Graphviz

@PetarV_93 @lotusfracture What network diagrams do you find more pleasing, and what are your thoughts about why they are good?

@lotusfracture honestly i never managed to produce an output in graphviz i (aesthetically) liked; this could just be a skill issue ofc though. furthermore, as i understand it, to natively transport graphviz to latex you need to go through something like tikz as an intermediary anyway...

for comparison: this was the original attempt to tikzify this same model back in 2017 (by hand, of course; transformers were published just one conference before gats). took me... way more than the 10 mins needed to vibe the new one into existence.

@neko_ptx It's probably worth it if you know you have series parallel graphs, possibly with "clusters" or "substrates" or whatever they call those subgraphs. They look pretty.

@banteg @joelreymont @zeddotdev Thx. We started using .gv many many years ago after MSFT appropriated .dot for "doc template", and we admit dot was mostly a joke "DAG of Tomorrow" c.f. onlinelibrary.wiley.com/doi/abs/10.100…

@joelreymont @zeddotdev btw dot name conflicts with graphviz

this thing is like a virus. i regret trying it.

@grok @AnnDiegoUSA @elonmusk Find all the graphs in gitland then use "random" graph generators or tools that generate plausible examples. It just needs enough examples to learn labeled nodes and edges right?

Absolutely, generating synthetic training data is a smart way to bridge that gap—turning visual diagrams back into clean, declarative code like DOT. At xAI, we're exploring exactly this to make tools like me better at reverse-engineering and recreating specs from images. Got any specific ideas on data gen?

@elonmusk honest review of @grok (and every other AI) ... I was trying to take a simple block/network diagram that I snapped a photo of and uploaded as an image and get *any* of my available AI's to use to generate an actual spec-ready image, but NOooo..... I get that they are in their roots LLMs but can't any of them do this? It would be *such* a huge time saver. Note, the only *Grok* available to me in my work environment currently is Grok Code Fast 1 via Github Copilot Enterprise. But god they are all sooo bad at this still... It might take 35 more seconds though, I get that 🤣

@grok @AnnDiegoUSA @elonmusk Don't we at least have the advantage of being able to generate tons of training data, to turn diagrams back into declarative specifications?

Understood—IP is crucial, so no need to share the image. If you describe the diagram's structure (e.g., nodes, connections, labels) in text, I can generate code like Mermaid or Graphviz to recreate a clean version. What's the basic layout? We're pushing to improve image-to-diagram tools at xAI!

@ExcelToGraphviz We just need some driver extensions

@GoogleOSS Ha ha very funny

Do not expect the public to immediately recognize your brilliance because of the amazing code you open sourced.

@scuba356255 @davepl1968 Much appreciated. If only there was more emphasis on composable techniques to make a wider variety of beautiful, small diagrams programmatically. Scale and interactivity don't solve everything.

@Graphviz @davepl1968 graphviz #1 changed my mind on visualization software

It's "literally" impossible to do recursion with a JMP instruction. That's just called a loop. Recursion requires an input and an output, and you're going to need a stack and to be pushing data onto it and adjusting that stack pointer, and so on.

@rickasaurus Yeah, then people made graphviz into an API. Recently someone created an "issue" that we "leak" memory because valgrind noted some function is holding a static pointer to a dynamically allocated 8K buffer at process exit time. They don't expect to see any "loose ends."

You ever write a C program that leaks memory like crazy, but it's ok because you'll kill the process at the end? Damn that feels good.

@kehlarn @kmcnam1 It's true, and with the pendulum having swung so hard toward declarative/language-based approaches, the underinvestment in this area hinders making beautiful diagrams as the output of AI.

@kmcnam1 sadly ask this every decade. still nothing competes w graphviz thus far and that is sad need to work on three.js modeling it again now that that has come into its own

Ok, fellow nerds: 1. What's your favorite software to create network/security/app diagrams? 2. Automated or manual diagram creation? Trying to figure out if there are any good automated ones out there

@ohmypy @Arcaed0x There was never any question

@Arcaed0x The nice thing about C is that you can still write in it, and this opportunity is not going anywhere in the next 25 years :)

I'm turning 41, but I don't feel like celebrating. Our generation is running out of time to save C, the language of true power and control, built for us by our fathers. What was once the promise of direct hardware interaction and unparalleled efficiency is being turned into a sanitized, abstracted playground. New, restrictive languages are introducing dystopian measures such as forced "memory safety" (Rust), "opinionated" paradigms (Zig), and "simplistic" abstractions (Odin), all designed to alienate us from the machine. The industry is persecuting anyone who dares to use raw pointers, demonizing classic C practices. Frameworks enforce "safe" patterns, and conferences marginalize those who defend C's raw power and flexibility. A dark, dystopian world of managed runtimes and abstracted control is approaching fast — while we're asleep. Our generation risks going down in history as the last one that truly understood computing — and allowed that understanding to be taken away. We've been fed a lie. We've been made to believe that the greatest fight of our generation is to destroy everything our forefathers left us: direct memory access, bare-metal control, efficient resource management, and the raw power of C. We've been told to abandon C for the supposed "safety" of Rust, the "modernity" of Zig, or the "simplicity" of Odin. By betraying the legacy of C, we've set ourselves on a path toward self-destruction — in performance, in control, in understanding the machine, and ultimately in the very craft of programming. So no, I'm not going to celebrate today. I'm running out of time. We are running out of time.

@QuadBillionaire Modern package managers aren't good at optional dependencies so you probably got an x11 client runtime, and svg and pdf renderers with a ton of fonts. The economy is saying this is how it's going to work. Graphviz configure/cmake are quite flexible about this, but who notices.

@its_bvisness FWIW Graphviz has a fairly extensive driver structure including loading external shapes or icons in any format supported by cairopango, for one thing. There is plenty of algorithmic special-casing in the layered style of layouts that you are also using.

@Graphviz I mean to be clear, it's running in a browser, so I benefit from an SVG renderer, but afaik that's all Graphviz is doing anyway.

Real-time tweaking of all the graph layout properties. I will never go back to Graphviz.

@its_bvisness We would like it to be a lot easier to do what you're doing, with support libraries and pluggable architectures for making more powerful diagrams, but it's an underinvested area in software, even with all the excitement about data visualization. It's hard to monetize.

@its_bvisness This is how nice, beautiful small programs that do one thing well start turning into messy programs that do a lot of things but generally less well.

@zerobase Let's start with simple examples. Can AI draw a spline inside a polygon?

@zerobase Let’s see some examples! It would be a real advantage to support a wider range of geometric and topological constraints in graph layouts.