The current cohort of the ML Alignment & Theory Scholars Program, MATS 6.0, had a unique application process and its broadest selection of mentors yet, with 40 mentors to apply to. I was invited to interview with twelve mentors and was accepted by five (which I later learned was an unusual number of interviews and offers in this cohort). Along the way, I challenged preconceptions I had about what AI safety research mentors look for in candidates, while becoming more familiar with various research areas.
In describing my experience with the application process, I hope this post is useful for anyone interested in applying to similar AI safety research programs or those involved in candidate evaluation. (Ultimately, I’ve decided to pursue another opportunity instead of MATS.) Note that I’ve written this relatively quickly and my experience may not be reflective of others.
Some updates I had from the interview process:
Brainstorming good research ideas is a major part of MATS interviews.
AI safety research experience matters a fair bit, especially experience relevant to a mentor’s research interests.
When I was rejected from prior rounds of MATS, I wondered whether I should have spent more time applying to more mentors. But a greater bottleneck was probably having more exciting AI safety research experience in my application.
Written answers to the mentor selection questions are less determinative than I previously thought of whether you get accepted.
It was pretty normal to receive a rejection email following an interview even when I felt I did reasonably well, given that each mentor could only accept a handful of candidates. After a while, I started to approach interviews with more equanimity.
I’ll divide the MATS application process into four phases, which I describe in more detail:
Initial intake
Mentor selection questions (with three waves of mentors)
Follow-up questions (optional)
Interviews (not all mentors)
Note that other mentors have a different process than what I experienced; for example, Neel Nanda commented:
For the avoidance of confusion, my MATS stream has a very different admissions process, that is heavily based on a work task and doesn’t have interviews (and weights quite different things), see more details here: https://tinyurl.com/neel-mats-app
Application process
Initial intake and mentor selection questions
For those who aren’t familiar, the MATS Summer 2024 application involved an initial Airtable form which asks for your resume and other background information. This was quick to fill out.
Then, weeks later, I got an email inviting me to complete mentor-specific questions on Airtable. The form asks you to choose which mentors to apply to and if applicable, answer written questions that they have. For example, part 1 had 18 mentors, 10 of which had questions. Some questions asked for a few sentences; others asked for several hundred words.
Questions topics included:
Describing your experience in machine learning or in research
Proposing follow-up research in response to prior papers or a general topic
Conceptual questions related to AI alignment
Implementing an experiment in a Colab notebook
Of the mentors that had mentor selection questions, I ended up writing responses to 13 mentors (not including two mentors that asked follow-ups), and I received interviews from three of these. Interestingly, the questions may have technically been optional, since I ended up interviewing and accepting an offer from a mentor I didn’t answer questions for (because of lack of time).
Follow-up questions
For a few mentors, I received an email asking me to complete additional written questions and/or a CodeSignal coding assessment or timed assessment in Google Colab. I had already done the CodeSignal assessment recently as part of applying to Anthropic, so I submitted my previous score and didn’t have to spend time retaking it.
Interviews
Over the course of a few days, I was invited to schedule an interview with individual mentors within the next week. Interviews were usually 15 to 30 minutes, though I did have one that was 75 minutes.
Interview invitation emails sometimes included information like:
The types of topics they wanted to discuss during the interview
Whether there were resources they recommended reading ahead of time
How many people they were interviewing and how many people they could accept (e.g., interviewing 9 people out of 200 candidates and accepting 1)
During the interviews, we discussed some of the following things, roughly sorted from most to least common:
Research ideas for a specific question they’re planning to research (e.g., related to deception, honesty, robustness). I was often asked follow-up questions to make these ideas more concrete.
My career plans
Logistics (e.g., “are you interested in the MATS extension program?”)
Questions I had about the research project
Technical machine learning questions
My prior experience
Interviews varied among mentors; some focused more on brainstorming research ideas, while others didn’t ask me questions and were just offering time to answer my questions.
Unlike other technical interviews I’ve had before, in MATS I was not asked questions like:
Behavioral questions (e.g., “Why are you interested in my stream?” “Tell me about a time when you overcame a challenge.”)
Mathematical questions (e.g., “What’s the formula for KL divergence?”)
