that sounds like what you'd like to hear the most.

How did the training of ChatGPT make it do this?

And could a different training avoid this?

Bahaha and Google does?? Paid content and garbage.

Reply on news.ycombinator.com