Local LLM Runs a Party's Laser Wall; Guests Started Performing for It Within 78 Seconds

vanadiuz · reddit · 2026-09-27

At the ZOLAK art residence in Vilnius, the author is building a "dynamic AI party" where a local LLM decides in real time what a laser-projected wall shows — "the room is the prompt."

Hardware split (perception separated from the LLM to avoid VRAM contention):

Coral Edge TPU for person/motion detection on camera feeds
RTX ADA 2000 for live speech-to-text from mics
RTX 3090 running the LLM that turns "who's moving and what's said" into wall actions
CPU composites frames and drives the laser, closing the feedback loop

Observations from the first 78-second test:

Echo alone hooked people — throwaway words ("basically", "crazy") reappeared as huge projected type within seconds
Guests immediately tested for control, e.g. raising hands to see if the wall brightened; human behavior in a responsive room proved more interesting than the visuals
The loop changes the input: once people know the model is listening, they speak for the wall, so the model no longer sees a natural room

Open questions: how to keep a room-scale feedback loop from collapsing onto the loudest person, and how to honestly tell guests the room is listening (all processing is local) without killing the mood.

Original post →

Local LLM Runs a Party's Laser Wall; Guests Started Performing for It Within 78 Seconds

More from coding & agent