In retrospect, I think I should have focused my interview prep more on brainstorming good research ideas and follow-up proposals. I probably spent too much time carefully reading each mentor’s past papers and trying to understand the details, as if I might be quizzed on specifics. In some cases, the interviews surprised me by how a mentor’s current research interests were unrelated to their existing work. That said, reading their past work is a good thing to do, and I think it was helpful to read up on (for example) adversarial robustness and cloud compute governance.
Mentors seemed most interested in two particular projects on my resume: the WMDP benchmark for unlearning, and working on general autonomy evals at METR. I wasn’t asked much about various other things like my software engineering experience or more conceptual work like a survey paper on AI deception. One update I made is that machine learning research experience relevant to the mentor’s research interests seems to matter much more than other types of experience.
Once MATS mentors are done interviewing candidates, they send the MATS team a ranking of candidates they want to accept. Then, the MATS team invites the top-ranked candidates based on the number of available slots, and then sends out additional invitations if earlier candidates decline the offer.
Time spent
I generally like to track my time using Toggl Track. In total, I tracked 31 hours on the application process for MATS 6.0. (Most people probably weren’t spending this much time on MATS applications, as I had an atypical number of interviews.) Having nine interviews during the week of May 12 was a little intense, but I eventually got the hang of it and stopped worrying as much about individual interviews. I would have appreciated it if the interviews could be spread out instead of having to be scheduled largely within one week, but I think that logistically wasn’t feasible for the MATS team this time. Here’s a breakdown of the time I spent applying to MATS:
April 28: 2 hours on mentor selection questions (part 1), where I wrote answers for four mentors
Week of May 5: 13 hours
2 or 3 hours on mentor selection questions (part 2), where I wrote answers for four mentors
9 hours on follow-up questions from three mentors and interview prep
Week of May 12: 14 hours
12 hours across nine interviews and related interview prep
2 hours on mentor selection questions (part 3), where I wrote I original answers for ~three mentors
Week of May 19: 3 hours across two additional interviews and related interview prep
Parting thoughts
If I were to give my past self some advice for applying to MATS, here’s what I might say:
Focus on developing strong machine learning research experience in areas that I would want to continue working on in the future.
Spend more time thinking about what good follow-up work to research might look like.
Thanks to Dawn Lu, George Ingebretsen, and Ryan Kidd for helpful comments and feedback. All mistakes are my own.
My experience applying to MATS 6.0
The current cohort of the ML Alignment & Theory Scholars Program, MATS 6.0, had a unique application process and its broadest selection of mentors yet, with 40 mentors to apply to. I was invited to interview with twelve mentors and was accepted by five (which I later learned was an unusual number of interviews and offers in this cohort). Along the way, I challenged preconceptions I had about what AI safety research mentors look for in candidates, while becoming more familiar with various research areas.
In describing my experience with the application process, I hope this post is useful for anyone interested in applying to similar AI safety research programs or those involved in candidate evaluation. (Ultimately, I’ve decided to pursue another opportunity instead of MATS.) Note that I’ve written this relatively quickly and my experience may not be reflective of others.
Some updates I had from the interview process:
Brainstorming good research ideas is a major part of MATS interviews.
AI safety research experience matters a fair bit, especially experience relevant to a mentor’s research interests.
When I was rejected from prior rounds of MATS, I wondered whether I should have spent more time applying to more mentors. But a greater bottleneck was probably having more exciting AI safety research experience in my application.
Written answers to the mentor selection questions are less determinative than I previously thought of whether you get accepted.
It was pretty normal to receive a rejection email following an interview even when I felt I did reasonably well, given that each mentor could only accept a handful of candidates. After a while, I started to approach interviews with more equanimity.
I’ll divide the MATS application process into four phases, which I describe in more detail:
Initial intake
Mentor selection questions (with three waves of mentors)
Follow-up questions (optional)
Interviews (not all mentors)
Note that other mentors have a different process than what I experienced; for example, Neel Nanda commented:
Application process
Initial intake and mentor selection questions
For those who aren’t familiar, the MATS Summer 2024 application involved an initial Airtable form which asks for your resume and other background information. This was quick to fill out.
Then, weeks later, I got an email inviting me to complete mentor-specific questions on Airtable. The form asks you to choose which mentors to apply to and if applicable, answer written questions that they have. For example, part 1 had 18 mentors, 10 of which had questions. Some questions asked for a few sentences; others asked for several hundred words.
