AI safety pipeline
(re: https://80000hours.org/2016/12/has-80000-hours-justified-its-costs/#technical-ai-safety-research-pipeline-of-50-people)
Hi Rob. I think this is very impressive, and also something that's largely been neglected so far (Matthew Gentzel was planning to do it on a large scale, but it didn't work out AFAIK). However, I think the program would probably be more effective if there were better ways for researchers to coordinate once they'd been recruited. Eg., about six months ago, MIRI and Google each put out machine learning safety research agendas, but it's not clear how people interested in those problems would talk to each other (apart from low-discoverability, very-high-latency journal papers). I asked some people a few months ago:
"I liked reading the machine learning safety research agendas put out by Google and MIRI a few months ago, but am not sure where I'd go to read about proposed solutions, or where I'd post ideas if I had one myself (short of publishing a full paper). Is there a forum/meetup/mailing list/etc. where people can flesh the basic proposals out in more detail?"
The only suggestion was MIRI's IAFF, which is high-quality but low-traffic and mostly dedicated to more abstract topics (and which also might be problematic for those with large disagreements with MIRI).
FLI has sort of been trying to fill this role with occasional conferences, but they have also been trying to act as a PR/outreach group, and I think the two aims are incompatible - research discussion requires floating low-confidence, likely controversial ideas, while PR requires strict filtering and message control.