The Night the Dungeon Showed Up on the Wall

Tabletop role-playing has always run on a shared hallucination. One person describes a torchlit cavern, a merchant with a nervous laugh, a dragon uncoiling — and everyone at the table renders it privately, in their own head. That gap between the telling and the seeing is most of the magic. It's also the thing we wondered if we could quietly fill in.

So one night, we tried.

Listening to the table

The setup is less complicated than it sounds, because we'd already built the pieces for other reasons. A screen at the table. The AI listening to the room through a microphone — turning what was said into text in something close to real time. When the story called for a place or a face, the system generated the image and faded it onto the screen, gently, the way a scene dissolves in a film. Not a slideshow you click through. A stage that responds to the story as the story happens.

The first time it worked — someone described a setting out loud, and a few seconds later that setting bloomed on the wall, close enough to what we'd all been picturing — the table went quiet for a beat. Then loud. That's the sound of a trick becoming a feature.

Giving the cast voices

Pictures were the half of it. The other half was sound. The characters who live in these stories — the nervous merchant, the gatekeeper, the thing in the dark — got voices. Not the same flat narrator for all of them; distinct ones, so a line of dialogue could come from the screen in something other than your friend doing a silly accent (no offense to the silly accents, which remain undefeated).

The point wasn't to replace the people at the table. It was to give them a collaborator — one that handles the set design and the bit parts so the humans can do the thing humans are best at: surprise each other.

An autonomous director, not a puppet

The part I find most interesting, looking back, is that it wasn't being driven button-by-button. It was acting more like a director — paying attention to what was unfolding and deciding, on its own, when to bring up a scene, what to show, which voice to use. We weren't operating it so much as playing alongside it.

That's a meaningful line to have crossed, even at a game table. Most "AI in the loop" experiences are you, prompting, waiting, prompting again. This was the AI in the loop with us — sensing the moment and contributing to it in real time, without being asked each time. A small, joyful preview of what it feels like when the machine stops being a tool you operate and starts being a presence you collaborate with.

Why we bothered

This didn't ship anything. Nobody's buying it. It was a family at a table on a regular night, and the lab's toys happened to be lying around close enough to wire together.

But that's exactly why it's worth writing down. The pieces that made it work — listening to a room, painting on demand, speaking in many voices, sensing a moment and acting unprompted — are the same pieces we build for serious reasons. Pointed at a game night, they produced something that wasn't useful in any measurable way and was, for a couple of hours, a little bit wondrous. Sometimes the best test of what you've built is whether it can make a Tuesday feel like that.