I was inspired by the way ChatGPT writes bullet lists, then invites you to "delve" deeper.<p>This is an interface that reifies that rabbit-holing process into a tiling layout. The model is instructed to output hyperlink-prompts when it mentions something you might want to delve into.<p>Lots of features to add (sessions, sharing, navigation, highlight-to-delve, images, ...). Would love to hear other usecases and ideas!
This is very good. I can't put my finger on it, but it seems more important than a mere "gimmick." I noticed that if you click on a topic already explored, it won't open again. That's cool, I'd make it snap back to the pane where it's open.<p>Kudos! This is an interesting perspective on how we really need to put a little more effort into the UX of LLMs.
Taking a step back: The UX/UI for LLMs in general are very immature. We're in the very early days of to best interact with these tools. We need more experimentation like this to help figure out what works, and doesn't work.<p>Kudos!
If "delve" was meant to be an in-joke, I just wanted you to know: I got it.<p>I also have a Custom GPT "AutoExpert (Chat)" [0] that several reviewers have called "the perfect Rabbit Hole GPT" due to the way it leads users through learning a topic. You might dig it, especially since free tier users have access to these now.<p>[0]: <a href="https://chatgpt.com/g/g-LQHhJCXhW-autoexpert-chat" rel="nofollow">https://chatgpt.com/g/g-LQHhJCXhW-autoexpert-chat</a>
Love it! I like that the site is straight to action, but I think it could really benefit from a walkthrough. Here’s an idea!<p>It would be great if we had an introduction to the site right in the prompt! to help understand its main purpose right from the start.<p>It'll be great if the first thing you see is [Explain what "delve into" is] as a prompt suggestion. Next, it will reply with, "It’s for exploring topics deeply, similar to going down "rabbit holes" where one interesting thing leads to another. Here are some examples ..."<p>Then, you guide the user through the functions step-by-step. Something like, "Click on option X to start a new thread, then choose from the suggested prompts or create your own. Follow the flow to see related threads and dive even deeper."
I really like this.<p>How about enabling user to select any piece of text and use it the way the links work now?<p>For instance, I've noticed it doesn't linkify peoples' names, and one thing I love on Wikipedia is that you can easily lookup people mentioned in the text. So, rather than having thousand links in the page, it would be handy if I could just select the name in the text, click some button (or right-click menu item), and get the new prompt based on the selection (user having to click a few extra times for this custom use-case wouldn't be a problem IMO)
Great idea. I also see inspiration from Andy Matuschak's notes [0], of which I'm a bit fan<p>[0] <a href="https://notes.andymatuschak.org/zB74H9CuWrosEuqve7jZyCo?stackedNotes=z5E5QawiXCMbtNtupvxeoEX&stackedNotes=zDcuS8A5uxGR8hQygsqP83A" rel="nofollow">https://notes.andymatuschak.org/zB74H9CuWrosEuqve7jZyCo?stac...</a>
This is great and something that I've wished existed. Thanks for making it! Right now, the tiles are linear. E.g. if tile A links to tiles B and C, clicking either B or C will open a new tile directly next to A (and only one of B or C is visible at a time). What do you think about making more of a tree layout where B and C both branch from A and can be viewed simultaneously?
I like it a lot. Feels like idle Wikipedia link surfing but with the key difference that each new step keeps track of the previous context. To me it is both novel and useful.
Wow, this is really neat! I usually don't comment on Show HN's because I'm rarely impressed by them and I don't want my lack of enthusiasm to be a detterent for people showing their work, but occasionally one like this comes up that is very cool. I also really appreciate that absence of tracking other than Cloudflare Insights (which seems very reasonable to me).<p>There's an old truism in the business, that the more "suggestions" people give about your idea, the more they like it, and it's absolutely true. Solid work!<p>Do you have plans to monetize and/or open source it?
