[bmacho]

> Yeah; and the merging is correct in all cases.

No, it's obviously not.

I'd say most of the time the merging is incorrect, that is, it is in contrast to both of the parties intention. Unlike git, which actually notifies them, Loro just silently modifies both changes into something nonsense/wrong.

[josephg]

My correctness point is that Git can end up with commits which spuriously conflict with themselves. Git's merge algorithms simply aren't very good.

CRDTs (and REG) are correct in that everyone always ends up seeing the same document state. But I take your point - which is essentially that Loro (and automerge, yjs, diamond types, etc) don't correctly preserve intent in all cases. As you point out, if two users concurrently edit the same word, you can get nonsense.

In practice thats usually fine when users edit in realtime - since users notice and fix it. But its often not ok while users edit asyncronously (eg while offline). Thats exactly why I want a CRDT type system which can emit git style conflicts.

[cubefox]

A while ago we would have said a computer program obviously can't guess the intent behind edits. But now we do have such programs: Language models. I wonder whether anyone has already used one to automatically solve merge conflicts.

[josephg]

That’s a good idea that I’ve heard a lot of people suggest. But as far as I know nobody has actually implemented merging using a llm yet.

Part of the problem is that CRDTs want to have strong eventual consistency - so, after merging everyone should be looking at the same final result. It’s hard to guarantee that if you pass the conflicting data to a LLM.

[rubymamis]

Exactly. This process of putting users in control by notifying them about changes and allowing them to decide how to merge is essential when dealing with text.

It sounds like Loro can do that too, but then again, is it worth the complexity over git?