These are my working notes for a (possibly new) project. It’s a draft notebook of open questions I have and ideas I think are interesting, including thoughts and perspectives I might not necessarily agree with.

<aside> ✨ AI Summary

This document discusses the author's working notes and ideas for creating better interfaces for dialogue. The author explores concepts such as branching, spatial interfaces, searching in the context of 2D branches, quotations and annotations, and provides a draft sketch of their proposed design. The document also references prior art and provides visual illustrations.

</aside>

Motivations

Stable Diffusion XL: “3 or 4 boy and girl friends talking to each other with animated expressions, a web of hundreds of floating mystical 3D vibrant colorful speech bubbles with textlines glowing above their heads, dreamlike from Studio Ghibli, color-grading, intimate candlelit fireside lighting”

I started building a personal interface for dialogue with language models called Dual, because I wanted a proxy I could use to switch between speaking with ChatGPT and my own self-hosted language models. Then I realized I could make the interface itself more interesting than the one commonly found.

I’m constrained by my own API, which provides a single endpoint, /chat, that takes messages in an existing conversation thread and generates one new message from the language model. This seems general, but prevents use cases like:

• Generating potential human responses
• In-filling conversation transcripts (generating messages that may have come before others)

Some ideas I started with:

• Branching, tree-like UI

  • Spatial canvas in which to arrange different branches of a tree-like conversation

• I like the texture of a dialogue that happens on a shared surface like a doc rather than a sequence of separate speech bubbles.

  • I used use Google Docs often for conversations with friends, because you can build any arbitrary data structure for your conversation as it’s suited to the interlocutors’ use cases and structure of the dialogue.
  • I like that when the dialogue happens on a single shared surface rather than in message bubbles, the conversation feels as if the responder and I are building a shared artifact together, rather than throwing little texts at each other.

• Searching past threads throughout a conversation

  • As I navigate a conversation, I want past related conversations to surface nearby.
  • What should be the “unit” of a search result? A thread? A message? Some smartly subdivided chunk?

• What exactly is a reply? There are two ways to look at a conversation:

  1. A conversation is horizontal first. First, I lay out N comments, and then you respond to any of them by going deeper. Here, the dominant axis is horizontal, the “all the replies to this message” axis.
  2. A conversation is vertical first. Our messages follow each other in a line, and when either of us has multiple responses, we “offshoot” to the side. Here, the dominant axis is vertical, the “main thread of the conversation” axis.

  In particular, I don’t like that in many existing threaded conversation UIs (e.g. Slack or iMessage), having more than one responds to a message means that only one of those responses must be deemed the “main” response, and all other sub-threads become relegated to some deeply buried “thread” UI. I want all responses to a message to be deemed equally meaningful in the interface.

Working notes

Chatting with Glue

Conversations are fundamentally nonlinear, and forcing linearity on it often leads to strange workarounds and cowpaths.

Glue also makes many references to Outlines, a specific interesting kind of data structure for textual information.

Spatial interfaces