Branching conversations are great for a whole bunch of reasons. I posted a demo of a prototype:
<a href="https://x.com/ajdegol/status/1788689011302682657" rel="nofollow">https://x.com/ajdegol/status/1788689011302682657</a><p>And Jake Collins just announced he’s open sourcing an obsidian plugin which has a ton of features: <a href="https://x.com/JacobColling/status/1795462258258002255" rel="nofollow">https://x.com/JacobColling/status/1795462258258002255</a>
Perhaps have it not scroll down as it generates the text? Invariably I have to scroll back to the top to start reading. You could have a mini-hud (growing line, with a small rectangle at the top showing the first page of text) which would let you see at a glance how much text is being generated, without interrupting reading. Or not; ChatGPT just keeps on vibrating the phone (iOS app) during text gen, with hovering arrow in the middle-bottom as a shortcut to jump to the end.
Reminds me of Andy Matuschak's UI as well: <a href="https://notes.andymatuschak.org/z9C7piFz8mthkGUUd1W2CR?stackedNotes=zExJhXA1doKnxyQQ4ZjX4dV" rel="nofollow">https://notes.andymatuschak.org/z9C7piFz8mthkGUUd1W2CR?stack...</a>
This is pretty great, provides a nice set of breadcrumbs for a deep dive into any rabbit hole.<p>Once thing that threw me off was when I went to the original panel and clicked a second topic, it cleared out the panes that I had explored off the first topic. I had to discover they weren't really lost by re-clicking. I think it would be better if there was some visual indicator they were still there - perhaps the topic (and sub-topics) get collapsed but are still visible with the heading of the selected topic?
For some reason I feel like the Wikipedia should use this (or something like this) as a backend to serve a more "dialogish" UI (without replacing the current static UI). The name "delve" is spot-on, but it lacks Wikipedia's intelligence and interconnectedness.<p>And maybe add some locally stored "Microsoft Recall"-like feature to revisit paths you've made. It would be text-only, so use up almost no space, and be quickly searchable.<p>Well done it could even work in a terminal.
I can't believe no one has mentioned Andy Matuschak's work on his notes, "Andy-mode:" <a href="https://twitter.com/andy_matuschak/status/1568032773025431552?lang=en" rel="nofollow">https://twitter.com/andy_matuschak/status/156803277302543155...</a><p>It's a really clever use of the UI, and I think he'd be happy with it.
I really love this. A book was recommended to me that I'm not going to have time to read, but this UI is an amazing way to figure out the main ideas and dive deeper into the interesting ones.<p>No idea if the things it's telling me are true or not, but that doesn't matter quite as much in this case.
You may find interesting to look at Google's AI Kitchen and the early versions of Bard (the LaMDA version) because they were specifically optimizing user scenarios where the user wants bullet points (though intuitively, your tool seems already much better than what Google did).
This UI wrapper looks fantastic! Great job on making the GPT API more accessible. Are there any plans to roll it out as a UI component that can integrate with existing systems, or does it require some backend work too? I see huge potential for clients like GPT4All and Prompta to use pluggable UI components like this, allowing people to keep their data locally.<p>I believe we might see AI clients evolving into open-source editors like Atom or VS Code, with plugins and packages that can be shared and iterated upon rapidly.<p>Also, I received the message 'You've gone too deep!' and I'm wondering what the context length limit is. Thanks!
Really fun! I realized each delve carries the context of the previous ones. So I got to StarCraft II from the initial example of "Faster language models", but it mostly talked about how SC2 can be used for reinforcement learning. It'd be nice to have a key I can hold down to start a new delve on the topic (bonus points if you can stack multiple delves so you can keep going deeper on the old track as well!)<p>Another thing that would be interesting is if there was minimal markup for the LLM to indicate "here should be an image of [search term]" or maybe even interactive code blocks etc. But obviously this is scope creep deluxe.
This is really cool! I love the rabbit hole stuff you can do when you give GPT more capabilities. I was playing around with this stuff and found I was most often wanting to use it when wanting to learn about something so made Instaclass: <a href="https://myinstaclass.com/">https://myinstaclass.com/</a>. It finds videos, images, makes quizzes and gets more relevant web links for you to keep exploring, and structures it like a class (basically a list of bullet points like you mentioned). Try it out and lmk what you think!
