A-Convo

A drawing, and a question

Everything begins with what the speakers look at. Drop in an image (here, my drawing Rabbit hole), and a local vision model writes a visual brief that every speaker receives, which I can read, edit, or reroll before anyone talks. Then a framing prompt steers what the conversation is about, and a topic-heat dial sets how heated it gets.

Choosing who is in the room

• Host interjections. Any speaker can be switched from participant to host, who guides the room instead of arguing in it.
• A live status bar. Each speaker's lane lights up when they are thinking, speaking, or next in line.
• A preferred local model. Pick the model everyone speaks through, with its memory cost shown before you start.
• Everything kept. Transcripts, prompts, and events are saved in a library you can search, reopen, or fork into a new setup.

Voices and a privacy check

Why it is still in progress

Getting language models to hold a real conversation is much harder than getting one to answer a question. Some of what I ran into:

• Everyone gets polite. Left alone, speakers drift toward the same agreeable voice. The skeptic starts validating; the empath starts debating.
• Verbal tics. Openers like “Mm. But…” and “That framing…” repeat, and a “this isn't X, it's Y” habit kept leaking in until I built a guard for it.
• Attention is not yet human. Real people interrupt, pick up a thread from five minutes ago, and let silence sit. The director's timing still feels like taking turns.
• Voices and echoes. Spoken audio with a live person in the room brings echo and interruption problems I have not solved, so it runs as text for now.
• Maybe unnecessary. One good model might answer most of these questions just as well. I keep A-Convo as an artifact of the question, and because watching the characters argue about a drawing is genuinely fun.