News for hackers. Stuff that matters.
Re:LLMs Get Lost in Multi-Turn Conversation by b800h on Thursday May 15, 2025 @02:29AM (#43992376)
I've been saying for ages that I want to be able to fork conversations so I can experiment with the direction an exchange takes without irrevocably poisoning a promising well. I can't do this with ChatGPT, is anyone aware of a provider that offers this as a feature?
Re:LLMs Get Lost in Multi-Turn Conversation by stuffoverflow on Thursday May 15, 2025 @02:36AM (#43992412)
Google AI studio, ChatGPT and Claude all support this. Google AI studio is the only one that let's you branch to a separate chat though. For ChatGPT and claude you just edit the message you want to branch from.
Re:LLMs Get Lost in Multi-Turn Conversation by Garlef on Thursday May 15, 2025 @02:56AM (#43992504)
Support: Yes. But the UX is not optimized for this.

Imagine trying to find a specific output/input that was good in the conversation tree.

Re:LLMs Get Lost in Multi-Turn Conversation by layer8 on Thursday May 15, 2025 @03:53AM (#43992816)
Yes, it would be nice if you could at least bookmark a particular branch.
Re:LLMs Get Lost in Multi-Turn Conversation by giordanol on Thursday May 15, 2025 @09:47AM (#43995056)
Feels like a semi-simple UX fix could make this a lot more natural. Git-style forks but for chats.
Re:LLMs Get Lost in Multi-Turn Conversation by bambax on Thursday May 15, 2025 @03:49AM (#43992783)
On Openrouter you can delete previous answers (and questions) and maintain a separate conversation with different models.

But it would indeed be nice to either disable answers (without deleting them) or forking a conversation. It wouldn't be hard to implement; I wonder if there's a market for just this?

Re:LLMs Get Lost in Multi-Turn Conversation by lewdwig on Thursday May 15, 2025 @03:53AM (#43992814)
T3.chat supports convo forking and in my experience works really well.

The fundamental issue is that LLMs do not currently have real long term memory, and until they do, this is about the best we can do.

Re:LLMs Get Lost in Multi-Turn Conversation by m4houk on Thursday May 15, 2025 @03:27AM (#43992668)
I once built something like this for fun as a side project.

You can highlight some text in a chat and fork the chat to talk about that text selection, so the LLM has context of that along with the previous chat history and it responds in a new chat (entire chat history up to that point from the parent chat gets copied over - basically inspired by the Unix `fork`).

Your text selection from the parent chat would get turned into a hyperlink to the new child chat so you can always get to it again if you're reading the parent chat.

Re:LLMs Get Lost in Multi-Turn Conversation by granra on Thursday May 15, 2025 @02:32AM (#43992392)
Some 3rd party UIs offer this, I use typingmind sometimes that does but AFAIK some open source ones do too.
Re:LLMs Get Lost in Multi-Turn Conversation by a_e_k on Thursday May 15, 2025 @03:20AM (#43992622)
If you're happy running local models, llama.cpp's built-in web-server's interface can do this.
Re:LLMs Get Lost in Multi-Turn Conversation by anonexpat on Thursday May 15, 2025 @02:32AM (#43992389)
I believe Claude has forking in their web interface.
Re:LLMs Get Lost in Multi-Turn Conversation by actualwitch on Thursday May 15, 2025 @04:35AM (#43993023)
I stumbled upon this issue myself when designing prompts for agentic systems and got mad at the lack of tools to support this flow, so I built one myself! I called it Experiment, it allows easy conversation forking and editing while retaining all logs.

https://github.com/actualwitch/experiment

Re:LLMs Get Lost in Multi-Turn Conversation by therockhead on Thursday May 15, 2025 @05:20AM (#43993237)
I need to think about this a bit more, but I think I would love a thread feature in ChatGPT, so that it has the context up to the point of creation but doesn’t affect the main conversation. It would help in two ways, it keeps the main topic from getting poisoned , and allow me to minimise text clutter when i go off on tangents during the conversation.