Listen to Episode Here
Show Notes
Roman Yampolskiy, Professor of Computer Science at the University of Louisville, joins us to discuss whether we can control, comprehend, and explain AI systems, and how this constrains the project of AI safety.
Topics discussed in this episode include:
- Roman’s results on the unexplainability, incomprehensibility, and uncontrollability of AI
- The relationship between AI safety, control, and alignment
- Virtual worlds as a proposal for solving multi-multi alignment
- AI security
You can find FLI's three new policy focused job postings here
Have any feedback about the podcast? You can share your thoughts here
Timestamps:
0:00 Intro
2:35 Roman’s primary research interests
4:09 How theoretical proofs help AI safety research
6:23 How impossibility results constrain computer science systems
10:18 The inability to tell if arbitrary code is friendly or unfriendly
12:06 Impossibility results clarify what we can do
14:19 Roman’s results on unexplainability and incomprehensibility
22:34 Focusing on comprehensibility
26:17 Roman’s results on uncontrollability
28:33 Alignment as a subset of safety and control
30:48 The relationship between unexplainability, incomprehensibility, and uncontrollability with each other and with AI alignment
33:40 What does it mean to solve AI safety?
34:19 What do the impossibility results really mean?
37:07 Virtual worlds and AI alignment
49:55 AI security and malevolent agents
53:00 Air gapping, boxing, and other security methods
58:43 Some examples of historical failures of AI systems and what we can learn from them
1:01:20 Clarifying impossibility results
1:06 55 Examples of systems failing and what these demonstrate about AI
1:08:20 Are oracles a valid approach to AI safety?
1:10:30 Roman’s final thoughts
Paper's discussed in this episode:
Unexplainability and Incomprehensibility of Artificial Intelligence
We hope that you will continue to join in the conversations by following us or subscribing to our podcasts on Youtube, Spotify, SoundCloud, iTunes, Google Play, Stitcher, iHeartRadio, or your preferred podcast site/application. You can find all the AI Alignment Podcasts here.
