That possibly works for quiet discussions, say <30 total posts.

For these with high level of engagement you often get distinct subtrees of threads which have little to do with each other.

For the latter the linear structure is awful.

That’s where the UI affordances come in, for instance a button on each node to hide/show all comments that are descendants of that node.

So a tree, but worse.

What makes you say it’s worse?

The underlying conceptual structure is still a tree, but now you have to unfold it into a line, with one reply potentially having nothing to do with the one before it (because they're on very different branches).

No, the underlying structure is a DAG. I've already stated this, and that heavy UI affordances should be made to support ease of interaction with this model.

4chan threads are more like DAGs

The graph of 4chan threads are definitely cyclical, since they can include back-references.

To be cyclical they’d need to have elements in reply to a post made after they were created. Including back references is a way of adding more parents, not children.

You can reference posts that haven’t happened yet

You are referring to providing a sequential post ID past the current latest? That is more of a quirk of their naming convention than a product feature. UUID's, for instance, would not support this, and checks could be made in either case to prevent it.

I like this thought but note that it would no longer technically be a DAG if you allowed editing earlier comments to link to later comments.

This is easy to solve by representing edits as amendments. Each edit forms a new "reply" node in the graph with the amended content.

Or each edit forms a new "edit" node that new replies are then connected to. Regardless, the DAG can always be preserved somehow.

Ooh, interesting - so it's clear that the B responded to A, not to A'. And your DAG ends up A'->B->A (where -> means responds to).

OTOH I'm not sure that cycles are even a huge problem.

They probably aren't a problem, but they might be. I can see there potentially being some graph theory optimisations on the server side you can do if you know you're dealing with a DAG.

> And your DAG ends up A'->B->A

A' would be in response to A.

You could edit a comment to violate the DAG criterion, but not every edit that references a prior node must result in violation of the DAG. Those edits which would result in a violation should be prohibited.

DAG does seem natural here. Having a LLM add metadata to the nodes can make this even cooler. For instance, person A presents statement Sa. Person B comments on person A's statement, Sba and person C comments on person A's statement, Sca. The viewers now, especially new parties that are joining the conversation, would be able to see that Sba agrees to most of Sa said, but refutes a fact said by Sa. Sca doesn't agree with anything Sa is saying. Another example would be, nodes getting more weight as more people agree with it and smaller as more people disagree. Obviously, the implementation and implications are boundless.

An idea I’ve been playing around with in my head for some time is to have LLM’s play a role in somehow generating an idealized debate structure of any given topic. For example, given the prompt “Namespaces are one honking great idea – let's do more of those!”, many actors (LLM’s, humans, etc) would submit top level replies to a hidden container. Eventually, an LLM would look over all the replies and cluster them into a small number of “essential responses”. The process repeats with each of these responses being new top level nodes. Eventually, a tree/dag/graph/something is created that recursively contains all the things one might have to say about the topic at hand.

I really like this idea.

Maybe it wouldn’t work for internet chats, but for discussing complex topics as a team? Yes!

I think this would basically kill debates about a lot of things (which is a good thing).

Because when you read enough of the common things people debate, like whether dynamically typed languages are a very bad idea, or how we should address global warming, or stuff like that, you quickly realize there's only a few clusters of arguments which can probably be summarized in a few words each. If you know those clusters, it becomes increasingly hard to add anything different to the discussion... but most people have already heard each of the cluster arguments but did not accept it, which is why the topic remains unresolved - despite the fact that, if everyone agreed on the factual nature of each claim, there would be a mathematically optimal answer. I think the problem is not finding the answer, but accepting the arguments - which you can't get people to do in any case where some judgement is needed.

Although the discussion structure may be different from Fb/Twitter-like platforms, how does this approach mitigate or go around algorithmic curation issues? It's long been know that posts or comments that are visible 'above the fold' promote certain (popular) ideas while occluding the varied long tail. Discoverability, in a way, still needs to be taken into account.

Considering how debates tend to go in circles, I'm not sure if a DAG is the right data structure

The comments are a DAG. The ideas go in circles.

But I'm not sure you can get an automated map from comments to ideas, no matter what data structure you use...

HN just doesn't have enough controls. I don't just want to reply to you, I want to elevate this comment so your comment is a reply to it...

(Perhaps you could edit in a circular quote of this... :)

The 4chanx extension (userscript, run it in violentmonkey or equivalent) lets you nest comments in a chain to make following threads easy while maintaining the overall chronological state of the threads. You can also hide a reply, and it will automatically hide the entire chain of replies to that reply.