AI safety is extremely bottlenecked on grantmakers by lukeprog

Last month, An­thropic an­nounced Mythos Pre­view, the most pow­er­ful cy­ber­weapon in his­tory, ca­pa­ble of find­ing and ex­ploit­ing zero-day vuln­er­a­bil­ities in ev­ery ma­jor op­er­at­ing sys­tem and web browser. Mean­while, many fron­tier AI com­pany em­ploy­ees in­creas­ingly ex­pect full au­toma­tion of AI R&D in the next year or two, fol­lowed by the rapid au­toma­tion of thou­sands of other im­por­tant tasks and jobs. This pace of tech­nolog­i­cal change is un­prece­dented, and the ...