Yong
Tips for Cracking the AI Safety Technical Interview
With regard to citations, it’s one thing to complain that a paper is only citing two out of three of the relevant pieces of prior work—but it’s another thing to complain that a paper seems blissfully unaware of an entire relevant body of prior work. This is especially problematic if the prior work persuasively establishes some limitations on or reasons to be skeptical of the author’s preferred data or methodology.
There are also a lot of papers that just cite without properly engage––some even mischaractererize––with the work. I notice that this problem is worsened with a lot of less-experienced researchers simply relying on LLMs for writing, especially on related work or discussion section.
Many times I went through some of the cited work, and found that they almost have identical findings or even entirely opposite findings. I usually make it a point in my own writings to write something like “The closest work to ours is by XXX, who found …”.
IF they cannot set the refusal classifier well for cyber and bio, what gives you confidence that they would classify “frontier LLM development” well? Not only that, you would need to second-guess the response you get. I’d rather they just do overrefusals for frontier LLM development questions (since they clearly don’t care about overrefusals).