Columbia EA is starting an Advanced AI Safety Fellowship this summer, which will involve one 3-hour meeting per week with a group of people seriously considering careers in AI safety who have some background in both ML (at least working knowledge of linear algebra, multivariable calculus, probability and stats, and neural networks) and AI safety (usually the Cambridge AGI Safety Fundamentals program or a similar reading group we run at Columbia). Each meeting will involve reading about a topic selected the previous week and discussing it, and may also include exercises (like these). While we do not have capacity to invite people from elsewhere to join our meetings, we would like to invite others to follow along with our exploration of the field! You can find our running meeting notes here, where we will write down our meeting agendas, readings, and other notes (maybe key takeaways, responses to exercises, questions, etc.).

The hope is that this can serve as a small, easy win, by motivating at least a few people to set aside time to dive deep into AI alignment topics with little additional effort on our part. Participating in an alternative program with a similar premise, in which the organizers put more effort into making it a good experience for you, would likely be better than following along with us. If such a program already exists, please let me know! I’m not currently aware of any programs that satisfy all of the following properties:

1. Have the basic premise of being a deep dive into AI safety for people with some background in ML and AI safety who are seriously considering careers in the field
2. Are open to anyone with that background
3. Have organizers dedicated to making this a good experience for many interested people from various places

Given that we are not satisfying 3, you could dispense with “following along” and just structure your entire plan yourself if you feel that would work better for you. There are two main benefits we hope to provide for people following along:

1. Reducing the amount of time you spend selecting readings / exercises.
2. Providing an external anchor that you are motivated to keep up with.

We don’t have any strong takes on what it would look like to follow along with us, but here is a suggested path:

1. Comment on this post that you are interested in following along - if others have also expressed interest, consider forming a group with them so that you can have partners for accountability and discussion.
   1. Even if you are not looking for a group, I (Rohan Subramani) would like to have a sense of how many people are following along, so a comment expressing interest (or any other way of letting me know, like via Facebook Messenger) would be appreciated.
2. Look at the notes doc (at least) once per week. We are currently planning to meet on Sunday mornings ET, so most notes for each week should be completed by Sunday afternoons ET.
   1. See what we’ve been reading. Read whichever parts of that seem interesting to you, and supplement with other AI safety papers / blog posts that you want to read. To match our plan, spend 3 hours each week on a balanced combination of reading, thinking, taking notes, summarizing readings, carrying out exercises, and discussing the content (if you have discussion partners).

Thanks to Berkan Ottlik for co-running the fellowship and providing suggestions for this post. All mistakes are my own.