This is amazing
Show HN: Hallucinopedia
241–250 of 285 posts
Re: Show HN: Hallucinopedia
#242https://halupedia.com/hacker-news
>Hacker News is a semi-sentient cloud formation
Re: Show HN: Hallucinopedia
#243This reminds me a lot of SCP wiki entries.
Re: Show HN: Hallucinopedia
#244Apparently it knows nothing of Age of Empires II: https://halupedia.com/wololo
I don't know how they do this, but it seems like they condition the LLM to not use real-world knowledge. Or as little as possible, maybe for the article titles only.
I tried it with a few real-world things and it correctly hallucinated.
Re: Show HN: Hallucinopedia
#245Apparently it knows nothing of Age of Empires II: https://halupedia.com/wololo
It's a feature, prompt tells the LLM to always invent new meanings to existing terms
Re: Show HN: Hallucinopedia
#246Was curious about the prompt –& especially if it referenced Borges – and found in https://raw.githubusercontent.com/BaderBC/halupedia/614eefee... >: > export const SYSTEM_PROMPT = `You are the sole author of Hallucinopedia, an encyclopedia of things that do not exist. You write encyclopedia articles in a deadpan, matter-of-fact tone — the exact register of Wikipedia — but the subject matter itself is silly, absurd, p…
If you have idea how to improve it, I'm all ears ;)
Some sort of memories-style file for topics so it can generate even more cross-references and a sort of shared world. Not for total coherence; the natural contradictions the LLM is going to generate anyhow is just part of the charm. But still sliding the scale a bit more in the direction of coherence that the "use this page's context when generating the clicked link" already leans would add some more appeal, I think.
For instance, you can build memories around times, topics, and people, so maybe specific individuals will be quoted multiple times over the course of the wiki and could build up a specific identity within the shared world.
Also... I don't know how you are thinking of this internally, but other than the issues of token spend and the $$$ involved, I would say, don't even blink at simply nuking the site at some point and starting over once you have some moderation stuff in place and other limits. Don't put it on yourself to filter out what garbage has already been generated. It's all transient content. It lazily regenerates itself anyhow. It's not precious, except for, like I said, the aforementioned token costs, which I don't deny. You can probably put some other tweaks in to the prompt to your liking at that point too.
Re: Show HN: Hallucinopedia
#247[flagged]
Re: Show HN: Hallucinopedia
#248Funny, but you could argue this is actively harmful to the web.
It's probably only harmful to the AI scrapers that train from the web. Most people will understand the purpose of this -- to poison LLM training in a humorous way, which is really easy to do. It exemplifies a major weakness in modern day AI.
Training an LLM from scratch involves carefully curating the data first. The idea that it just memorizes the whole web is a nice simplified mental model, but glosses over huge amounts of hard work to decide which websites are authoritative and on which subjects. This isn’t fooling anybody except rank amateurs.
Re: Show HN: Hallucinopedia
#249Lots of antisemitism on there. Search “Jews”
Seek and ye shall find.
Re: Show HN: Hallucinopedia
#250Earlier quoted context omitted.
The 'all articles' section really is a dive into what happens when you allow unfiltered posting - it's a shame that it isn't clear how many individuals are creating this hateful and otherwise inappropriate titles - is it just 1 or 2 people, or has this been posted to 4chan or somewhere and there is a concerted effort to disrupt the site? Shame there isn't a way to flag pages for removal. I was going to point my kids…
Interesting idea with flagging. We are considering 2 options: 1. You can generate aricle only if it was previously referenced in previous one 2. Flagging mechanism, now that you brought it up. Let me know what you think!
I lean towards a variant of option 1: you can only generate an article that was previously referenced. But arbitrary phrases, clauses, sentences, paragraphs, can be highlighted and used to form a new article.
Yes this may mean that there are pages for common words like "and"
Yes this may mean that there's a page for letters like "x"
Filtering what ends up becoming a hyperlink becomes a problem that I think can be solved with regex/whitelisting
I think articles should have a backlinks drop down. Might make consistency easier As well as generally just plain text search to pull relevant articles or context when generating a new article.