AI alignment should leave room for social values to evolve

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 12:00

   On sort of agents, I saw your paper from July, which is called AI Value Alignment for Evolving Social Norms. And here, it's an interesting approach because it's rooted in social physics, which maybe you can provide a high-level understanding of. um but yeah so how do you guys use simulation and also like analytical methods to like figure out how you're going to safely deploy um ai systems and avoid this thing called value lock-in um i think briefly also to contextualize a piece of work there is a saying that goes uh all models are wrong some are useful uh this isn't necessarily saying that their work that work was uh you know extremely useful and biased because i've worked on it um but the purpose of it was to help us contextualize and think about certain issues that may arise in a more tangible way because these

2. 13:01

   issues of the consequences of taking certain technical approaches and applying them at scale in society they're very hard to otherwise reason about there is a lot of qualitative research that's done on these topics within the fields of social science and political science etc philosophy right of technology um but then when something is purely qualitative it makes it hard to measure the extent to which something happens which is a quantity is something very prevalent is something just a possibility um these things obviously matter when one considers risks because you can have a a very large let's say harmful outcome that is very unlikely or maybe less harmful outcome but is very likely so the likelihood of an outcome matters in contextualizing potential long-term harm is not just harms also benefits right so we developed some set of mathematical models that

3. 13:58

   would help us think about the consequences of different approaches to alignment when it comes to values and i guess the tldr of all of that with you know many bits and pieces of analysis that were contained within is that the thing that we are somewhat worried about is that alignment is often treated as this fixed set of universal values and rules and then the technical problem becomes how do we best align our models or agents to that fixed set of rules however you represent And then there are many ways, technically, of dealing with this.

4. 14:33

   Whereas in reality, what we often see in societies is that they change over time, they will over time. How we think about things today is not the same as how we thought about them two years ago or five years ago. And we as humans are working our way through and adjusting as we move along. So what we want to have is technology that doesn't restrict and hamper individuals from their own social evolution on this trajectory, but leaves us free enough that we can explore and not be in a filter bubble of sorts, right?

5. 15:11

   Because filter bubbles were obviously quite prominently researched in the context of things like social media in the past. we all know what their cons are so we do not want to be trapped too tightly maybe by technology but obviously we do want to have alignment and that brings up this interesting question of how does one achieve this alignment while not being unnecessarily restricted perhaps