MyHub-ChatGPT integration (newsletter 1, November 2023)

Over the summer I put the "ai" into myhub.ai, turning my Hub into a personalised GPT wrapper. I've been experimenting ever since.

Briefly

• Today I'm sharing my first results with you via a massive wiki - essentially my experimental notes, published online (see About this site).
• Pretty soon I plan to open a pilot programme so "early adopter" Hub Editors can play for free. But if you have a ChatGPT API key let me know and I'll fast-track you right now.
• Either way, you might want to Hub some more resources to give yourself more knowledge to play with.

Experiments to date

My current experiments focus on finding the "sweet spots" where a large language model like ChatGPT can best help Hub Editors extract more value from their collections of knowledge and notes. Since last week, moreover, I've also been investigating how your visitors could chat with your Hub via GPTs on OpenAI's upcoming GPT Store. 

For now, this short video sets out how the pilot integration currently works, as well as the fuller integration planned next:

I've launched six different experiments so far, focusing on:

• experiment 1: summarising resources for newsletters,
• experiments 2-4: creating knowledge visualisations to better understand the main themes connecting a set of resources
• experiment 5: creating new ideas by combining related resources together.
• experiment 6: the same as experiment 5, but using a GPT.

Experiments 4-6 are still underway so I haven't written up all my findings yet, but so far I have:

• a reasonably good prompt for creating simple newsletters
• promising ideas for creating new ideas from a collection of resources
• leads on creating useful visualisations of a collection of resources.
• a first glimpse (as of last Friday) of how GPTs could be integrated for both Hub Editors and visitors.

But I have even more prompts which produce results which are impressive at first glance but then, upon careful examination, are more trouble than they're worth. More on that later this year.

Examples: one week's resources labelled "llm"

There are now ~250 resources on my Hub tagged #AI. Most of those added recently are also tagged #llm, for Large Language Model.

This is certainly the case for the 8 resources I Hubbed one week in October. An AI summary, courtesy of one of the prompts tested in Experiment 1, can be found in the experiments wiki: C-5-AllNotes-response-P1b (scroll down to ChatGPT's response).

But while I would have loved to give you an AI-created editorial summarising those resources, results have been reliably unreliable. However, that's not always a bad thing:

"ChatGPT's hallucination, in other words, was a connection that the author did not make. While that makes it a false summary of the author's work, that doesn't mean that it's wrong as a stand alone statement. But it also doesn't mean that it's right."
- Experiment 1 analysis > editorial content

Experiments 5 & 6, still underway, build on this to focus on exploiting ChatGPT's creativity by asking for ideas created by combining resources. Examples ideas from last week's resources ranged from AI-Enhanced Creative Writing Tool ("a browser extension similar to HARPA AI that integrates with AutoGen and leverages its conversational capabilities... generating creative content, helping with research, and even suggesting conversational dialogue for characters in stories") through to AI-Enhanced Fact-Checking Browser Extension ("combines the fact-checking capabilities of knowledge graphs with Google Bard's information retrieval abilities. Users can verify the accuracy of information on webpages in real-time..."). All 10 ideas can be found in C-5-AllNotes-response-P5.md.

Much more is on the way, from a post I plan to publish on Medium written by my first GPT through to more progress on knowledge visualisation. So hang on, this is going to get interesting.