3 ms·
Learning human objectives by evaluating hypothetical behaviours
- carapace 7y agoThese systems (large, automated, computer-controlled, and capable of unsafe states) should be trained in such a way as to leverage each others' knowledge. E.g., each brand of auto-auto (self-driving car) should share information with the others in re: safety and unexpected conditions. This should happen in (near-)real-time.