I already talk about my views pretty freely in public, but I do think people at labs tend to be a bit cagier and less blunt about these things than is ideal, so in the interests of normalizing such behavior: I think Earth’s approach to AI development is incredibly reckless, a competent and sane civilization would not permit the trajectory we’re on, and a global moratorium on all aspects of AI capability progress for the next few decades would be a substantial improvement over the status quo (though it’s probably not my optimal policy given the background rates of non-AI x-risk and value lock-in, and I’d likely favor a controlled 5-10x slowdown with pause optionality).
I think my work on capabilities at Anthropic is probably good on net because I think it’s valuable for them in particular to have greater influence, but I consider the acceleratory effects of such work to be a cost. If I thought large fractions of people at other labs felt similarly and I wasn’t playing against defect-bot, I would probably quit for decision theory reasons, but in practice I don’t think meaningful coalitions elsewhere are running anything like my decision process.
FWIW I pretty regularly express such views internally and do not find doing so to be particularly socially costly, though I wouldn’t expect this to be true of the social incentives at other labs.
a global moratorium on all aspects of AI capability progress for the next few decades would be a substantial improvement over the status quo
Saw some shrug reacts on this so wanted to elaborate a bit—I’m not super confident about this (maybe like 70% now rising to 80% the later we implement the pause), and become a lot more pessimistic about it if the moratorium does not cover things like hardware improvements, research into better algorithms, etc. I’m also sort of pricing in that there’s sufficient political will to make this happen; the backlash from a decree like this if in fact most ordinary voters really hated it seems likely to be bad in various ways. As such I don’t really try and do advocacy for such changes in 2025, though I’m very into preparing for such a push later on if we get warning shots or much more public will to put on the brakes. Happy to hear more on people’s cruxes for those who think this is of unclear or negative sign.
I already talk about my views pretty freely in public, but I do think people at labs tend to be a bit cagier and less blunt about these things than is ideal, so in the interests of normalizing such behavior: I think Earth’s approach to AI development is incredibly reckless, a competent and sane civilization would not permit the trajectory we’re on, and a global moratorium on all aspects of AI capability progress for the next few decades would be a substantial improvement over the status quo (though it’s probably not my optimal policy given the background rates of non-AI x-risk and value lock-in, and I’d likely favor a controlled 5-10x slowdown with pause optionality).
I think my work on capabilities at Anthropic is probably good on net because I think it’s valuable for them in particular to have greater influence, but I consider the acceleratory effects of such work to be a cost. If I thought large fractions of people at other labs felt similarly and I wasn’t playing against defect-bot, I would probably quit for decision theory reasons, but in practice I don’t think meaningful coalitions elsewhere are running anything like my decision process.
FWIW I pretty regularly express such views internally and do not find doing so to be particularly socially costly, though I wouldn’t expect this to be true of the social incentives at other labs.
Saw some shrug reacts on this so wanted to elaborate a bit—I’m not super confident about this (maybe like 70% now rising to 80% the later we implement the pause), and become a lot more pessimistic about it if the moratorium does not cover things like hardware improvements, research into better algorithms, etc. I’m also sort of pricing in that there’s sufficient political will to make this happen; the backlash from a decree like this if in fact most ordinary voters really hated it seems likely to be bad in various ways. As such I don’t really try and do advocacy for such changes in 2025, though I’m very into preparing for such a push later on if we get warning shots or much more public will to put on the brakes. Happy to hear more on people’s cruxes for those who think this is of unclear or negative sign.