I know this might not be an immediate feature, but I am curious how people are thinking about the idea of integrating OpenAI’s GPT-3 into the rich text editor? Or maybe automatically generate a field based on the prompt from another field?
I don’t actually think it would be best in Rich Text for Fibery (although for personal reasons, since I use Fibery as a PKM and “Tool for Thought”, I would be happy to have it available there!). Something more like this seems like it would bring the most immediate utility for lowest effort:
Others are doing similar things:
So imagine an Advanced Field for GPT3 I picture it having a simple formula-like setup that allows you to reference other fields, so e.g. “What is the address in [input field]?” or whatever, is passed as the prompt, with the value of [input field] for each Entity. Another way to use it, then, would be to reference one field to process, and the other as example output for GPT3 to reference. It should be really simple to use though, not a full-blown formula editor IMO. And it should ultimately be natural language-based, that’s part of the whole appeal of GPT3. A bonus could be to have a GPT3 function added to the overall formula system.
Agree with you on both fronts. I think this could be the future of the productivity market, in general, to leverage the advances in AI to speed up tasks and create new insights.
Interesting, Notion just announced a clearly GPT-3 based (or at least very GPT-3 like) AI writing function:
It makes me start to wonder… I typically think of Fibery as a very clear and direct Notion competitor (a better one IMHO ), yet somehow I wasn’t feeling like Fibery would or even should go down the AI-in-free-text route. But here is Notion showing that there are at least some reasons to do so in the context of their users.
Now there are a couple of things that are top of mind for me with this. The first is that in my view their implementation of this is probably in large part a marketing play. Not to say that AI is not of use in Notion, but I suspect it was prioritized for implementation because A: existing APIs make it pretty easy to do and B: AI is a super hot topic right now and implementing it will get them a lot of press! Just from reading the Twitter comments on their announcement it’s clear that a lot of people have no clue that the Notion AI feature set is almost exactly the same as 10+ other tools, all of which (no surprise) are based on GPT-3. The real magic, of course, is in OpenAI’s large language model and the fact that they make it openly available via a nice API.
The second thought I had is that it’s arguable that Notion’s customer base is much more heavy in individuals and small content creators (YouTubers, bloggers, etc.) than Fibery is. This is not so much because either one is fundamentally more suited for a different customer base, per se, I think it’s more about marketing and reach, and it ties-in to some discussion we had in the “Lack of work management features” topic. Michael himself seems to feel Notion focuses more (or at least is more used for?) a writing-oriented use case:
Functionally speaking Fibery is not far behind Notion for Wiki uses though. It only lacks the breadth of embedding options to really get there, IMO. But… Fibery is not very focused on that use case. Although there is perhaps some room for that to change:
But outside of the free-text cases that Notion outlines (idea generation, summarization, proofreading), the GPT-3 type of AI does not really have as much wow factor. Implementing it in databases as an advanced type of Field would definitely be great, as I mentioned before. But now I’m starting to wonder if Fibery should just implement it in free text in the / menu just to tick the “Has AI!” box, largely for marketing purposes.
If my understanding is correct it’s pretty trivial to implement via the API, it could be a few days of work, and would make a fun announcement from Fibery team. I can just see it in their typical, dry, sarcastic style: “We saw everyone talking about AI so we decided to try it too. Here’s some AI stuff you can do now in Fibery. Give it a try, maybe it won’t disappoint you.”
Sure, AI is a good marketing these days. To be honest, we didn’t think about it in depth and so far have no good ideas how to improve Fibery with AI. What I saw in Notion does not impress me, but maybe I am wrong, we’ll see. We started to accumulate info, but I doubt we will jump into it in the next 3-6 months. So far our priorities are more basic:
- Rich Edit improvements (Views, Embeds, Files, …)
- Collaboration (new comments, chat, notifications)
- General polishing and unification (better whiteboard, new tables, …)
Those are certainly good priorities. Did you look at Promptloop, etc. examples I posted above?
I think the priority makes sense. But I just want to say that personally I find the GPT3 in writing quite helpful. I used a plugin called “Obsidian Text Generator” a lot, when I write in Obsidian. That plugin uses GPT-3 (you need to register and setup account at OpenAI).
A post was merged into an existing topic: Auto-Generate linked diagrams in Whiteboard
Mem.ai has a lot of decent examples of AI use case in free text:
We are already having some fun internally
Summarization of Intercom conversations works pretty good I’d say.
I can imagine you can add a special field like Summary to Conversation database and summarize it automatically in future based on special Command we will add in automation.
I am also not so impressed by the way most productivity tools (seem) integrate GPT-3. All it does it generate text based on prompt for a rich text field or so. Depending on what you do, that might be useful, but it feels like it’s just a glorified way to not have to sign up and give it a prompt to answer that you just copy over.
This essentially means any place it’s used manually. Therefore, where it can shine is when integrated with automation. Since Fibery has a fantastic knowledge network and good automations, this is where I can imagine it’d be most useful.
