Research update: Accepted to CHAI internship!
Ram Rachum
Hi everyone,
After last month's breakthrough, I found myself exhausted from all the work. I decided to pause working on the core of the research, and instead I've been doing a lot of other things. Some of them have been personal things that I've been postponing, like doing my taxes and asset declaration for 2023. A couple of research-related things I've done is (a) improve my shell scripts I've used in the research and (b) get interviewed for both MATS and CHAI.
ML Alignment & Theory Scholars (MATS) is a 2-month research program for AI Safety researchers at Berkeley. CHAI is the Center for Human-Compatible AI at UC Berkeley, and it offers an internship program. The two programs are similar in that they're both in Berkeley, providing basic funding, mentorship and networking with other participants. They're different in a few ways, one of which is that CHAI is part of the University of California, Berkeley, whereas MATS is an independent program that's just located in the city of Berkeley.
Thankfully, their timelines don't overlap. MATS is between January and March while the CHAI internship starts in May and goes on for 3-4 months. This means that if I were to get accepted to both, I could participate in them one after the other.
Both MATS and CHAI sent me programming tests, the kind of tests that you do by yourself in a web interface, with a time limit of 90 minutes. The MATS one was straightforward, and similar to the kind of tests I've done when I was interviewing for Google, so I got a great score on it. The CHAI test was much more difficult.
CHAI sent me a choice of 3 possible programming tests I can take. I chose the one they said is the hardest. It's about Monte-Carlo Tree Search. I chose it because it's the closest to my interests. I prepared for the test for a week, implementing MCTS on a game of 3D tic-tac-toe. It's quite a difficult algorithm. After a week I finally started the test. It was a lot more challenging than I thought. My code wasn't that great and after 80 minutes, I assumed I've failed miserably. I submitted the code I've written and emailed CHAI, giving them polite feedback about the difficulty and withdrawing from the process. They suggested that I do one of the other tests, but I declined because I already feel bad enough. They accepted my withdrawal.
Three days later they sent me this email: "We have now finished grading the programming tests and actually your score on the MCTS test is considered passing. Also, there is at least one mentor who is interested in interviewing you. Would you be interested in proceeding with an interview?"
I said yes obviously. The mentor I was matched with is Cam Allen. Cam is doing a postdoc at CHAI and his research interests overlap with mine. We had a chat and hit it off. A couple of weeks later I got the email that I was accepted! This means I'll be working on my research with Cam, at Berkeley.
I'm very excited. It's 6 months away so there's lots of time to plan the quite complicated logistics. I'll stay in Berkeley for 3-4 months. I'll also need to sublet my apartment in Tel Aviv which would be a complex operation. If you're around Berkeley and want to meet up, let me know!
See you next month, Ram.