Agree with your soft max idea but “net positive EV” is too soft—as I’ve said elsewhere donating to the university you went to or your local animal shelter is still net positive EV.
I do agree net positive is too soft, but I don’t think this is what anyone is seriously advocating for here.
The main implicit theory of impact for event venues is venues → events → {ideas / intellectual progress / new people engaging with the ideas / sharing of ideas}
I think in domains like “thinking about future with powerful AI” or “how to improve humanity’s epistemics” or “what are some implications of digital minds existing” it seems the case that noticeable amount of thinking and discussion is happening at various in-person gatherings.
If you overall make the bet these domains are important (which seems softmax-reasonable), you are faced with a range of bets which differ on criteria such as how good the bet is, what’s the variance, and how scalable it is. To showcase some extremes - e.g., I think it’s an obviously great idea to pay Paul Christiano, allowing him to work on alignment. while this is way more effective than buying events venue, it’s not really scalable—if you send twice as much money to Paul, you won’t get twice the output. - it’s less clear what’s the sign and impact of ‘pay a lot of AI safety community builders’ : plausibly good, but high variance - I think venues are plausibly good softmax type of bet, with a lot of positive uncertainty, and not that much direct downside
The idea that EAs should take only actions which maximize EV according to some sort of straightforward calculation is wrong.
The argmax(EV(action)) is a stupid decision strategy and people should not be criticized for not following it.
Agree with your soft max idea but “net positive EV” is too soft—as I’ve said elsewhere donating to the university you went to or your local animal shelter is still net positive EV.
I do agree net positive is too soft, but I don’t think this is what anyone is seriously advocating for here.
The main implicit theory of impact for event venues is
venues → events → {ideas / intellectual progress / new people engaging with the ideas / sharing of ideas}
I think in domains like “thinking about future with powerful AI” or “how to improve humanity’s epistemics” or “what are some implications of digital minds existing” it seems the case that noticeable amount of thinking and discussion is happening at various in-person gatherings.
If you overall make the bet these domains are important (which seems softmax-reasonable), you are faced with a range of bets which differ on criteria such as how good the bet is, what’s the variance, and how scalable it is. To showcase some extremes
- e.g., I think it’s an obviously great idea to pay Paul Christiano, allowing him to work on alignment. while this is way more effective than buying events venue, it’s not really scalable—if you send twice as much money to Paul, you won’t get twice the output.
- it’s less clear what’s the sign and impact of ‘pay a lot of AI safety community builders’ : plausibly good, but high variance
- I think venues are plausibly good softmax type of bet, with a lot of positive uncertainty, and not that much direct downside
It would be helpful to see your thoughts on community building fleshed out more in a post or longer comment.