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.

AP
drop in any time
formal process
AP
anyone can join
clearly interested in AI safety

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:

  • Some people are intimidated by the idea of an application process and might not apply
    • A current organizer who did fundamentals in Fall 2026 mentioned that his friends thought of club applications like these exclusive business clubs
  • More people will go to our fundamentals course if they try it first
    • They will tell their friends to later apply after liking it
  • We might be bad at predicting who will do well at our intro course
  • We might be able to persuade people who would otherwise be working on normal ML research into working on AI safety

Arguments for having an application/being selective:

  • Making an application would make people more committed
  • More prestige for the program
  • Applications might filter for people who are more into AI safety
    • People who are good at writing applications will fare well later on, as much of AI safety involves applying to things
    • We do not want people to treat us as a generic ML upskilling place
  • Our impact and our funder's theory of change likely comes from a few exceptional people rather than a lot of people

Alternatives we are considering to our current structure:

  • Make our foundations course open to anyone for the first 2 weeks then require application
  • Frame it as an interest form rather than an application
  • Accept more applicants and run more cohorts

We would appreciate your thoughts on our selection process, or any experiences and thoughts you had with your existing group.

4

0
0

Reactions

0
0
Comments4
Sorted by Click to highlight new comments since:

You're probably right that your impact and funders' theory of change comes from the most capable and engaged people, but those people are often also busy or have many interests. An underrated cost of applications is just that they add friction, which selects for people who have time and motivation to fill out applications. You mention "intimidation" but it also just takes time and activation energy to fill out an application without a clear payoff.

The main costs or risks from reducing friction are if the group becomes so large and low engagement as to distract from the intended content.

(I started an EA group at UIUC when I was a student 5 years ago. On summer logic I didn't run a "fellowship" structure at all, against the advice of CEA, and instead spent their money on food to run weekly meetings themed around brief slides and discussion aligned to attendee interest. At least two out of <20 regular attendees work full time on AI safety, though I'm not sure what the counterfactual was.)

My opinion is that AIS groups should:

  • Have an application form
    • to (1) filter and (2) get people to commit early (as opposed to an interest form, where the applicant doesn't need to conciously commit). 1 gets you better fellows, 2 gets you more-committed fellows
  • Make it short.
    • You should test it out yourself (or better yet, get a friend to do it). Make sure it's under 10 minutes (preferably 5).
  • Make it informative, even if it's short.

    • Mainly, you want people to just write a few short sentences about the questions you think matter. Restrict the response to just a few sentences, and you'll be able to quickly tell certain qualities without taking much of their time (or yours!)
    • "Short yet extracts useful info" is best explained by an example IMO; see Penn's application. Also see real examples of short-yet-good-and-informative responses to "What currently most concerns you about AI?" below
      • Good response that you can tell is good even though it's just a few sentences #1: "I'm most concerned with the gap between how fast AI is growing and how slowly the regulatory frameworks are being developed. It interests me that there is no government agency or sector dedicated to regulating AI, and that's something I'm very open to learning more about and discussing further."
      • Good response #2: "I think what most concerns me about AI is the existential risk, not necessarily of rogue AI but governments using it to fight each other and that leading to death and destruction."
      • Good response #3: "At the moment, I am most concerned about the application of AI in the natural sciences. I believe that AI can accelerate the pace of discoveries but can also lead to more mal-intended use and exploitation of AI to create more existential risks such as engineered pandemics or even physical and chemical weapons."
    • You'll notice that response 1 isn't familiar with existential rilsk but they seem to (1) geniunely hold a belief in the/a risk of AI, and (2) that risk seems very relevant to the fellowship. That's all you need! And a very short application gets you that.
      • Most bad responses are always just people mascrading or trying-too-hard at explaining why they care. When people ask me, "how do I get into the fellowship!" I tell them, "just write few simple sentences about why you care; let me ask you now: why do you care?" and whatever answer I say, "Exactly! Just write that. Write what you think, like literally your inner-voice, word-for-word, give or take or whatever." Some will make up reasons for why they care, and you can spot insincerity from just a few sentences. Others have concerns that are not super-relevant, and then it's your choice if you have capacity to include people that the fellowship isn't exactly for.

    (Also, if you have thoughts on structuring short applications, I'd love to hear them!)

Jamie Huang
2
1
0
60% disagree

We should be selective with our intro course

I think that if someone who isn't initially interested in AI safety joins our course, maybe they will become interested through it. Unlike research I don't think our foundational course is that helpful for people going into typical ML.

Jamie Huang
1
0
0
50% disagree ➔ 0% agree

We should have an application-like process

I'm personally of the opinion that we should rename our application into an interest form.  I also think that writing out what they want to get out of their program will make them more committed to finishing it. However, I also don't want to scare the newbies away

Curated and popular this week
Relevant opportunities