Hacker News new | ask | show | jobs
Show HN: LLM-Generated Wikipedia (canonica.ai)
6 points by TernovskiiA 991 days ago
Hi there,

I've decided to jump on the AI train and put something together with low effort & high reward, to see if it can get any traction.

What do you think?

Is it a promising area?

Do you guys have ideas for me?

There is obviously going to be sea of LLM generated content out there and one project adding up to it might not necessarily be what world needs.

In the same time there is something intriguing about the area.

Well, please play with it and let me know what y'all think.

Much appreciated.

6 comments

I've had an idea on my mind from prior experiences with projects where it was community wikis.

People have bits of information but they don't want to write.

Maybe take in sets of facts, put them in a list for the llm and say "make an article out of this"

Or go through batches, "please group these by category" " please sort these facts chronologically"

Then for the image header "generate a highly stylized 8bit image of what this person probably looks like"

I was thinking moreso in a specific type of wiki though, my favorite one that almost got traction was a wiki for independent music artists.

Hmm, that's interesting! It didn't even occur to me :) I am your target audience, because I love to read wiki because of format. The thing is I don't trust AI in complicated things: history, dates, physics, maths... And a question: for example, I made "Kazan" article and saw AI-generated unreal pictures. Is it ok and that's the idea?
I enjoyed the AI images in the articles, they were crisp and beautiful (mostly). I wouldn't trust this to research anything, but it could be interesting as a way to generate a fictional world and a Wiki to document all the details.
This is horrible idea. LLMs can't preserve factual information for 100% and never will.
Why do you think that?

There are no fundamental limitations in LLM technology that precludes it from having accurate recollection of the facts, isn't that so?

Actually, LLMs can't store 100% accurate factual information, just as our brains can't. All NNs, whether biological or algorithmic, suffer from the same flaw - the probabilistic adjustment of neurons to each other. This is what makes it impossible for the information to be recalled with 100% accuracy.
I am not an expert in LLM implementations, but I don't know if that is something that is necessarily impossible to solve.

In other words, is it an inherent theoretic limitation, or a practical one, and will be solved in time?

It is impossible for now. LLMs need external factual information to be kept in check and decrease hallucinations. That is why phind.com is a good approach for improving LLMs quality, bad is that you can't limit search results to the factual only, all related and unrelated trash found by bing and advertised for money will be topped in results.
Looks interesting! Are you planning to add an AI generated videos as well?