Why would 30 years tell us anything? If we can’t build AGIs without risking a nuclear war how do we learn how they fail? You have to have an example of a failure happening before you can possibly design a fix.
And Eliezer says start a war instead of using the open agency model or some other reasonable modern swe approach? Stateless microservices are the primary reason some modern software is extremely reliable, even if it is spread across many computers. Google search being the premiere example but its also how spaceX avionics work.
Stateless microservice based AI could be extremely powerful—thousands of times more capable than current systems—and reliable and safe so long as no one is able to do the obvious and make them stateful...
I think world governments should demonstrate that they can cooperate and coordinate on a relatively simple problem like not building unaligned TAI, before they tackle the more difficult problem of coordinating to build aligned TAI.
If everyone suddenly got their act together and started coordinating effectively, I don’t think it would take 30 years before we could start building again, in a more controlled way. It might not even take 1! But a civilization should be able to pass easy tests, before it is allowed to attempt hard ones.
I dont know how informed you are about politics but most governments dont give a hoot about AI at the moment. There are smaller issues that are getting more attention and a few doomers yelling at them wont change a thing. On my model they dont care even at the moment the die. But I didnt predict the recent 6 month “ban” on training runs either so its possible this becomes a politically charged issue say by 2030 when we have gpt7 and its truly approaching dangerous levels of intelligence.
Why would 30 years tell us anything? If we can’t build AGIs without risking a nuclear war how do we learn how they fail? You have to have an example of a failure happening before you can possibly design a fix.
And Eliezer says start a war instead of using the open agency model or some other reasonable modern swe approach? Stateless microservices are the primary reason some modern software is extremely reliable, even if it is spread across many computers. Google search being the premiere example but its also how spaceX avionics work.
Stateless microservice based AI could be extremely powerful—thousands of times more capable than current systems—and reliable and safe so long as no one is able to do the obvious and make them stateful...
I think world governments should demonstrate that they can cooperate and coordinate on a relatively simple problem like not building unaligned TAI, before they tackle the more difficult problem of coordinating to build aligned TAI.
If everyone suddenly got their act together and started coordinating effectively, I don’t think it would take 30 years before we could start building again, in a more controlled way. It might not even take 1! But a civilization should be able to pass easy tests, before it is allowed to attempt hard ones.
Yeah but that’s never happened and won’t happen so long as you are talking about baseline humans.
I dont know how informed you are about politics but most governments dont give a hoot about AI at the moment. There are smaller issues that are getting more attention and a few doomers yelling at them wont change a thing. On my model they dont care even at the moment the die. But I didnt predict the recent 6 month “ban” on training runs either so its possible this becomes a politically charged issue say by 2030 when we have gpt7 and its truly approaching dangerous levels of intelligence.