Good AI alignment online class?

post by Alexander (alexander-1) · 2021-10-11T00:48:19.529Z · LW · GW · 3 comments

This is a question post.


I am fascinated with AI alignment, and I am interested in diving deeper into the area. I am going through the recommended readings listed on MIRI Research Guide, but I wish there were structured online classes covering similar content. I enjoy reading textbooks and learning, but having a structure and a community of people with similar goals would improve my learning experience and my grasp of these powerful ideas.

My employer gives a yearly allowance for structured learning (not self-directed learning). I have wanted to take advantage of it and use it to do some paid online apprenticeship on AI, focusing on alignment. However, the closest thing I have found is these online masters degrees in CS from Georgia Tech and other universities. Given that I work full-time, doing a masters degree would be infeasible. Is there anything worthwhile and high quality that can address my need? Preferably with a low time commitment so I can realistically manage it alongside full-time work.

The reasons I am looking for an online course rather than an in-person course are the COVID-19 pandemic and the fact that I live in Australia. Hence, the options for studying AI alignment in-person are limited or non-existent.

Academic background: I graduated with an undergraduate degree in computer science two years ago.


answer by abukeki · 2021-10-11T20:24:21.238Z · LW(p) · GW(p)

A bunch of links here and here.

answer by Ozyrus · 2021-10-11T15:30:43.806Z · LW(p) · GW(p)

I could recommend Robert Miles channel. While not a course per se, it gives good info on a lot of AI safety aspects, as far as I can tell.


Comments sorted by top scores.

comment by Charlie Steiner · 2021-10-11T05:23:31.014Z · LW(p) · GW(p)

I'm not sure there's anything that's a perfect fit. There have been some attempts at AI alignment survey courses (I think MIT had one that looked promising spring of 2019, but I can't find it so maybe I'm hallucinating) but they seem to be aperiodic.

If I might recommend imperfect things, I figure some choices are to be pragmatic and learn about useful things like the infrastructure behind modern machine learning, or be more idealistic and learn about useful things like algorithmic information theory.