Please don’t share without permission.

This document describes how HAIST runs our intro fellowship; it’s meant to be useful both to HAIST and to other university AIS groups who want to run similar programming. It has all of our resources compiled, as well as general comments on strategy and tips & tricks. Good luck, and feel free to reach out to us ([email protected]) with questions! -Xander, Sam, and many others.

HAIST’s intro fellowship is a technical introduction to AI safety. We work hard to attract the best possible applicants, and are selective in order to improve the experience of those we accept. We only have in-person readings, and break into (intentionally formed) cohorts facilitated by people with experience in AIS research. Meetings are two hours, usually at dinner time (say, 6 – 8 pm). We alternate between reading and discussing (more in Facilitator Guide). Please read these two posts, which provide some initial context on HAIST and the intro fellowship. Note links to our outreach copy and our pubbing targets. You might also find the HAIST member handbook useful.

<aside> 📚 Content

To suite our in-person reading style (as well as some differences in taste), HAIST uses an adaptation of Blue Dot Impact’s AGI Safety Fundamentals (original by Richard Ngo, adapted by Sam Marks).

• Curriculum (in-person readings)
• Drive of readings formatted for printing (previous one here)
• Facilitator Guide (previous one here) </aside>

<aside> 📄 Forms & Systems

• Application. Make sure to ask for resume, and I’d default to copying this form exactly. We like using AirTable, which we also use to track attendance.
• Website Page (with FAQ)
• Attendance tracker [To Do]
• Feedback forms [To Do] </aside>

<aside> 🧑‍🏫 Facilitator Selection

Semester Timeframe: Week -2 – Week 2 Estimated Time: 5 hours, varies wildly.

Since the fellowship is targeted at highly talented technical students, facilitator quality is a key priority. We’ve so far only had facilitators who have experience conducting AIS research (though we recognize this is a luxury afforded to us by having lots of involved members).

“I'm not sure what the exact facilitator-quality-vs-value-added curve looks like, but my guess is that you should have fewer sections facilitated by a more elite group of facilitators than you'd think by default. The most important thing is that the facilitators have context in alignment, that they can explain the relationships between various parts of the curriculum and connections to current work in alignment (which it's good if they're relatively abreast of).” - Sam Marks

One way to help attract high-quality facilitators is to make their lives as easy as possible: make sure that their job is just to show up and facilitate a room full of carefully chosen students; try to avoid asking them to do prep work, printing readings, picking up food, booking rooms, etc. (See operations below.)

We currently pay all of our facilitators. We pay generously (reach out if you want the exact figure), which we think is reasonable since:

• Comparable jobs like tutoring often pay still more than our rate.
• Makes facilitators a bit more likely to accept the job, and to keep doing it in future semesters.
• Makes facilitators take the job a bit more seriously + get more into it.

See here for the copy we used to ask people if they’d like to facilitate (though note we already knew them, so it’s quite informal).

</aside>

<aside> 🔍 Outreach

Semester Timeframe: Week 0 - Week 1.5, ideally starting a bit before week 0. Estimated Time: 15 hours.

We’re very deliberate about our outreach process, and I think it’s one of the things HAIST does best. This is a key place to apply optimization pressure, and really try to get the best people to apply. Here are some key considerations:

• Our copy is technically interesting and exciting. We emphasize the importance of the problem, but don’t concentrate on altruism (note that HAIST is not affiliated with Harvard EA as a club). I think it’s also well written, with a nice level of formality. See here for our copy.
• We spend a while thinking about where are the best places to reach the most talented technical students, including lots of advanced classes (via emails and/or Slacks and/or discussion boards and/or asking prof if we can tell the class about it directly), CS/math/stats majors mailing lists, clubs (like HMMT or Women in CS), etc. See here for the list of places HAIST publicized to in Spring ’23, which should serve as useful inspiration. Note: I think this step is one of the key places to think like an optimizer, and try to spend some time generating a list of all the best places (mailing lists? signs in math centers? ask some professors? what’s that math/cs/econ/etc class full of very smart students?).
• We benefit a lot from friend networks, and encourage members to refer friends / help us with outreach.
• It’s important to get on this early in the semester; we do lots of outreach in the first week of classes, and sometimes do our outreach before other clubs (which I think is great).
• Note that we are asking people to apply for the fellowship directly, as opposed to asking them to fill in “interest forms” or get on our mailing list; I think this is very important, since you lose people in between successive emails. HAIST very rarely asks people to express general interest.
• HAIST & MAIA receive around 100 applicants each; see if you can beat us! </aside>

<aside> 🛠️ Operations

Semester Timeframe: Week 3 – Week 11 Estimated Time: 1.5 hr/week * no. sections

Room. At HAIST, we have an office, so we use that. Previously, we tried to book the nicest rooms we could on campus (and think paying attention to room selection was useful).

