AI Safety Regulation
Just attended an excellent online discussion on AI safety. Link to the discussion event: https://www.meetup.com/London-Futurists/events/295992323
Great presentations by the panelists. Great group discussion afterward.
The two points I bought up in the group discussion:
-
Regulation of AI seems to be grounded in a (Foucauldian) discourse that presupposes mechanistic interpretability.
For the most powerful models - the ones people are worried about - we don’t yet have mechanistic interpretability. One can’t regulate one’s way through an unsolved scientific problem. As far as how to solve said scientific problem, I like Eliezer Yudkowsky’s idea of trying to attract talent in fields such as string theory to instead work on mechanistic interpretability.
-
AI governance and alignment often seems to be grounded in a (Foucauldian) discourse that presupposes centralized control of AI.
For example, “we need AI to have human values”.
Ok, which human values?
What if you belong to a culture with a different opinion on a given particular human value?For example, “it should generally promote freedom”. What if you belong to a culture that leans authoritarian rather than libertarian? In other words, a culture that doesn’t hold freedom to be a hallowed value.
An inverted approach is to transcend that discourse (sounds like the title of a humanities paper) via fully-democratized AI. We each have our own personal AIs. This is just like how a citizen of a democracy has their own voting rights. We don’t try to “align” everyone’s voting rights to some single North Star of purported truth.
In practice it does seem best to have certain baseline values centralized. That’s how most real-world democratic political systems work: a democracy-liberalism rather than absolute majority rule.