For manual inputs though, a way to also give some input to the generated output, like ask for corrections, improvements, add/remove aspects, etc. would be nice rather than a shoot-and-done static response.
- Generate text based on multiple variables (e.g. greeting card, like tweet shown posted by @Oshyan)
- Be used to populate fields other than just Rich Text
I haven’t worked much with AI, and I doubt Chat GPT-3 can do this, but a cool scenario would be the possibility to use your own data for training sets and let the AI have ability to work with Fibery. You could automate the task of creating references for instance.
An example would be Intercom chats with customers that links to Entities or so, or for bugs. You don’t have to know if an entry for X bug exist or not, the AI will try to link it in the text for you.
A much simpler example of this is what was shown in the Tweet; automatically set a single/multiple select based on a variable (e.g. Rich Text / Text Field / Formula result, etc.), possibly make new options in the dropdowns.
For example when new entities are created via Form or via API, say I have feature request Database with several avenues of entities being created, it could find existing feature Enity and reference it (for back-references) and/or increase a “# times referenced” field for popularity ranking—or create a new feature request that it from there keeps track of.
In our photography company we could potentially use it this way:
- after X amount of time after a order is created, trigger automation
- receive Y list of products bought as input, along with their name and email from our Customer database
- automatically create a new Entity with a personalized email that they can still buy images, and Z products go well with Y products
- one of our employees check for new Entity and looks over if it is okay, make necessary edits if any, and click button to send its current state (or mark OK for next scheduled batch of emails, if optimized day/time of day for emails to be read)
These are some interesting ideas and points. I think GPT-3/3.5 is already capable of a lot of this, but e.g. in the follow-up sales email example the trick is integrating it into a system like Fibery in such a way that it can be used fairly generically and broadly by many types of users. Making it part of automation makes a lot of sense there, of course, but the responses of GPT3 can be somewhat variable and there is currently no live automation output view or way to easily test, so it could be challenging to setup and troubleshoot such a workflow. But definitely intriguing!
I also had the perhaps wilder/less (currently) possible idea of being able to describe in natural language a workspace setup comprised of multiple interconnected databases, and having GPT translate that into a literal workspace setup. E.g.:
I want a CRM database that tracks the most common information about contacts, along with support interactions. The support activity should be tracked in a separate, connected database. Both databases should also connect to the issue tracker to attribute issues to the original reported user and prioritize issues based on their subscription level. The support database should provide context to the issues database.
Then the AI sees that 3 databases are mentioned and creates those, it looks at phrases like “the most common information about contacts” and references its broader internet-derived knowledge about common fields for contacts (it can already do this, try asking ChatGPT
"What are some common fields used in CRM systems?") and sets up fields, it interprets the “connected database” terms to be Relations and makes guesses about what kinds of connections, etc. It could be hard to make it do something totally useful, but it does have a lot of good examples out there across the internet already. And yes there are Templates that do something similar-ish, but this is definitely different and more customized out of the box. At the very least even if it was just a recombination of existing templates in the back-end, it could be a good marketing headline: “Natural language AI workspace configuration”.
The thing with all of this stuff is that the fundamental API connections and integrations are to some degree something you can build knowing that the AI models will improve as you do so, such that by the point you have a working integration, the back-end language model may already be capable of doing all this stuff and more. At the same time it could be difficult to build features and integrations on top of a “capability model” (via the API) that is constantly changing (though fortunately mostly growing/improving).
I’m quite curious to see what the team comes up with!
This is something we are thinking about as well. Fibery domain is a static JSON format, so it is relatively easy to train the network and produce relatively good results we hope. Here is an example of Comment in Fibery schema. All custom Types/Databases have same format for sure.
"fibery/short-description": "Communication made easy with comments",
At the early stage we did not change schema from UI, but did it from JSON like this to check how it behave
Wow, I did not expect to hear this, but it’s great!
Here’s another interesting, related capability to consider (and it ties into a lot of other big ideas Fibery is working on, “augmenting org intelligence”, and integrating chat):
Skip to the end if you need the TL;DR, which is essentially that one way to handle knowledge fragmentation is to AI-summarize changes and information, decisions, etc. from across tools and then make it available in a single place as “summaries” which link to their sources for details as-needed. Fibery already pulls in a lot of stuff, so it arguably already has a head-start there on such a feature, in some respects.
Yeah, I read this article some weeks ago. We do have ideas around it since Fibery fetches info from many systems. For example, we envision it will be possible for this use case:
- Fetch Intercom chats, Discourse threads, etc
- Ask questions like “What are the top 10 issues our customers are struggling with?”, “What our customers really like in general?” etc
Great! Given the “fluid” nature of AI content handling and content generation, and the largely externalized computation, I encourage you to get some tests into Experimental Mode ASAP so we can help you determine what is actually useful across more use-cases than just yours!
2 posts were split to a new topic: “Similar Entities” AI Action
Some inspiration for use cases → GPT for Sheets™ and Docs™ - Google Workspace Marketplace