Hacker News Reader

Comment by creatonez | Hacker News Reader

These AI chatbots on websites are beyond useless, and LLMs won't make them any better. When you go to the chat-based support interface on a website, it's usually because you've read the entire contents of the website and didn't find what you are looking for. Now you will just get a hallucinated answer, with no indication as to whether it's a human or AI you're talking to.

View on HN

> When you go to the chat-based support interface on a website, it's usually because you've read the entire contents of the website and didn't find what you are looking for.

Because chatbots were so far utterly useless.

It doesn't seem crazy to think that given a good enough chatbot, users might prefer to ask their question directly rather than have to find the specific piece of information they need from a dense docs website.

> > When you go to the chat-based support interface on a website, it's usually because you've read the entire contents of the website and didn't find what you are looking for.

> Because chatbots were so far utterly useless.

My guess is that you are correct. I have been thinking that rebranding site chatbots will be needed and inevitable. I wonder what that will look like.

As someone who spends a lot of time answering user questions in chat support, I sincerely wish more people bothered to read the website or at least try the search box first.

Disagree. I now defer to chatGPT instead of reading raw documentation. Even if it can hallucinate an answer it's still way faster and better for discoverability.

Even better. When provided with your context (use case, models,…) it can adapt its answer to it.

Although that can also lead to hallucinations, it’s still quite wonderful.

Even better, just ask your stack

https://ask-your-stack.vercel.app/

(It uses official docs to provide answers with context)

Quite nice, but by definition limited to the stacks it covers, isn’t it?

It's open source, so you could certainly add any missing pieces, or even custom code which docs are not public, to your ask-your-stack* version.

At the end of the day with LangchainJS + LlamaIndex anything is possible

Those ones are using inferior models. I am using ChatGPT for a lot of stuff: personal, work, hobbies. Anything where what I say isn’t private. And it is wickedly helpful.

Extending this to sites makes sense. Eventually this service will need to compete against Google or Bing chat based search with regular indexing and it’ll probably get put out of business unless it pivots into tailor made models or something else the big guys can’t offer en mass.

If it knows the answer to the question the likelihood of hallucination goes down from what I understand. There is a lot of knowledge on many specialised forums hundreds of thousands of question and user generated answers getting chatgptlike bot be trained on those would I personally feel very useful.

Reply on news.ycombinator.com