
TRC #663: Interview with Dr. Roman Yampolskiy About The Threat Of Advanced AI
The Reality Check
00:00
Is Deception Part of Multivariate Alignment?
"I think it's absolutely necessary that you don't remember entering the simulation," he says. "If they think everything in their world is simulated, they may just not find meaning in life." The main idea is to figure out if the problem with multi-agent alignment is so obviously impossible, how can we reduce it to a single agent problem which may be solvable? It's still difficult. You're still not the same person at different time points,. But it's way easier than doing it with, you know, multi-agents set up.
Transcript
Play full episode