If I can plug my own API key into this and/or run Llama locally, that'd be great.
It reminds me of a tool I saw recently called Heuristica [0]. Would like to try it but I don't like being tied to a subscription and the free plan seems quite limited if I can't even plug my own key in. Don't see why this can't do what Heuristica does! :)
Hey there! Thanks for mentioning Heuristica. I would love to find out how to make the free plan for Heuristica to be more permissive (without destroying the incentive to subscribe for willing users). Feel free to send me your suggestions.
At one point, I also worked on making it work with a personal API key. However, this added a lot of complexity. It felt like I was building and maintaining two separate branches of the same app, so I had to put the idea on hold. I might revisit it in the future.
Hey! To be honest, I am not sure my feedback would be very valuable. I'm probably your worst type of user -- perma free plan -- simply because I'd only use your tool sporadically. I don't do literature reviews that often, but often enough to think about using a tool like this I suppose.
I totally get it though, it's a difficult thing to balance. If I was doing lit review and deep research daily 6.99 is an amazing deal.
I ended up at Tldraw's London office a few weeks ago for a thing, and I remember afterwards being like 'ahh, now I understand how they end up just casually doing random cool shit and attracting the kind of talent they do'.
They should be extremely proud of the culture they've managed to foster and I genuinely hope to see them succeed as a business.
Very much this! I was also at a thing at their office a few weeks ago (some thing? "Local Thirst"), and Steve gave a demo of this. It is incredible.
I've joked before that the last generation of human machine interfaces ware invented at Xerox park, and the next generation is being invented at TLDraw of Finsbury Park. But it's not really a joke, I genuinely believe it.
I agree. Looking at this, it seems to be exactly how I want to use LLMs. Describe a small transformation of data I don't want to work out now, connect it to other components. As the needs become more-defined, replace each part with a faster, more-reliable, well-defined data transformation. I could actually see developing a system this way...
Ha yeah, that was the same thing! The night it rained sideways.
So this is the demo people were talking about at the end of the night! I was quite annoyed I missed it, makes sense now. I think I was nerding out over current-gen HIDs while eyeing up their very tastefully equipped coffee station (ozone roasters ftw)
I got to see this demo'd at a conference in Sydney recently, and it's really cool.
It's not super serious, but it's not meant to be -- it's not pitching to be your enterprise AI strategy. However, even though it's presented in a playful way, I suspect it's quite powerful, and expect The Internets will build some cool stuff atop it.
It's a fun and creative way to explore playing with LLM's, and it's brilliantly executed! Happy to see it here on HN.
I want to use Tldraw as a simpler alternative to Figma. I want to drag and drop Web Components (or React components) into the canvas to play around with different UI ideas. Maybe a built in library of Shadcn components I could mock up an UI with.
I'd like to echo the impressiveness of tldraw. At the BigBlueButton project, an open source virtual classroom, we built tldraw into the core. It has saved us countless development hours as we stopped trying to build our own whiteboard and instead stood on tldraw's (very) wide shoulders. We've never looked back.
After reading the blog post, I don’t really understand what this is. The blog post seems to go out of its way to only make allusions and not explain what this actually is and does.
I'd love a "Code" component where you could enter arbitrary code. After following the tutorials I asked myself "What would I like to make?" And I imagined a tweet-bot - grab headlines from Wiki news (or somewhere), combine with an instruction to generate text and another couple instructions to generate an image, post to twitter (or bluesky).
This seems easy enough if I have a code component that could execute arbitrary code. I could just write a couple small component (take API key, text, post to twitter/search wikinews) and add them to the workflow. If the components I needed were generalizable I could share them on some kind of community repository - so the next person who needed a "Post to twitter" component wouldn't even need to rewrite it.
Looks like an interesting project for sure, but they make it really hard to try out. At first, this submission links to a blog post with information and in the middle, a bunch of demos laid out in a grid.
I clicked on one of them, which took me to the actual app. When trying to click the button to generate something, it asks me to login/sigup. Fine, I signup. Then once finishing the sign up, I land on some sort of index page with tutorials and "my projects", but not at the demo I first wanted to see.
So I go back to HN, to this submission, click on the link but it still takes me to this index view instead of the blog post/page I first read. I just wanted to see what happened when I clicked the button inside the app...
That redirect is a bug for sure, thanks for pointing it out. We have all those examples on the index page for logged in users but I can see how it would seem like they’d disappeared.
I was thinking of developing something similar, but it ended up being one of the thousands of ideas that never end with a line of code. I'm glad to see it here.
Visual programming is a tempting idea I love. It rarely works, but this might be the case.
I think there is a lot of room for AI UIs - between chats (the simplest and most prevalent) and arbitrary code (even if it is "just API calls", it is only people with at least some software inclination).
One thing I am keeping track of is WordWare (https://www.wordware.ai/), which makes it easy to create a sequence of operations. It feels like an "Excel formulas of AI".