I'm not sure how to use this, I just see a blank screen with the word "delve" at the top and typing doesn't do anything. I'm on Firefox on macOS.
Well this is incredible. Love the interface and the speed. My only wish is that it gave more links, right now it only seems to generate a few of them for each response.<p>Also it seems to be way faster than ChatGPT, yet just as intelligent, how is that possible? Can you elaborate a little bit on the architecture you're using and how it all works? What model are you using? Is it just a straight up API to ChatGPT, or are you also using additional embeddings and fine-tuning?
I've been craving a gpt ui where I could fork conversations that stem from a genesis thought.<p>Put this on a canvas too so I can zoom out and look at the footprint I left to retrace my steps
crrriispy ui dude. This is like that custom prompt people were sharing for gpt where it would preempt three relevant follow up questions.<p>what heuristic are you using for making words clickable?<p>i recommend making the links just hardcode old-school blue and purple. make it obvious you can click these things. "dive on in" <a href="https://www.youtube.com/watch?v=DElxVXS7PD0" rel="nofollow">https://www.youtube.com/watch?v=DElxVXS7PD0</a>
Interesting concept, but in practice the experience is just a simplified Wikipedia hole. I enjoy all the hyperlinks in Wikipedia articles: they let me decide how deep I want to go rather than get a short summary with only a few pre-generated links.<p>This adds a lot more barriers to knowledge exploration (and the GPT is likely trained on Wikipedia anyways) and doesn't provide sources.
awesome work! I've wanted to explore the same idea - glad to see it getting worked on. The chat interface into language models clearly works but it frequently feels like an inefficient way to explore the latent knowledge space of the model. Hypertext (the www) has also been shown to be a great way to explore a massive knowledge space. What this is doing is applying something like a hypertext layer as a way to navigate the model's latent space. Very cool. It could become something of a dynamically generated personalized wikipedia. I'm curious what the prompts look like that you are using to generate subsequent "pages". It could be as simple as "write a wikipedia style summary of <x>" but I think there is a lot of potential in including the context of my current "rabbit hole": "explain <x> in the context of <y> with a learning goal of <z>", etc. Another idea: grounding this kind of hypertext exploration with rag on a specific dataset, e.g., wikipedia or hn.
This is pretty neat.
What I really like is the tiling layout.<p>I subscribe to phind, which provides a nice search/answer service, which also suggests followup questions, which works fairly well:
<a href="https://imgur.com/a/WfHSzdk" rel="nofollow">https://imgur.com/a/WfHSzdk</a><p>But if it was in a tiling format, that would be pretty awesome for the flow, especially on mobile.
Nice work, this is really solid!<p>I've had something like this on my mind for a while. I really think there are some great use cases for AI around supporting/enhancing human cognition rather than trying to outsource our thinking. In this case of this, being able to rapidly "expand" your working memory with whatever is present in these cards is promising.<p>I look forward to seeing what you do with this.
This is fantastic, like an encyclopedia that knows what context you are learning about as you skip pages. Nice work!<p>My minor recommendation is to highlight somehow that the input fields accepts any topic and the suggestions are just random try-it-out topics, it's wasn't immediately clear. Instead of 'write a message' it could say maybe, 'enter a topic to learn about'.
This reminds me of what <i>Perplexity . AI</i> 's interface was initially like, 2023Q1 (when the hyperlinks actually listed at the bottom of each reponse, instead of just as unlisted superscripts).<p>Nice clean interface, OC, <i>please keep it this way</i>. Thanks for sharing.
I used to go down rabbit holes on Wikipedia all the time--could spend hours doing this.<p>This to me seems like an Infinite Wikipedia! Really cool use-case!
I love the left to right view. Super cool. I got stuck thought like 5 levels deep.<p>I really like this for source level information. Like drill down on research studies and then drill down to authors and concepts. Would be cool if it was also building a mind map or semantic tree that you could see. Like how to separate the topics by their level of generalization.
