From math grad school to AI alignment

I get a lot of emails from folks with strong math backgrounds (mostly, PhD students in math at top schools) who are looking to transition to working on AI alignment / AI x-risk. There are now too many people “considering” transitioning into this field, and not enough people actually working in it, for me, or most of my colleagues at Stuart Russell’s Center for Human Compatible AI (CHAI), to offer personalized mentorship to everyone who contacts us with these qualifications.

But this is exactly why we need more people actually working in the field. It’s a chicken and egg problem, and the only way to solve it is to start incubating yourself. So, I’ve written this post about how math grad students specifically can work toward shifting their own researching more toward AI alignment and control.

Note that I’ve previously written about how folks who are seriously interested in AI alignment should try to get into Berkeley. However, PhD students in mathematics at other top schools might not want to uproot and transfer to Berkeley at the drop of a civilization. (You’ve gotta have a sense of humor about x-risk if you’re gonna think about it every day!). So here are some steps for getting closer to doing AI alignment research without having to physically move:

1. Get acquainted with basic arguments about AI and existential risk, by reading:
   1. (short article) Yes, We Are Worried About the Existential Risk of Artificial Intelligence by AI professor Stuart Russell and political scientist Allan Dafoe.
   2. (full book) Superintelligence: Paths, Dangers, Strategies by Oxford philosopher Nick Bostrom, the subject the of the article above.
2. Get acquainted with nascent technical work on AI control and alignment:
   1. Browse CHAI’s bibliography page, and
   2. Read 3-10 papers from it that catch your interest.
3. Develop relevant technical writing skills, either by
   1. publishing something (anything) in top AI/ML conference,
      ***or***
   2. develop your own independent thoughts for technical (mathematical/algorithmic) methods for controlling superintelligent AI, by:
      1. writing about them in a precise but humble/non-authoritative way so as to get feedback/criticism from others, and then
      2. sharing your best thoughts (after several rounds of critical review from thoughtful friends/colleagues) with someone at CHAI, FHI, MIRI.