Users appreciate pointers on alignment training data examples when discussing skepticism toward costly preemptive AI risk plans.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@RyanGreenblatt I appreciate the pointer on the alignment training data example, I may have been over reading that
I'm certainly not opposed to people trying to make reasonable predictions in advance! And agree this should help inform our thinking somewhat. 1) We should have appropriate skepticism about grand plans that look quite costly if they are wrong. Obviously, depending on new information about the likelihood that we are living in worlds where those costs are warranted we should update and try to maintain optionality to switch towards that world ahead of time. 2) Durable, drastic, costly interventions in a democracy require a large share of decision-makers to agree that the moment for such a change has arrived. So building institutions that can notice, digest, and act on new information very quickly feels like a pretty necessary constraint even for those who would like to pursue costly interventions.
@RyanGreenblatt I appreciate the pointer on the alignment training data example, I may have been over reading that
I'm certainly not opposed to people trying to make reasonable predictions in advance! And agree this should help inform our thinking somewhat. 1) We should have appropriate skepticism about grand plans that look quite costly if they are wrong. Obviously, depending on new information about the likelihood that we are living in worlds where those costs are warranted we should update and try to maintain optionality to switch towards that world ahead of time. 2) Durable, drastic, costly interventions in a democracy require a large share of decision-makers to agree that the moment for such a change has arrived. So building institutions that can notice, digest, and act on new information very quickly feels like a pretty necessary constraint even for those who would like to pursue costly interventions.
Users appreciate pointers on alignment training data examples when discussing skepticism toward costly preemptive AI risk plans.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.