Many are saying…
Installing brakes is a no brainer (figure out how to do a verifiable slowdown if needed), which is different from shooting bullets at random parts of the engine (banning datacenters in different states)
Users praise the brakes vs random bullets metaphor for making the distinction in debates over verifiable AI progress brakes clearer than most discussions.
Re: https://www.anthropic.com/institute/recursive-self-improvement
Many are saying…
Installing brakes is a no brainer (figure out how to do a verifiable slowdown if needed), which is different from shooting bullets at random parts of the engine (banning datacenters in different states)
FWIW I am personally not that RSI pilled compared to some of y’all + think we’re more in need of verified agreements on safety/security mitigations than slowdown
But similar tools are needed either way + we should be doing more to prepare for various scenarios + I could be wrong
I will note btw that it is not as if this is a new concept but good for them to talk about it.
Many have been trying to figure out verification for AI for years (see some of my related publications here https://milesbrundage.substack.com/p/some-recent-stuff-i-wrote, not exhaustive!).
I will note btw that it is not as if this is a new concept but good for them to talk about it.
Many have been trying to figure out verification for AI for years (see some of my related publications here https://milesbrundage.substack.com/p/some-recent-stuff-i-wrote, not exhaustive!).
@Miles_Brundage i think brakes would be great if we can make sure there’s not an incentive to misuse them. extinction does probably seem worse than societal collapse i suspect but many possible worlds in between that make this a very hard problem.
@Miles_Brundage the surpassed by super intelligence and not brought along outcome seems far more clear than the stagnation and collapse one though
@Miles_Brundage brakes vs random bullets is a good way to put it
the metaphor makes the distinction way clearer than most of this debate