Thanks for the comment! We looked at autonomy more generally with the toy model as only finetuning on one of the autonomy and weaponry was easier and would give more interpretable results (not changing two variables together).
With more time we’d have looked at it, as well as just regular weaponry, especially if the privacy erosion results didn’t elicit EM. Given the stronger privacy erosion dataset did elicit EM, we felt the point was sufficiently made so didn’t dive too much further. We also think there’s broader risks from autonomous weaponry than EM specifically that this might distract from.
Thanks for the comment! We looked at autonomy more generally with the toy model as only finetuning on one of the autonomy and weaponry was easier and would give more interpretable results (not changing two variables together).
With more time we’d have looked at it, as well as just regular weaponry, especially if the privacy erosion results didn’t elicit EM. Given the stronger privacy erosion dataset did elicit EM, we felt the point was sufficiently made so didn’t dive too much further. We also think there’s broader risks from autonomous weaponry than EM specifically that this might distract from.