Crossposted from the EA Forum.
We are AI Alignment @ Illinois. We got ~80 applicants for our fellowship program, with 40 accepted people. We are currently considering how selective we should be for our spring cohort. Since we added many new organizers, we would probably have the capacity to run more cohorts this time.
Here are some things we thought during our meeting (mentally add maybe in front of every bullet point):
Arguments for not having an application/not being selective:
Arguments for having an application/being selective:
Alternatives we are considering to our current structure:
We would appreciate your thoughts on our selection process, or any experiences and thoughts you had with your existing group.
The part about an interest form confuses me since I'm not sure what the alternative is. If the options are:
I definitely don't recommend (3), since getting people to make some sort of basic commitment is very low effort and helps with engagement. For deciding between (1) and (2), you should not cargo-cult practices from prestigious programs. They are prestigious because of what you get out of them, not because their applications are good. The main question for (1) vs (2) is whether you need to filter people (limited resources, planning, etc.).
For reference, Lens Academy and BlueDot both have reasonably low-effort applications, and only ~25% of people signed up for my Lens Academy cohort showed up to a majority of sessions. I'm taking a BlueDot course now, and attendence was so bad that they disbanded the cohort entirely and had the three of us who were actually showing up join different groups.
Make our foundations course open to anyone for the first 2 weeks then require application
If you can afford to this, you can probably get most of the signal by just counting attendance and accepting everyone who shows up.
We might be able to persuade people who would otherwise be working on normal ML research into working on AI safety
I think this is underrated, and it's kind of ridiculous that BlueDot's intro course on why you should care about AI safety rejects people for not caring about AI safety enough. Like sure, I can see why MATS doesn't want to train capabilities researchers, but what's the downside of accidentally teaching e/accs why AI safety is important?
Looking at your website, it makes sense that you have applications + filtering for the research fellowship. For your fundamentals class I would have a simple form that makes people check a box committing to doing the work (whatever that is), but it would be strange to ask any complex questions or filter people here.