Printed (1) Core readings, (2) Further readings, and (3) Facilitation Guide. We always print readings, so that there are enough copies of the core readings for every participant, a reasonable number of further readings (enough for maybe half the number of participants for each reading), and one copy of the facilitation guide waiting in the room before the start of the meeting. Note that we don’t make facilitators do any of this work, and would encourage you to try to do the same.

Food. We always serve food at our meetings, which we think probably helps a bit with attendance & general enjoyment (for both participants and facilitators). We usually have a mix of vegetarian, vegan, and meat.

Prioritize Smooth Operations. Try hard to notice who is experiencing what annoyances/friction, and make sure you’re okay with that. At HAIST, we try really hard to make the experience very smooth for facilitators and participants, and wouldn’t make them do things like pick up the food or book the rooms.

</aside>

<aside> 🖇️ Processing Applications & Cohort Formation

Semester Timeframe: Week 2 – Week 2.5 Estimated Time: 15 hours

We spend lots of time and effort processing applications, in order to (1) review carefully the 100+ applications we receive, (2) think hard about how high our bar should be (we’ll likely decide it should be higher than our initial intuition) and who meets it, and (3) think about facilitator <> participant and participant <> participant matches (along with availability) to group participants into cohorts.

• This is a difficult constrained optimization problem, which consistently takes us a surprisingly long time to do well.
• We’ll often email participants who indicated problematic availability and pitch them on doing a specific time they haven’t indicated.
• We prioritize the people we are most excited about in (1) accommodating their availability, (2) grouping them with their friends, and (3) matching them to the facilitators we think are the best match.
• Useful copy here for acceptances / logistics. </aside>

<aside> 🏕️ Retreats

Semester Timeframe: [To Do]

Estimated Time: [To Do]

Useful copy here.

[To Do, add details]

</aside>

<aside> 🕺 **Extras

Slack.** We’ve made a HAIST/MAIA intro fellow Slack which we are using for communicating announcements, socials, etc. We hope that the Slack will be active / useful, but aren’t sure how this will go.

Socials. We run weekly HAIST socials. After a few (3?) weeks, we’ll specifically invite the intro fellows.

Member interactions. Last semester, we regretted not encouraging more member <> intro fellow interactions. We may try to setup a 1-1 matching system at some point during the semester.

Noticing awesome people. We’ve gotten value from trying to notice people who seem especially engaged and cool, and then working extra hard to make sure they are having a good time and getting value out of our programming. This has included pairing people to talk 1-1 to AI safety researchers. Some AI safety orgs would be interested in connecting with promising intro fellows. In particular, I’d be interested in hearing about particularly exciting intro fellows on behalf of Redwood Research; feel free to email me at [email protected].

Checking in with facilitators. We encourage regularly checking in with facilitators, and looking for ways to make things go better. This is also useful to make sure you are noticing awesome people (per above).

Being dynamic. Attrition is normal, if sad. Likely, some groups will drop to only a few people. It’s worth being dynamic in response to this, and merging groups / dissolving as needed. More generally, it’s important to monitor how things are going and try to adapt your plans to make things go as well as possible.

Monitoring vibe. Try to make sure that the vibe around the intro fellows is good. This can include things like making sure the room is well set-up and nice, and encouraging members to hang around the office during times when there is a nearby intro fellowship.

</aside>

<aside> 🤝 Next Steps

Winter / Summer Programming. CBAI and Redwood Research both ran winter programming that served as useful next-steps for intro fellows, especially CBAI’s ML bootcamps. We encourage intro fellows to apply by sending emails and (importantly!) specifically talking to people we thought might be good fits.

HAIST Membership. For the first time this semester, we asked prospective members to complete a long (~one day) membership application, which includes both technical and non-technical question. More information on membership is available in our handbook.

</aside>