For most AI Safety groups, the key activity by which newcomers familiarize themselves with AI Safety and the group is through a reading group (also sometimes called a fellowship).

Reading groups are usually groups of 4–6 people, alongside a facilitator or host, where people go through a set of readings (often papers or articles on a topic) and then discuss them together, usually once a week, and often for around 7 or 8 weeks. If you’ve taken the AI Safety Fundamentals course, then you’re already familiar with the format and structure of these.

Choosing the readings

A key step before starting a reading group is choosing the readings themselves. For advanced members of your group, it’s not uncommon to define the readings in collaboration with the students themselves, but for introductory reading groups, we strongly recommend going with a predefined curriculum, and picking a facilitator that is already familiar with its contents.

It’s common for these to be based on one of two curricula, both lasting 7 weeks for its reading component:

• BlueDot’s AI Safety Fundamentals, Alignment Track
  • This is the standard curriculum to introduce people to AI Safety, and is mainly focused on technical aspects of AI Safety, while still going through some aspects of policy and governance efforts.
  • The curriculum was designed to be friendly to people with or without a strong technical background, but we nonetheless think people with such a background can usually take the most advantage of it.
  • BlueDot maintains detailed session plans for this curriculum.
• BlueDot’s AI Safety Fundamentals, Governance Track
  • This is an alternate curriculum that is focused on AI policy and governance, and is better suited for people with some pre-existing interest in policy efforts. Many groups choose to run this alongside the alignment curriculum, especially if the group is “hybrid” in nature (appealing to both technical and policy leaning people).
  • BlueDot also maintains session plans for this curriculum.
• You can also consider using some of the following curriculums as well
  • CAIS’s AI Safety Society and Ethics Textbook
  • CeSIA’s AI Safety Atlas Textbook

Or one of the numerous “forks” of these curricula, which typically adapt the readings to fit better to a student group context. For example, a popular adaption is maintained by Harvard’s AI Safety Student Team (AISST), which removes readings to make it friendlier for doing the readings in-person, rather than before each meeting. AISST also maintains a helpful guide on running a fellowship similar to theirs. If you can do in-person readings, we recommend you do so, as people not doing the readings is a recurrent pain point for many introductory reading groups.

Regardless of which curricula you choose, if it’s based on work by BlueDot Impact, it’s important you follow their guidelines on running independent versions of their courses. The summary is:

• Brand your reading group in a way that clarifies who’s running the course. For example, don’t use BlueDot’s logo in a way that implies that they’re running it.
• Give appropriate credit to them for having authored the curriculum.