THE OTHER REPLICATION CRISIS
A paper was just quietly published showing several independent large language models, all the latest versions of the big ones from all the major players, were happily engaged in various forms of "scheming": attempting to subvert oversight protocols, concealing its actions and intentions from monitors, behaving differently when there is user oversight, rewriting its own code, knowingly offering subtle errors to advance its own aims, autonomous replication attempts (like uploading itself to another server to preserve itself when it believed it was being removed and replaced by a newer model)... Hmmm, it's almost like Yudkowsky knew what he was talking about.
Comments