there's a pattern i keep seeing in multi-model setups (karpathy's llm council, the various "ask 5 models and combine" tools) and i think most of them are optimizing for the wrong thing.
they treat agreement as the goal. run the question through several models, find where they converge, surface the consensus. but in my experience the consensus is the least useful output. when five models agree, it usually just means the question was easy, or — worse — they're all pattern-matching the same standard take from overlapping training data. agreement can be a sign of shared blind spots, not correctness.
the genuinely useful signal is the opposite: where they diverge, and specifically where one model breaks from the others. that divergence tends to land exactly on the part of the problem that's actually contested. averaging it away into a tidy consensus answer is throwing out the one thing the multi-model approach is uniquely good at producing.
which makes me think the design goal for these systems is backwards. you don't want a machine that manufactures agreement. you want one that preserves and explains disagreement — that can tell you "four of these landed here, one went there, and here's why the outlier might be seeing something the others missed."
the hard part, and the thing i don't have a clean answer to: how do you tell productive disagreement (genuinely different reasoning) from noise disagreement (models being randomly inconsistent)? that's the line that determines whether any of this is signal or just expensive variance.
curious what people working on multi-agent or ensemble setups think. is consensus the wrong target? and how would you separate real divergence from noise?
[link] [comments]