This is so cool! I can see myself using this all the time for research work. Would love to manage multiple sessions. Would also like to see if the individual tiles could be resized. I would like to decrease the size of slightly less important tiles perhaps. Also, how are you managing the API costs right now?<p>Great work!
Follow up questions generated with the answer is easy picking. Much like Google (traditional search) does. To spice it up, there can be a slider using which - questions would change. The axis of slider itself can be an interesting one eg: simple to advanced. It will depend on the topic being discussed.
This is cool, it's almost like a wiki you can talk to. I also wanted to make a thread-based UI for LLM chat since I realized that's how I typically interact with them (almost like git branches) but hadn't gotten around to it yet. Neat to see others are interested in branching conversations as well!
<a href="https://delve.a9.io/" rel="nofollow">https://delve.a9.io/</a> no longer give answer.
Please give me code/self-host frontend, or let me input my api key
Very cool!<p>It seems to dead-end unexpectedly on some topics. For example, I delved "path tracing" -> "importance sampling" and it output a section on probability distribution functions (PDFs) but didn't offer any links to explore those further.<p>Highlight-to-delve would probably fix that.
It would be great to implement a browser extension that lets you highlight a term or phrase on any webpage and open a GPT rabbit hole for that term or phrase.<p>@maxkreiger - if I were to build one as a proof of concept, would you object to me having it hyperlink to your UI?
Since the page did not provide any information on the language model used, I asked it and it said it was Anthropic. An About page would go a long way for this project.
I see there are many comments about Wikipedia. I can't find the link, but many years ago (like 7+), there was a concept Wikipedia redesign that proposed the same UX. This was a "marketing" project of a non-Wikipedia related designer. Does anybody remember?
This is great, I love it.
I'd love to have it be keyboard only. Maybe using a command syntax like "/link title", so it knows we don't want to continue the chat in the same panel, but want to follow the keywords.
You've got something great here!
Reminds me of the Smalltalk IDE and browser.<p>Also, WorldSim could use something like this, and perhaps web browsers.<p>Anyway, I suspect this resonates for anyone that has to do research on the web or in GPT. I often end up with multiple threads on GPT anyway trying to learn about something.
This reminds me of Andy Matuschak's notes website [0]. I dig this view for much better understanding.<p>[0]: <a href="https://notes.andymatuschak.org/" rel="nofollow">https://notes.andymatuschak.org/</a>
I dig this. Reminds me of column mode in MacOS's finder, which is similarly helpful in "rewinding" an exploration of a file system.<p>Would be interesting to rabbit other rabbit hole resources like Wikipedia or IMDB in this way too.
Neat! I made a free browser extension with a somewhat similar behaviour (<a href="https://hoverflow.io" rel="nofollow">https://hoverflow.io</a>) Inspired by nested tooltips in Crusader Kings 3.
For coding it would be nice to add task to links. So when a link is clicked you could simply choose to follow it or create a new agent with the link. Each agent will tune the output as one goes down the rabbit hole.
I like it! Would like to have some sort of map/visualization of what topics i've explored for when diving into parallel sessions. Right now the UI only supports diving into one branch at the time.
There should be something a UI like this but it just sources data from Wikipedia. Could just be the Abstract and allow you to open a wikipedia iframe or something. You should add this as a mode to this app
This is cool. I'd love an option to make the output elaborate more, if I'm interested in a topic it's a bit disappointing that I can only read a few sentences. Other than that, nice work!
First impression, a fast and neat interface. I went into data indexing rabbit hole as that has been my obsession these days.<p>Cudos! Particularly impressed with the lack of clutter and the speed.
I REALLY like how snappy it is. I've always been impressed with how fast Wikipedia managed to stay over the years, but this is even better. Really nice work.
Super cool!<p>What’s the logic used to determine what words are links? I found that it was possible to feed it new vocabulary and that it would turn those into links
some feedback:<p>* can you make it so we can share links of sessions?<p>* can you describe on the homepage or in a link from the homepage what it does.
great work! We are also trying to design a new different way to use LLM: <a href="https://flowith.io" rel="nofollow">https://flowith.io</a> you can generate stuff on an infinite canvas.