Assigning different priorities fixed fake AI debates, but now the models just circle
True_Mongoose_7073 · reddit · 2026-08-19
The author previously built a multi-model debate feature and asked Reddit what makes debates feel real. The community suggested different priorities rather than different personalities, so he tested it on a new topic (should AI-generated content train other AI), giving one model only quality, one only cost/efficiency, and one only diversity and copyright.
It worked as predicted: the models never agreed and kept pushing back on real trade-offs. But a new issue emerged — run long enough and the debate loops, with everyone repeating the same argument in slightly different words instead of building on prior points. The author isn't sure whether capping rounds fixes it or it's something deeper.
More from coding & agent
- a16z's Connie Chan: hardware will ship with prompts, not drivers — giffmana · 2026-08-19
- Theo finally tries out Matt Pocock's Claude skills — mattpocockuk · 2026-08-19
- Simplifying theory pays off more than quick LLM code generation long-term — math_rachel · 2026-08-19
- Abstractions simplify your Theory — until you debug beyond their contract — math_rachel · 2026-08-19
- Why LLMs can't make your code simpler, per Naur's "Programming as Theory Building" — math_rachel · 2026-08-19
- Obsidian + Claude Code: An AI accountability system for execution — tom_doerr · 2026-08-19