changing the composition of people s minds. arguably we should have a phase where they have to be approved and they have to be tested, perhaps with focus groups and control groups to see, after three months of use of this new algorithm, let s say, recommending newsfeeds or be videos, do we see noticeable changes? if we do not understand how to get the objective specified correctly or we build systems that are different that do not pursue an objective so one of those two things has got to happen for us to move forward. is there an alternative to giving a! a strict subjective? the alternative is build ai systems that know that they do not know what the full objective is. when they know they don t know it so they will try to avoid messing with those things or if they have to mess with them, they
to move forward. is there an alternative to giving a! a strict subjective? the alternative is build ai systems that know that they do not know what the full objective is. when they know they don t know it, so they ll try to avoid messing with those things or if they have to mess with them, they will ask permission, so the system that might fix out carbon dioxide problem by turning the oceans into sulphuric acid, if it knows that it doesn t know our preferences about the oceans, then it s going to ask. i ve got this plan but it turns the oceans into sulphuric acid is that ok? we would say, actually, no, don t turn the oceans into sulphuric acid. and then it would ask, is it ok if i remove all the oxygen from the atmosphere? no, don t do that either,
of people s minds. arguably we should have a phase where they have to be approved and they have to be tested, perhaps with focus groups and control groups to see, after three months of use of this new algorithm, let s say, recommending newsfeeds or be videos, do we see noticeable changes? if we don t understand how to get the objective specified correctly or we build systems that are different that do not pursue an objective so one of those two things has got to happen for us to move forward. is there an alternative to giving a! a strict subjective? the alternative is build ai systems that know that they do not know what the full objective is. when they know they don t know it so they will try to avoid messing with those things, or if they have to mess
a strict subjective? the alternative is build ai systems that know that they do not know what the full objective is. when they know they don t know it, so they ll try to avoid messing with those things or if they have to mess with them, they will ask permission, so the system that might fix our carbon dioxide problem by turning the oceans into sulphuric acid, if it knows that it doesn t know our preferences about the oceans, then it s going to ask. i ve got this plan, but it turns the oceans into sulphuric acid is that 0k? we would say, actually, no, don t turn the oceans into sulphuric acid. and then it would ask, is it ok if i remove all the oxygen from the atmosphere? no, don t do that either, we really want to be alive and we need the oxygen to breathe. on a slightly more upbeat note, what are ai s biggest successes? things like healthcare. even in countries that have
is there an alternative to giving a! a strict subjective? the alternative is build ai systems that know that they do not know what the full objective is. when they know they don t know it, so they ll try to avoid messing with those things or if they have to mess with them, they will ask permission, so the system that might fix out carbon dioxide problem by turning the oceans into sulphuric acid, if it knows that it doesn t know our preferences about the oceans, then it s going to ask. i ve got this plan but it turns the oceans into sulphuric acid - is that ok? we would say, actually, no, don t turn the oceans into sulphuric acid. and then it would ask, is it ok if i remove all the oxygen from the atmosphere? no, don t do that either, we really want to be alive and we need the oxygen to breathe. on a slightly more upbeat note, what are ai s biggest successes?