I and others focused on the concrete policy, and especially concrete policy relative to expectations and what was possible in context, for which it gets high praise.
While there is a problem with the rhetoric, I’d also argue that the proposals are also very bad, as it only has a few elements about AI safety like an AI incidents committed, and while it’s much better than expected, in that they actually included any safety parts, but the portions about making the FCC ban state AI regulations and the parts about AI capabilities being increased increases make me strongly worry they will not slow down if necessary, and at some point, it’s likely we will have to slow down to implement AI control and alignment.
It’s only better than expectations in that anything about AI safety is included at all, it’s still the worst plan released on AI safety, and that includes company plans, as they at least state when they may slow down on stuff.
Indeed, if I condition on a post-Trump Republican party’s AI policy being anywhere similar to this AI action plan, I’d substantially update to Eliezer’s and Nate’s and Rob Bensinger’s view about the best path being pivotal acts (because we are not getting a slowdown), and more generally if this becomes a standard, I’d favor plans that were more uncooperative in AI safety, because cooperation isn’t getting us anything much of value.
And the most important action for a Democratic president after Trump is to completely repeal and replace this AI action plan, almost regardless of the level of partisanship (for example if there’s a part of the plan that is much more AI capabilities enhancing than safety enhancing, and Republicans threaten or might oppose AI regulations in the future if the part is removed, remove it anyways and let others deal with the partisan fallout later).
While there is a problem with the rhetoric, I’d also argue that the proposals are also very bad, as it only has a few elements about AI safety like an AI incidents committed, and while it’s much better than expected, in that they actually included any safety parts, but the portions about making the FCC ban state AI regulations and the parts about AI capabilities being increased increases make me strongly worry they will not slow down if necessary, and at some point, it’s likely we will have to slow down to implement AI control and alignment.
It’s only better than expectations in that anything about AI safety is included at all, it’s still the worst plan released on AI safety, and that includes company plans, as they at least state when they may slow down on stuff.
Indeed, if I condition on a post-Trump Republican party’s AI policy being anywhere similar to this AI action plan, I’d substantially update to Eliezer’s and Nate’s and Rob Bensinger’s view about the best path being pivotal acts (because we are not getting a slowdown), and more generally if this becomes a standard, I’d favor plans that were more uncooperative in AI safety, because cooperation isn’t getting us anything much of value.
And the most important action for a Democratic president after Trump is to completely repeal and replace this AI action plan, almost regardless of the level of partisanship (for example if there’s a part of the plan that is much more AI capabilities enhancing than safety enhancing, and Republicans threaten or might oppose AI regulations in the future if the part is removed, remove it anyways and let others deal with the partisan fallout later).