I always feel that AI Strategy not deeply informed by a technical understanding of the alignment problem is misguided (i.e. your models of and uncertainties around the alignment problem determine ~90% of the variance in strategy). This is to be contrasted with quetions like “Whose values ought to be imbued in the AI?” or “Should industry be allowed to create AI?”. Whether industry is a moral authority or not is one question, but simply asking whether its leaders have deep security mindset is a better question, not least because it immediately zooms you in on the small set of possible AI projects that could ever be alignable.
i.e. your models of and uncertainties around the alignment problem determine ~90% of the variance in strategy
I don’t think that’s the case in most groups. There is a huge amount of disagreement even holding fixed views about the difficulty of the alignment problem, with more of the variance explained by questions about what different groups of people might agree to, the dynamics of different kinds of conflict, the internal dynamics of a group with a nominal commitment to safety, etc.
I think that if we’re talking about modern-day policy, the alignment problem isn’t really relevant yet—we have no idea of how the alignment problem should inform policy at this stage—wheras questions like “whose values should be imbued in AI” (or rather, in ML applications) are becoming very relevant.
I always feel that AI Strategy not deeply informed by a technical understanding of the alignment problem is misguided (i.e. your models of and uncertainties around the alignment problem determine ~90% of the variance in strategy). This is to be contrasted with quetions like “Whose values ought to be imbued in the AI?” or “Should industry be allowed to create AI?”. Whether industry is a moral authority or not is one question, but simply asking whether its leaders have deep security mindset is a better question, not least because it immediately zooms you in on the small set of possible AI projects that could ever be alignable.
I don’t think that’s the case in most groups. There is a huge amount of disagreement even holding fixed views about the difficulty of the alignment problem, with more of the variance explained by questions about what different groups of people might agree to, the dynamics of different kinds of conflict, the internal dynamics of a group with a nominal commitment to safety, etc.
I think that if we’re talking about modern-day policy, the alignment problem isn’t really relevant yet—we have no idea of how the alignment problem should inform policy at this stage—wheras questions like “whose values should be imbued in AI” (or rather, in ML applications) are becoming very relevant.