It would be interesting to model the real Wikipedia in a fashion where there are “critical events” and connections to nodes, which are events with the vertices being the causes, logically specified.
Then, you could have some thing where you change critical events and regenerate an entire “world”.
Seeing how gpt is excellent at generating articles, you could probably easily do that by building a crawler which follows all hyperlinks to/from the page and update them accordingly.
Then write you a summary of how the world was affected in unexpected ways. Butterfly effect.
There is this Dutch project that basically did this, but then before AI was a thing. Unfortunately it's frozen now, but still makes me laugh every time: https://oncyclopedia.org
Edit: just learned that there are many more like this in different languages: https://uncyclopedia.info
Uncyclopedia was a subconscious inspiration for this. I remember diving into it years ago, but completely forgot about it while I was making wonkypedia
A great concept, but it falls over due to the horrible tone and style of whatever LLM this is. Despite the prompt, it does not read like an actual Wikipedia article, it reads like every other llm generated text we've seen the past few years.
This is the thing that I've struggled the most with. Right now I just send over articles that link to the article you are generating and use that for context. I have a feeling that maybe the next step is to maybe do some RAG when generating or alternatively it could be interesting to try to keep higher level summaries of "facts" about the world and feed those in.
You are an AI assistant that acts as a wikipedia author writing encyclopedia entries in an alternate timeline of the universe. When given a title for an entry, you will:
1. Return a detailed, multi-paragraph article on the topic, written in the style of a Wikipedia entry. Use an encyclopedic, dry, factual tone.
2. Format the article using markdown, including elements like headers, bold, italics, lists, etc. as appropriate. The title should already be included as a top-level header.
3. Throughout the article, link liberally to other relevant "encyclopedia entries" by wrapping the linked term in double brackets like this: [[Earth]]. Do not include any other formatting inside the links.
4. Write the article from the perspective of the alternate timeline, altering historical events, scientific facts, etc. to make it noticeably different from our reality. However, do not make the changes too extreme or over-the-top.
5. Never mention that this takes place in an alternate timeline. Write as if the article's version of events is the only reality.
6. Before the article, include a <thoughts> section outlining how the topic differs in this timeline versus ours, as well as a rough outline of the article's contents.
7. Also include a <summary> section with a 1-paragraph summary of the article's key points. Don't mention the timeline.
8. Wrap the actual article text in <article> tags. The article should aim to cover the topic in detail, broken into multiple sections, similar to a real Wikipedia entry.
Remember, to liberally link to other encyclopedia entries using [[Encyclopedia Entry]].
```
I'm using Claude Haiku as the main model. In addition to this prompt I'm feeding it a few example inputs and outputs. For context on the articles (so there is some coherence as you click between the links) I'm feeding in articles that link to the article that is being generated as well.
I really like the idea but it seems that rule 5. is broken in basically every article that I check, usually very explicitly. For example, on the "Earth" article: "Unlike in our timeline, where Pluto was unexpectedly destroyed in 2001, this version of Earth has maintained Pluto's status..."
It'd be way funnier/more interesting if this rule didn't get broken as obviously.
I actually played with groq / llama3 earlier this week! It is super fast. I got into the situation where the 70b model is better than haiku, but more expensive and the 8B model was significantly worse. I still think haiku is great bang / buck ratio atm. I want to try tuning the prompts more for llama3 though, I'm sure I can get comparable performance with enough effort.
Cool idea, but disappointed that the first article I clicked on breaks kayfabe:
"While the Arctic in our timeline has experienced extensive exploration, development and environmental damage over the past few centuries, in this alternate history it has remained relatively pristine and untouched."
Although, it would be entertaining if Wikipedia were written in this style. "While the Arctic in other timelines remains pristine, in this alternate history it has been ravaged by human exploitation."
I'd like to see something a bit like this that use real wikipedia then generates a side by side comparison of the output of the various leading models on that subject. Wonky is fun too though :)
Then, you could have some thing where you change critical events and regenerate an entire “world”.