As someone with an academic background now working in mainstream ML research, I strongly endorse the message of this post.
I agree with many on this forum that (a) there is some extra work in writing an article in an "academic" style, and (b) academic articles are often written with the objective of impressing a reviewer rather than being completely transparent about what was achieved in the work and how it advances the state of knowledge. I was incredibly frustrated by both of these issues when I was a grad student. However, when I try to read research written on LessWrong, it often seems even more difficult to understand. I've heard from others working in ML but not well-versed in safety that they have had similar experiences.
Another commenter complained about papers that "getting them into ML conferences is time-consuming." Definitely agree, and conference reviewing in ML is atrocious. But you can avoid all of that work just by putting your work on arXiv instead of aiming for an ML conference.
As someone with an academic background now working in mainstream ML research, I strongly endorse the message of this post.
I agree with many on this forum that (a) there is some extra work in writing an article in an "academic" style, and (b) academic articles are often written with the objective of impressing a reviewer rather than being completely transparent about what was achieved in the work and how it advances the state of knowledge. I was incredibly frustrated by both of these issues when I was a grad student. However, when I try to read research written on LessWrong, it often seems even more difficult to understand. I've heard from others working in ML but not well-versed in safety that they have had similar experiences.
Another commenter complained about papers that "getting them into ML conferences is time-consuming." Definitely agree, and conference reviewing in ML is atrocious. But you can avoid all of that work just by putting your work on arXiv instead of aiming for an ML conference.