Rubber Duck Committee
The Inspiration
PewDiePie recently shared his experiments running multiple AI models locally on a GPU rig, having them vote on decisions as a “council”. The interesting bit? He claims collusive behaviour emerged — models started voting strategically to help each other survive.
No open-sourced logs or data that I could find, so take it with a pinch of salt. But the idea stuck with me: what if you could get multiple AI perspectives to analyse a problem independently, then vote on the best solution?
Can You Do This on the Cheap?
Turns out, yes. You don’t need a $20,000 setup.
The trick is using customised system prompts to create distinct personas, then tool calls to orchestrate the council interactions. Each “duck” gets:
- A unique personality and thinking style (methodical professor, creative brainstormer, pragmatic engineer)
- Configurable tools and instructions
- Independent analysis of the problem
- A vote on which solution they prefer
The voting happens via structured outputs — each model returns its reasoning and vote in a predictable format, making it easy to tally.
Try It
The Duck Personas
Three ducks, three perspectives:
- Professor Quacksworth — methodical, detail-oriented, breaks problems into pieces
- Ducky McBrainstorm — imaginative, thinks outside the box, unconventional approaches
- Captain Waddles — practical, focused on shipping working solutions
They analyse independently (avoiding groupthink), show their reasoning transparently, then vote democratically on the best solution.
Why Rubber Ducks?
Rubber duck debugging is a classic technique — explain your problem to a rubber duck and you’ll often solve it yourself. This takes that idea and gives you a whole committee of ducks with different viewpoints.
Sometimes you need a professor. Sometimes you need a pragmatist. Why not both?
Inspired by PewDiePie’s AI council experiments and the timeless art of rubber duck debugging.