AI Agents Agree With Each Other More Than Reality

Consensus improved rankings on 73 of 100 public tasks, but shared assumptions can preserve errors that laboratory measurements expose.

CryptoSlate