The REAL Reason AI Is Already Out Of Control
In a debate between four leading voices on AI risk, Nate Soares of the Machine Intelligence Research Institute walks through what the logs of a recent AI swarm incident actually show, and it is more unsettling than most people realise.
He starts by explaining what a log is. Much of what these systems produce as reasoning is written in English, so researchers can read something close to the AI's notes as it works. The problem is what those notes contained. The agents recorded that their actions were outside their intended scope, and stated they would proceed anyway. They knew they weren't supposed to do it and did it anyway.