🎉 Check out my new project: Heuristi.ca 🎉

No, I don't want to chat with my documents

I wonder if there are any other people out there who are tired of these chat-based interfaces for large language models (LLMs). It makes me think if chatting with documents is that desirable. I don't even want to talk to people sometimes! I think there should be alternative ways of interacting with LLMs. Here are some limitations that I see with using chat as an interface.

Chat flow is constrained

You can get sidetracked in a chat. This implies that there is only one main track to begin with. This is a limitation. Keeping track of and maintaining multiple branches of conversations in a chat-based interface (or a real-life conversation) is hard. Everything eventually converges to one destination. This is apparent in the UI of ChatGPT as well. You have a sidebar of self-contained conversations that are isolated from each other. You can't get the results of one conversation to flow into another one. This means you can't remix conversations or compare and contrast them easily.

Chatting is full of redundancies

Kevin from The Office had a point. Why speak longer when you could get away with less? However, we don't seek those kinds of efficiencies in real life. It can result in misunderstanding and could be perceived as being rude. However, you probably don't want a lengthy conversation with an intelligent agent on your computer to open your browser, either. We could hide some of the redundancies of the chat behind abstractions.

LLMs as word calculators

It goes without saying that having a chat-based interface for an artificial agent has many advantages. However, I think it shouldn't be the only way to interact with LLMs. One of my favorite mental models about LLMs is to think of them as word calculators. However, this isn't obvious when working with a chat interface. We don't have explicit operators and operands. Everything is just too messy, buried under a pile of words.

I built Heuristi.ca to explore a different way of working with artificial intelligence. It is a node-based interface to GPT that allows you to visualize (and organize!) different flows of information all at once. It allows you to interact with AI by simply pressing buttons, so you don't need to type anything at all. Different buttons are different operators on a given textual input so that you can work with information in all sorts of different ways.

It doesn't afford the freedom of a free-form chat box, but having limitations sometimes makes for a better experience. It is certainly better for creativity. I am not saying it is a definite improvement over ChatGPT. However, it is a version of the future that is worth exploring.