1 comment

[ 4.6 ms ] story [ 14.3 ms ] thread
Five months ago I built a system that autonomously produces satirical news content. Two consumer GPUs, local inference, no cloud bills. It crawls sources, generates scripts, renders audio/video, and publishes ~100 pieces (5 hours+) published every other day, limited due to YT upload quotas, can do more.

~6,000 hours published so far on all the platforms: https://doomscroll.fm https://youtube.com/@doomscrollfm

The interesting parts weren't entirely the models, but the job management system, and the TTS, that is always an interesting challenge.

I'm ex-security/devops. This started as a weekend experiment in autonomous systems and spiraled into something that just... runs and runs.

Solo project, no funding, minimal babysitting.

The content itself? It's satirical news. Some of it lands, some doesn't. The uncanny-valley feel left intentional, but the system works...

Happy to discuss the infrastructure challenges if anyone's curious about running this kind of setup on consumer hardware.