It’s basically a gpt-based-encyclopedia-like site.
All articles are generated by GPT3.5
Article/content is cached for a period after generation to save on costs.
Users can choose a few settings that alter the content so that it’s more personalized to them. For example:
Language
Language Level
Length
Expertise Level
The site tries to create a useful interface surrounding the actual generated text. This includes:
Automatically hyperlinking relevant terms in the content text.
Finding sub-sections and related topics.
Finding images from wikipedia or generating them using Dall-E (although quite slow and expensive).
I have been building it alone and haven’t gotten much feedback. Please give it a shot and let me know what you think. I can answer questions on this thread, and might even do a longer write-up if there is enough interest.
Do you think there is a future in this idea? or is the hallucination problem effectively a deal-breaker?
I just tried to search for “CreatiCode” on it, and it gave a fairly accurate and complete description. Nice work! I would suggest trying to add some links if the topic is an online entity, like the home page at creaticode.com, either in the content or in the reference.
Also, I wonder how you plan to differentiate it from Wikipedia. Being able to personalize the settings is nice, but it does not feel like a compelling reason for people to use it.
Right now I use a very basic prompt that asks GPT to describe the topic. I would need to do some work to determine if it’s a real online entity and link to it.
Good question. Right now it’s mainly a fun experiment. I don’t plan for it to ever replace wikipedia. The main differentiating aspects will be “personalized content”, better visual representations (coming later) and having articles on things that Wikipedia probably wouldn’t have. Since you can literally enter anything, it means all articles are generated on-the-fly.
I ask GPT to give me a list of “potential categories that are most relevant when organizing data and metadata related to {the topic}”.
Unfortunately I have not solved the problem of “disambiguation”. So in this case, GPT probably thought you were talking about biological trees for some reason. Thanks for pointing it out. Need to think of ways to solve that.
Well very cool stuff man. I’m sure there will be techniques that develop over time that will make it easier to narrow down. I personally think it has some real promise. And you made the interface super clean and minimal and those usually stand the test of time.
In some of my experiments I send prompts through a serious of queries first. I explain 1 method here. Maybe you can do something similar by asking the bot how likely something is to be confused with a different topic.
I made a quick example:
Debug version to get reasoning and making sure the bot understands the task.
Once you narrow down a pretty good ambiguity filter, then you can just ask for clarification or list a bunch of topics it could possibly be talking about. Single production examples will contribute practically nothing to your token usage.
Then of course there is always the embedding vector databases.
