I guess I would have liked more discussion on how this impacted Deepmind’s views on the topic. Could we get a discussion started on that in the comment section? I’ve added some thoughts to the article on the topic, but I haven’t sat down and done explicit calculations for how I should update yet, or how this ties into stuff like Johnwentworth’s research agenda, or how much more likely we are to be in Paul-verse or even if Paul and Eliezer would agree that this result favours Paul-verse.
Tons of small improvements can make big changes to performance, in a way that I did not think they would be able to. Or at least, not so soon. If there was just one change they made and it resulted in the same performance increase, I’d say that is way more Eliezer verse.
Paul seems down with progress being potentially quite fast. I’d be willing to bet he thinks a takeoff in 15 years is on the cusp of plausibility. I agree that the speed is maybe more Eliezer verse, but how it happened is more Paulverse. So I’d guess l(AlphaCode|Paulverse)>l(AlphaCode|Eliezerverse). But I don’t understand Eliezerverse enough to be confident about the numerical value of the likelihoods.
I guess I would have liked more discussion on how this impacted Deepmind’s views on the topic. Could we get a discussion started on that in the comment section? I’ve added some thoughts to the article on the topic, but I haven’t sat down and done explicit calculations for how I should update yet, or how this ties into stuff like Johnwentworth’s research agenda, or how much more likely we are to be in Paul-verse or even if Paul and Eliezer would agree that this result favours Paul-verse.
In which way does this news “favour Paul-verse”?
Tons of small improvements can make big changes to performance, in a way that I did not think they would be able to. Or at least, not so soon. If there was just one change they made and it resulted in the same performance increase, I’d say that is way more Eliezer verse.
Huh. Doesn’t this also push towards faster progress and thus faster takeoff though, and thus Eliezer-verse?
Paul seems down with progress being potentially quite fast. I’d be willing to bet he thinks a takeoff in 15 years is on the cusp of plausibility. I agree that the speed is maybe more Eliezer verse, but how it happened is more Paulverse. So I’d guess l(AlphaCode|Paulverse)>l(AlphaCode|Eliezerverse). But I don’t understand Eliezerverse enough to be confident about the numerical value of the likelihoods.
Idk. At this point I want to taboo “Paul-verse” and “Eliezer-verse” and use more specific, descriptive terms instead.