carado.moe

tammy's blog about
AI alignment, utopia, anthropics, and more;

also on twitter, lesswrong, rss, and at

AI risk plans

people have criticized my peerless plan on the grounds that it's too long-term/far-fetched.

while i don't disagree, i think that it is only one variable to be taken into consideration. here is a comparison of plans for addressing AI risk, with vague estimates.

plan	achievable before X-line¹	chance of U-line²	S-risk²
doing nothing	100%	<1e-6	<1e-6
direct alignment	.1%	5% → .005%	5% → .005%
the peerless	2%	10% → .2%	1% → 0.02%

• ¹: assuming significant effort is put behind the plan in question, what is the likelyhood that we'll have accomplished the work to what we believe to be completion? note that my current AI timelines are pretty pessimistic (we become more likely to die than not this decade)
• ²: if we believe to have completed the work; the latter number is adjusted by being multiplied with "achievable before X-line".

note that the numbers i put here are only very vague estimates, feel free to replace them with your own guesses. but my point is, in order for the peerless to be the plan we should be working on, we don't need it to be feasible, we just need it to be less infeasible than all the other plans. i think the peerless is more tractable than doing direct alignment, and only more risky because it has more chances to succeed. depending on how scared of S-lines you are, you should push for either doing nothing (and thus oppose direct alignment) or for my plan. (or come up with your own, and then compare it to these!)

not pictured: the plan to melt all GPUs, because it's just a modifier on what we do afterwards. but yes, melting all GPUs is a great idea if we think we can reasonably do it more than other plans.