Podcast: The Dawn of AI (Waking Up With Sam Harris)

A fascinating discussion with Stuart Russell on building AI “compatible with human well-being”.

You can find ‘Waking Up with Sam Harris’ on most podcast networks. Most shows are not related to AI or tech.

Particularly interesting piece at ~46:40 on wanting machine objectives to be “explicitly uncertain”. Stuart highlights that the majority of research in the field exists under the assumption that the objective specified for any machine by a human is assumed to be correct by default. This can cause all sorts of downstream problems as the machine interprets objectives in unforeseen ways (i.e. Nick Bostrom’s paperclip maximizer thought experiment)

The idea is that the machine would inherently assume that it may have misinterpreted the objective and be trained to constantly validate its understanding to course-correct. An example might be previewing it’s intended course of action, showing potential undesirable outcomes that result from it’s misunderstanding of its given objective.

It’s the first time I’ve seen the idea in my reading, and this strikes me as a promising way to counter-balance our inability to know if the machine has truly understood our intent.