MARS

MARS (Mentorship for Alignment Research Students) is a research program that pairs students with experienced mentors to work on an AI safety (technical or policy) research project for 2 to 3 months. There is an in-person kick-off week in Cambridge at the start of January, where students can work together, discuss research ideas with researchers and attend workshops. The program will start on 3 January and end on 3 March, with flexible start and end dates depending on the project. The general application (Phase I) deadline is 30 November at midnight UTC.

Projects

  • Representation Engineering

    Andy Zou (Center for AI Safety)

    Read More

  • High-Level Interpretability

    Arun Jose (Independent)

    Read More

  • Mechanistic/Developmental Interpretability

    Evan Ryan Gunter (MATS)

    Read More

  • Evaluating LM Agency

    Francis Rhys Ward (Imperial College London)

    Read More

  • Guarantees-driven Mechanistic Interpretability

    Jason Gross (Alignment Research Center)

    Read More

Program Details

The general applications (Phase I) will close on November 30 at midnight. Upon completion, you will receive a form to answer any mentor-specific selection questions and/or a coding test (Phase II) if they were required by the mentors you apply to (please check the project details). The mentor-specific application is due by December 5 at midnight. We will inform all applicants of the final decision outcome by December 12.

Applications are reviewed on a rolling basis, so earlier applications have a higher chance of being accepted. (More projects may be added, including AI governance/policy projects.)

  • Time Commitment

    Students should expect to each spend 6-15+ hours/week on their project, depending on the project chosen. (Check out the Projects for more info.)

    Weekly Check-Ins

    Throughout the duration of the program, you are expected to check-in:

    • With your teammates once a week. It'll be up to your team to coordinate meeting times or create coworking systems.

    • With your supervisor once a week/every two weeks

    The format of check-ins with your supervisor will be to their discretion.

    We will also have program-wide events and check-ins (during week 4-5 of each term). These will be announced later.

    MARS Research Symposium

    This event will take place during the start of Easter break, where groups can present their research. More information will be communicated later.

    Norms

    We expect transparent expectation-setting and communication between participants and mentors. Failing to do so may mean that we will not consider you for future opportunities.

    This message isn’t meant to scare you. CAISH and the mentors are providing this opportunity for you, and we want you to be respectful of mentors’ time. If you are ready to take on this commitment, we’re still excited for you to apply!

  • Each project has its own prerequisites—look through individual projects to find ones that match your experience and interests. Some researchers encourage applicants to apply, even without all of the prerequisites.

    We generally expect participants to have a strong foundation in programming and machine learning. Note that some projects may require additional background knowledge related to AI Safety/Alignment. Depending on the project, no prior research experience may be necessary, as long as you meet the specific technical requirements.

  • Virtual mentorship is possible, but we may give preference to those who can come to the in-person kick-off week in Cambridge.

    Mentorship time

    Mentors are free to choose when and how they meet with mentees. We expect mentors to commit ~10 hours to the project in the kick-off week, and weekly or biweekly check-ins with teams afterwards.

    Project

    The expected mentee is university students working part-time on this project. They will vary in background and research experience. The ideal project should allow mentees to reach meaningful milestones in 2-3 months of part-time work.

FAQ

  • This program is primarily for students based in Cambridge, but we welcome applicants from UK or elsewhere if they can come to Cambridge in January.

    We might support a remote option to participate in the program - if you can only participate remotely, you can apply to the remote stream in the application (we will notify you about this possibility at a later time).

    We expect graduates and promising undergraduates to be good fits, but are excited about applicants with a wide range of backgrounds. We expect to host ~25 mentees to Cambridge.

Email chloe@cambridgeaisafety.org or gabor@cambridgeaisafety.org if you have questions that are not answered.