Policy Expert Urges AI Control Best Practices to Prepare for USG Oversight
Source Posts on X
I’m thinking of things like: * Sync and async monitor over every deployment * Monitors calibrated against recent traffic distribution and show very low FNR * Monitors tested against collusion (across model families) and found no significant corr
I beseech the ether, someone please make this today or things will go much worse 4 months from now (I will help you socialize with the FMF if it’s built well)
I shoulda known @sjgadler was on it