Building a Fully Automated Satirical News Station on a Single RTX 3090
sysadmin420 · reddit · 2026-08-08
A developer shared INFORMANT, a self-hosted satirical news station running entirely on a single garage RTX 3090. The system measures social media posts against real market data, then automatically writes scripts, generates voices, renders videos, and publishes to YouTube with zero cloud inference.
Tech Stack & Engineering Highlights:
- Models: Uses qwen3.6:35b (MoE) for writing (chosen for fewest invented figures), qwen3-vl:8b for vision, and F5-TTS for voice cloning.
- Single-GPU Choreography: Dynamically squeezes the resident LLM's VRAM via ollama's numgpu option to free up memory for TTS jobs, then restores it—no model restarts or OOM.
- Anti-Hallucination (Core): The hardest problem is that local LLMs confidently invent numbers. The system enforces that every numeral must match licensed data, or it's replaced on-air with "some number." This fail-closed design prioritizes honesty over fluent lies.
- Failure Case: The anchor once learned from its own flawed past scripts and started nightly saying, "I am the some number who is not fine," which required adding a new filter to the memory loop.
More from coding & agent
- okf-rs: A Rust Tool for Turning Codebases into AI-Readable Knowledge Bases — adnan_hashmi · 2026-08-08
- AI Writes and Reviews Code: Are Humans Reduced to Mere Testers? — petergyang · 2026-08-08
- GitHub Hit: Agent Orchestrator Manages Parallel AI Coding Agents — tom_doerr · 2026-08-08
- Building a Coding Agent from Scratch: Harness Design and Tool Orchestration — dl_weekly · 2026-08-08
- Anthropic Sets Claude Code to Auto Mode by Default for Safety — The Decoder · 2026-08-08
- Moving Local AI Agents to Remote Boxes Controlled via Discord — mattpocockuk · 2026-08-08