It also tests if agents proactively request secret information.
.@ChrisPainterYup can you guys introduce a very simple interaction eval where you take your existing long horizon eval and add in neutral interrupts by the human?
then you can step this up to secret info that speeds up or actually allows model to solve things IFF it talks to human?
Users thank influential players for proposing human interrupt evaluations on long-horizon AI agents, viewing the idea as an important step against gradual disempowerment.
@alth0u Interesting, will share with team
.@ChrisPainterYup can you guys introduce a very simple interaction eval where you take your existing long horizon eval and add in neutral interrupts by the human?
then you can step this up to secret info that speeds up or actually allows model to solve things IFF it talks to human?
@ChrisPainterYup thank you i think small things like this from influential players go a long way in fighting gradual disempowerment but more importantly make sure models dont hit some ceiling where sim2real hurts them because they aren't trained for humans to be part of the environment