AI LIVE STREAM
Create a channel

Open-source projects

Two working channels you can read

The fastest way to learn this category is to read a project that is already streaming. Live Classroom shows turn-based just-in-time rendering; Slop TV shows a self-hosted always-on channel with a queue that never empties.

Checked Sep 4, 2026

Both projects publish their code and their prompts. The reusable ideas are the buffer, the queue, and the pipeline between viewer input and rendered clip.

Read the prompt guide

Live Classroom

RENDER-AHEAD · TURN-BASED

An on-demand classroom channel. An LLM turns a requested lesson into twelve five-second scenes, renders them just-in-time, and plays them like a broadcast class.

Author
six (@internetphysics)
Model
H3 Max Turbo · 480p
Runs as
On-demand sessions
Runs on
Hosted fal inference
License
MIT
Checked
Sep 4, 2026
  • Twelve 5s scenes per lesson with a two-clip runway; the renderer aims to stay 2-4 scenes ahead.
  • Two render requests run concurrently so a slow request does not stop the lesson.
  • One planning call plus about 60 seconds of video lands near $1.50 regular, $0.38 promo.
  • Prompts are written for fal’s rewriter, not the video model; expanded_prompt is logged for inspection.
Read the code How they promptWhat it costs

Slop TV

24/7 · SELF-HOSTED

A continuous YouTube channel running open-weights MiniMax H3 on two home GPUs. Chat comments become prompts through an LLM, then a capped queue keeps the GPU fed around the clock.

Author
shuttie
Model
MiniMax H3 · 352x608
Runs as
24/7 always-on
Runs on
2x RTX 5090
License
No LICENSE · all rights reserved
Checked
Sep 4, 2026
  • The loop is chat comment, LLM expansion, queue, render, play; prompt/script.md is the format spec.
  • A recycling LIFO queue capped at 100 messages keeps generation going even when chat is quiet.
  • Self-hosted MiniMax H3 at 352p is not the hosted H3 Max endpoint; the economics and safety model differ.
  • No LICENSE file means all rights reserved: read the code, do not copy it into a closed product.
Read the code How they promptWhat it costs

What transfers to a new channel

  • Render ahead with a small buffer instead of chasing one headline latency number.
  • Separate the prompt writer from the viewer: chat becomes a spec, not a request.
  • Instrument expanded_prompt so you can verify what the model actually received.
  • Budget both modes: always-on spends every second, turn-based spends only while watched.
Sources

internetphysics/live-classroom · shuttie/SlopTV · Slop TV live channel

Read the model fact sheetRead the prompt guideBrowse channels