Yet, I like the visual, graph-based approach of Tldraw.
users create workflows from blocks of text, images, and instructions. When run, information flows from one component to the next, with the output of each generation serving as the input to the next, creating powerful processes that branch, loop, and iterate to produce outputs
Even "vanilla" tldraw is super cool as a clean, functional, open-source html5 whiteboard, and the team have absolutely been killing it in their comms and use of LLMs. I honestly think they might be some of the most innovative people around when it comes to really novel UI for LLMs. Also, Todepond is just very cool.
does the cloud product’s “new project” button still trash your saved documents with one click behind a docstring something like “make sure you have saved your stuff before making a new project” where what they meant is “our cloud product does not save your projects to the cloud, it is in local storage actually and you can only have one project at a time so the new project button actually overwrites your old one, so when we say ‘save’ we actually mean export your stuff to a json file and save to local disk!! so you can re-import it back into the product later from local disk and overwrite it back!!!!” I did my VC seed pitch deck in tldraw along with a bunch of product mocks, ask me how i know this
So sorry Dustin. We'll have a new version of tldraw with user accounts in a few weeks that should improve things, but until then please no one clear your browser storage
In tldraw the text is always in the DOM, so the browser can use your native font rendering (for example ClearType on Windows). In Excalidraw the text is only in the DOM while you're editing it.
I didn't know Todepond worked on tldraw. That's cool.
> really novel UI for LLMs
Are you referring to Tldraw Computer or something else? Don't get me wrong, it looks really nice but not that different from other graph representations of LLM workflows, including live updates in the nodes themselves.
Do you somewhere have slides/recordings from the awesome AI tinkerers talk you did? Because that's what I had in mind when I made this comment and will be way easier than me trying to describe it
I'm curious how the internal prompting works in certain cases, and whether there's any way to customize a particular module's default or hidden prompt. Particularly with speech. I was trying to get it to sing a made-up Christmas carol, with generate lyrics and chords. I tried a bunch of different ways, but at best the speech module would only read it out. In one funny case, the speech module added on its own beforehand: "a spoken-word piece".
We had a bunch of fun putting this together so I'm really happy to see folks enjoying it. I'm not sure where the project is going but I've been waking up for weeks with a fresh "oh christ, we could do ___", so that's exciting.
Not immediately! This might turn out to be just a great demo, might be something worth continuing with, really depends on how the next few weeks go. Either way there might be something we can do with the developer community around data endpoints in the short term.
so exciting to see these ideas out in the world! i'm now imagining a Scratch-like playground for kids to explore end-user programming / AI in an accessible way, like some of the example apps you've shown
there's been a rising tide of academic HCI work in a similar space, wonder if there will be cross-pollination of ideas along these lines (many more papers i'm sure but some off the top of my head):
https://arxiv.org/abs/2305.11473https://arxiv.org/abs/2309.09128
I should really share the design doc from when we started putting this together. The core idea here was a) making a little hobby machine on the canvas and b) using LLMs to “power” the widgets. Somehow that ended up looping back to “ai workflows” but definitely not where we started. I’m only catching up on the space now!
In excalidraw I just have to click share session and anyone with that url can see my whiteboard and interact with it. I get tldraw has much more features etc. but how exactly is it making sharing whiteboard so easier compared to excalidraw?
I don't know maybe it was a skill issue from my side 2 -4 months ago , I felt as if I was forced to sign up back then
I am sorry I guess then for this comment , excalidraw also works great but I still just like tldraw because of how familiar I have become of this interface.
Shame that the licensing of tldraw is less permittive than excalidraw but I guess I am just a little bit okay with it considering its still open source and though I maybe wrong I had read the license , and it seems that it was focusing way more on that you had to have the name of tldraw / packaging of tldraw / copyright
here is the license restrictions
Not to disable, hide, remove, or alter the Watermark.
Not to disable, change, or interfere with the license key validation process that governs the display of the Watermark.
Not to remove any copyright or other notices from the Software.
Not to make the Software available under a license that supersedes or negates the effect of this License.
Not to distribute the Software or modifications of the Software as a standalone product, but only as part of another application.
To include a verbatim copy of this License in any distribution of the Software.
To comply with tldraw's trademark policy.
This is basically ComfyUI but for LLMs, is that right? I know tldraw as the open source Excalidraw competitor but this is an interesting product as well.
Excalidraw is more open source than we are at tldraw! We're both source available on GitHub but Excalidraw is MIT while the newer versions of tldraw are a kind of watermark-ware. (We still have an older MIT version available but not in development)
TLDraw relicensed about a year ago. It is under a permissive license, but no longer strictly open source ('watermark-ware").
https://tldraw.dev/ FAQ:
Is the tldraw SDK open source?
Our license is not exactly Open Source but you can view the source code on GitHub. We accept contributions from the community and work in public.
I'd like to know if I can use the SDK to build workflow/process diagrams that specify inputs, outputs, and side effects (ie, this process creates a pile of logs or documentation) and then export a process specification for use in another application.
My specific use case is process mapping and quality systems implementation in a hardware engineering setting.
That sounds really useful. There's no export yet here apart from images and it isn't something I've thought about much so far. Are there standard formats for these types of workflows?
No industry standard format that I know of. I presume I'll have to come up with my own (some kind of typed JSON, probably).