Questions topics included:
Describing your experience in machine learning or in research
Proposing follow-up research in response to prior papers or a general topic
Conceptual questions related to AI alignment
Implementing an experiment in a Colab notebook
Of the mentors that had mentor selection questions, I ended up writing responses to 13 mentors (not including two mentors that asked follow-ups), and I received interviews from three of these. Interestingly, the questions may have technically been optional, since I ended up interviewing and accepting an offer from a mentor I didn’t answer questions for (because of lack of time).
Follow-up questions
For a few mentors, I received an email asking me to complete additional written questions and/or a CodeSignal coding assessment or timed assessment in Google Colab. I had already done the CodeSignal assessment recently as part of applying to Anthropic, so I submitted my previous score and didn’t have to spend time retaking it.
Interviews
Over the course of a few days, I was invited to schedule an interview with individual mentors within the next week. Interviews were usually 15 to 30 minutes, though I did have one that was 75 minutes.
Interview invitation emails sometimes included information like:
The types of topics they wanted to discuss during the interview
Whether there were resources they recommended reading ahead of time
How many people they were interviewing and how many people they could accept (e.g., interviewing 9 people out of 200 candidates and accepting 1)
During the interviews, we discussed some of the following things, roughly sorted from most to least common:
Research ideas for a specific question they’re planning to research (e.g., related to deception, honesty, robustness). I was often asked follow-up questions to make these ideas more concrete.
My career plans
Logistics (e.g., “are you interested in the MATS extension program?”)
Questions I had about the research project
Technical machine learning questions
My prior experience
Interviews varied among mentors; some focused more on brainstorming research ideas, while others didn’t ask me questions and were just offering time to answer my questions.
Unlike other technical interviews I’ve had before, in MATS I was not asked questions like:
Behavioral questions (e.g., “Why are you interested in my stream?” “Tell me about a time when you overcame a challenge.”)
Mathematical questions (e.g., “What’s the formula for KL divergence?”)
In retrospect, I think I should have focused my interview prep more on brainstorming good research ideas and follow-up proposals. I probably spent too much time carefully reading each mentor’s past papers and trying to understand the details, as if I might be quizzed on specifics. In some cases, the interviews surprised me by how a mentor’s current research interests were unrelated to their existing work. That said, reading their past work is a good thing to do, and I think it was helpful to read up on (for example) adversarial robustness and cloud compute governance.
Mentors seemed most interested in two particular projects on my resume: the WMDP benchmark for unlearning, and working on general autonomy evals at METR. I wasn’t asked much about various other things like my software engineering experience or more conceptual work like a survey paper on AI deception. One update I made is that machine learning research experience relevant to the mentor’s research interests seems to matter much more than other types of experience.
Once MATS mentors are done interviewing candidates, they send the MATS team a ranking of candidates they want to accept. Then, the MATS team invites the top-ranked candidates based on the number of available slots, and then sends out additional invitations if earlier candidates decline the offer.
Time spent
I generally like to track my time using Toggl Track. In total, I tracked 31 hours on the application process for MATS 6.0. (Most people probably weren’t spending this much time on MATS applications, as I had an atypical number of interviews.) Having nine interviews during the week of May 12 was a little intense, but I eventually got the hang of it and stopped worrying as much about individual interviews. I would have appreciated it if the interviews could be spread out instead of having to be scheduled largely within one week, but I think that logistically wasn’t feasible for the MATS team this time. Here’s a breakdown of the time I spent applying to MATS:
April 28: 2 hours on mentor selection questions (part 1), where I wrote answers for four mentors
Week of May 5: 13 hours
2 or 3 hours on mentor selection questions (part 2), where I wrote answers for four mentors
9 hours on follow-up questions from three mentors and interview prep
Week of May 12: 14 hours
12 hours across nine interviews and related interview prep
2 hours on mentor selection questions (part 3), where I wrote I original answers for ~three mentors
Week of May 19: 3 hours across two additional interviews and related interview prep
Parting thoughts
If I were to give my past self some advice for applying to MATS, here’s what I might say:
Focus on developing strong machine learning research experience in areas that I would want to continue working on in the future.
Spend more time thinking about what good follow-up work to research might look like.
Thanks to Dawn Lu, George Ingebretsen, and Ryan Kidd for helpful comments and feedback. All mistakes are my own.