Hacker News Reader

Comment by catapart | Hacker News Reader

Sorry if I was unclear: I just meant that we could employ tensor-based learning methodology, like ChatGPT uses, to parse the graph representations and understand not only their visual layout (ex: no lines crossing), but also their contexts (ex: group functions that all emit events near each other on the graph). Having a machine be able to know what you're trying to do and then clean it up to based on that would be great. Even better, if it could learn how you tend to organize things and pick up "configurations" from that. Adding in "reroute nodes" and stuff when things start looking too loose for your personal liking, or whatever.

View on HN

I would also love to have an auto layout that considers more than just uncrossing lines. I spent some time trying to find a solution for organizing Nuke node graphs (for VFX) that would follow a few additional rules, but I didn't get very far. Not a perfect example, but getting something along these lines would be ideal: https://headjack.io/wp-content/uploads/2017/07/Huge_script-1...

This doesn't make any sense. There are already lots of graph UIs that can auto layout. Why would you need 'tensor-based learning' to uncross lines?

> There are already lots of graph UIs that can auto layout.

Do you have any references besides Graphviz? I haven't seen much that does anything interesting in terms of weighted or constraint-based layouts.

It's in a lot of content creation tools already.

https://saturncloud.io/blog/what-is-a-graph-autolayout-algor...

https://github.com/microsoft/automatic-graph-layout

https://github.com/kevtiq/digl

https://www.mindfusion.eu/graph-layout.html

https://reactflow.dev/docs/examples/layout/auto-layout/

What do you think 'tensor-based learning' or 'training data' would do for graph layout?

Those all look like the same, standard layout algorithms that don't seem to take into account any particular relationship between a group of nodes. Almost every graph layout algorithm I've seen just treats the nodes as individual things to be connected rather than additionally adding some contextual grouping context to various subsets of the nodes, which I think is what is really required for interactive graphs.

> What do you think 'tensor-based learning' or 'training data' would do for graph layout?

I didn't and don't have a comment on that.

Not to tell you how to comment, but those that don't seek answers with an open mind will often fulfill their own disillusion without your help at all. Which is to say: if the person you're responding to can't intrinsically understand what would be good about having a 'pair programmer AI' cleaning up your graph as you code it, based on things it learned about how you code, then that person is probably not going to benefit from whatever you can explain to them. If they don't have the problem, then their understanding of it will be just as academic as their naysaying of it. And, all things being equal, why bother trying to make someone understand something they've already represented a lack of utility for?

Not to tell you how to comment, but those who don't seek answers will not realize their dreams. Their understanding will be an academic projection of the future and they will guard themselves from truth. If people don't intrinsically understand you, maybe you need to go in to more detail so your explanation isn't so abstract.

Ergo, vis-a-vi have you ever actually tried making a graph that is rearranging itself while you're using it?

Also what is the "AI" doing that can't be done with a normal algorithm? Positioning nodes in a graph would just be determining their x and y positions. How much time have you spent working with node graphs? Arranging the nodes is easier than arranging your keyboard and mouse, it's trivial.

> If people don't intrinsically understand you, maybe you need to go in to more detail so your explanation isn't so abstract.

People did; you didn't.

> Ergo, vis-a-vi have you ever actually tried making a graph that is rearranging itself while you're using it?

No one said anything about "while you're using it". Linters don't fuck up your text as you type it. It's a utility like anything else. And, YES, I have built a working version of one. It doesn't use machine learning/tensors at all, and it's a fucking nightmare to develop. I can see exactly where those things would help me. It's too abstract, to you, to understand? That's a shame! But other people seem to get it, so if you are having trouble, maybe try approaching your ignorance with openness and curiosity instead of cynicism and trivialization ('uncross some lines').

If you want to know how much time I've spent working on node edge graphs, you wouldn't believe me. But I can at least point you to the start of the web-component version I'm porting: https://github.com/catapart/Magnit.NodeGraph It's probably two years out of date at this point, as I've been integrating some of the algorithms for the specific need I have for these graphs (interactive dialog), but even from that anemic demo you can see that I know how to build the damn things. Again, you're trivializing for...some reason? I don't know. You seem to think you know more about me on this, and that's fine, but you don't really ever demonstrate that so I can't see your commentary as anything other than intransigent ignorance. Which was obvious from your first response, which is why this is my first (and last) direct response to you.

I hope you feel better about whatever it is that is causing you to have this reaction to a pretty obvious benefit that you, for whatever reason, seem too stubborn to engage with.

No one said anything about "while you're using it".

Your comment: cleaning up your graph as you code it,

I have built a working version of one. It doesn't use machine learning/tensors at all

Your comment: we could employ tensor-based learning methodology, like ChatGPT uses

How would that work on a technical level?

You keep throwing around insults like:

those that don't seek answers with an open mind will often fulfill their own disillusion

if the person you're responding to can't intrinsically understand what would be good

It's too abstract, to you, to understand? That's a shame!

But you seem to be melting down instead of answering the most basic questions about what you're talking about. How and why would you use "tensor-based learning" to layout a graph?

Again, you're trivializing for...some reason?

Because graph layout features were invented essentially right after people started using node graphs. If a node to the right of another node is connected to an input to the left and the node to the left is connected to the input on the right, their wires will cross.

so I can't see your commentary as anything other than intransigent ignorance

I asked the most basic follow up question to what you said and you think you're being persecuted. Where does the expectation come from to make some sort of wild claim and have no one at least ask how it's supposed to work?

What is all this stuff about being "stubborn" and "closed minded" when you haven't explained anything? Just back up what you're saying somehow.

whatever it is that is causing you to have this reaction

You mean asking basic follow up questions?

> What do you think 'tensor-based learning' or 'training data' would do for graph layout?

I didn't and don't have a comment on that.

That's what I replied to and the question I asked.

If you want to group nodes by some context, why not just do it? This doesn't seem complicated to me. I don't know why there would need to be 'tensor based learning' 'machine learning' or even some one else's algorithm.

which I think is what is really required for interactive graphs.

Why would this be required for 'interactive graphs' and what does that mean? If you don't know of any graph that does this and you think it's required, do you have some definition of 'interactive graphs' that has never been done?

Wouldn't that need an awful lot of training data? And different people might have very approaches to laying out a graph, so all that training data would need to come from one person, wouldn't it?

Oh, for sure! It would need to be trained on the 'default' way of doing things, just like a linter has a default config. THEN it would need to train on how you code (pin ordering, reroute node layouts, loop layouts, etc), the entire time you work with it. And, of course, in order to train, you would need to be able to tell it "hey, don't do that; I know I did it here but don't mimic that" as well as "mimic this, every single time. Don't use the defaults". Those explicit commands are extremely important for handling the edge casing (similar to newline vs inline brace openings).

But, honestly, that kind of thing is the only valuable future for AI that I see, as a creative. I don't need it to write my code for me, but I could sure as hell use it to handle all the tedious bullshit that we've cobbled together as far as tooling, CI, formatting and any other "had to do it this way to make machines understand it" half-measure we've had to take in the past century. The AIs that we see, now, which learn from a huge dataset and then stay mostly inert until the next big update are more toy-like in my figuring. They are standards to measure against and proofs of all kinds of very interesting and useful concepts. But the lasting/useful AIs will be the ones that train alongside your work and ONLY on your work (or, let's face it, the work done in your company that will be shared amongst all corporate employees so that all work can be done more efficiently) in order to provide you with what you would be trying to do anyway.

Reply on news.ycombinator.com