At the end of the line these are just function definitions -- a black box that takes well defined inputs and produces well defined outputs, as well as calling out side-effects (I suppose these could just be more outputs).
Anyone else experiencing projects reverting back to their default state? I'm not sure if this is just after publishing... two of the cool things I did yesterday (including one linked here) were wiped out today and simply show the default new project.
This has been the best UI for my kids to get excited about AI and simple visual programming, they're having tonnes of fun playing around and creating cool image workflows with it.
So cool to hear how many folks are using this with their kids! My own 6yo got it right away when I was prototyping this and that was huge validation. (She quickly moved on to drawing the whole graph herself, combining images etc)
Question: Is there any way to induce randomness? Given two lists, instructions that say "Pick a random item from each list" consistently returns the same two items.
Seems like the same text/instructions are cached, so they give you the same result. Just a hack, but adding in a data box with current date+time fixes it as expected as the inputs are no longer the same. I wouldn't rely on this for true randomness though.
It's interesting how sometimes the instructions seem to specifically dictate JSON output and other times not. Even without changing the prompt, it seems that this aspect - the generated series of steps for each instruction - is pretty random each time you run it. Or maybe it caches for awhile. What would be really nice would be if there were a checkbox to request structured or unstructured output from instructions, or better yet just lock in a set of derived steps you were happy with.
Interesting product and obviously awesome execution, as expected from tldraw… but yeah… seems like a very strange departure from what Steve has been building the past few years.
The core product / pitch is still the same—an SDK for whiteboards and other infinite canvas stuff—and that's what we monetize through licenses. Computer (and our other demos) are basically marketing, R&D, and fun.
Well if you're looking for fun stuff... could you make a tool that lets me easily breadboard [1] an app, and then you GenAI it into a low-fidelity clickable prototype?
Actually in its present form. It would need to be supervised to help teach kids programming.
I just spent the last few hours typing in the specification of a Freshman programming project that I use to teach, Rock, Paper, Scissors.
Specifying the programming assignment using either Python, C or C++, using their parameterization feature. Parameterization is very cool by the way!
It would routinely miss putting any headers in for the C or C++ examples. Once in awhile it would generate actual working code in C and C++. But it was all unindented. The same for Python also. The Python code had no indenting everytime causing indentation errors for the Python environment. I tried many different ways of specifying indentation for the text output; it didn't work. Possibly b/c of the HTML output being generated. It's 90% there, but the user would actually have to have some knowledge of the programming language to make sense of the errors.
Tldraw computer – how does it work? https://www.youtube.com/watch?v=Kn1De5uwrlYSteve from tldraw gives a tour of how tldraw.computer handles AI generation on the canvas, including a peek at prompts and models.
TLDraw is fascinating, but I feel like when I see them tweet cool stuff it isn’t actually in the app. This
is likely me being dumb but since it happened a couple times now whenever I see their posts I assume the same.
Yeah we did a lot of work this summer that was really far out, maybe too far out, and didn't come together as a product. We shipped a lot of it in teach.tldraw.com though. Computer is exciting to me in part because it feels both very weird and also intelligible as a piece of software.
I think its decent considering it requires money , and even chatgpt in its early stages didn't allow for anonymous chats / unlimited chats and I remember going on all these chatgpt clones becuase of that.
I also understand the hilarious spin that you added considering tldr (too long didn't read) lmao. but still its worth your email.
Crazy how I realised that tldr meme after I had written the first paragraph
believe it or not I picked the tldraw name because I already owned the domain (I'd bought it for a different project called telestrator) and it was only weeks later when Francois Laberge complimented the clever name that I noticed the portmanteau
I'd appreciate if you didn't consider Firefox Relay emails as disposable email. The Firefox folks specifically have tried to make Relay anti-abuse.
It's an unkind thing to do to your prospective users.
From Bleeping computer's coverage the last time someone tried to dump Relay in with a disposable email blocklist:
> Back in November 2021, Firefox Relay's team lead had requested the maintainer of a separate burner email list, "burner-email-providers" to exempt the particular domain form the blocklist:
> "We are operating Relay with a number of features that I think mitigate the risks that these aliases pose," Mozilla's privacy and security engineer Luke Crouch explained in November.
> Firstly, if a @mozmail.com alias is disabled by the user, any emails sent to the alias are not bounced back but instead discarded with a 404 error message returned by the service's HTTP webook, stated Crouch.
Secondly, he explained, the anti-abuse protections built into Relay limit free users to a total of five aliases, and further rate-limit premium customers so they cannot abuse the service by creating large-scale throw-away aliases for, say, automated signups to web services.
> With that reasoning, mozmail.com was swiftly removed from that blocklist. And it appears, the creators of "disposable-email-domains" have also honored the clause, for now.
Oh sorry, that's a toggle in Clerk (our auth provider), it doesn't provide granularity around which are disposable and which aren't. I'll take a look and see whether there's anything I can do short of turning off that feature.
It reminds me of a tool I saw recently called Heuristica [0]. Would like to try it but I don't like being tied to a subscription and the free plan seems quite limited if I can't even plug my own key in. Don't see why this can't do what Heuristica does! :)
[0] https://www.heuristi.ca/
reply