How should multiple AIs reach consensus?
A developer is building "Chat Brawlers," a side project where three AI models independently answer, debate, and vote. Initially, the developer mistakenly treated model agreement as the best answer. For example, if Model A suggested "Use Postgres" and voted for it, the system would consider it. However, a trade-off exists: a judge might overlook important conditions, as "Use Postgres" and "Use Postgres unless X" are not always equivalent. So far, the judge has considered these details.
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年10月29日 19:43 UTC
收录当时偏移：UTC+02026年10月30日 00:00 UTC
- 发布
- 2026年10月29日 19:43
- 收录
- 2026年10月30日 00:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
I’m building Chat Brawlers, a side project where three models answer independently, read each other's responses, debate and vote. At first, I made one design mistake: treating their agreement on an answer as a way to find the best answer.
Originally, the debate stopped only when all three voted for the same model. But imagine:
- Model A: “Use Postgres.” Votes for A.
- Model B: “Use Postgres.” Votes for B.
- Model C: “Use Postgres.” Votes for C.
That's a tie, even though there’s nothing left to debate.
I changed the stopping rule: after each round, a separate judge compares their conclusions. If they give the same answer, the debate can end even when the votes are split.
There's a trade-off: the judge might overlook important conditions. “Use Postgres” and “Use Postgres unless X” aren't necessarily equivalent. So far, the judge has considered these details.