Overall, it was a very odd look into a discipline that I am familiar with, which leads me to think that the ideas are not very promising for any of the fields I don't have expertise in.
Whether AI is really more dangerous than, say, pandemics or asteroids, is left as an exercise for the reader.
The first question on their list is about the 'problem' of wild animal suffering - and I've personally seen EAs argue that, because some animals are carnivorous, nature should be destroyed.
That's not even the weirdest position EAs take. Look up Brian Tomasik. Specifically, his paper about the possibility that electrons might suffer.
Concern about superhuman AI is one thing; bullet-biting utilitarianism is another entirely.
(This isn't the only place where their philosophical framework is stuck in the British Empire; they also tend to take a teleological view of history and moral development, and believe that their views are the self-evident progression of ethical development that every culture and civilization will come to eventually. They may not be as bad about this now as they used to be - there are questions about China now - but I don't think they're quite to the point of coming to terms with cultural contingency yet.)
80k hours is more a cultural snapshot of the rationalist movement than anything.
Well, do Arden Koehler or Howie Lempel have 80,000 hours of CS Research experience? It looks like they have 80K hours of experience thinking about the best way to spend 80K hours. Woops.
On the contrary, the massive influx of applied research driven by deep learning hype means there's probably a lot more high-impact stuff outside of ML than inside of ML. Quantum computing, bullet-proof crypto implementations, safety engineering for robots, and even hardening the electric grid/key compute infra against solar flares all seem much more important. Also much less "sexy".
...and even most of the AI/ML problems are scraped from a few papers from corporate AI labs with the best PR skills, with an eye toward AGI risk. But 1) AGI risk is massively overblown and 2) trying to solve those problems in general has a "silver bullet" vibe. All of the solutions to "AI Safety" questions are going to boil down to "do good software engineering", which is often more about processes and org incentives than technology or science. A really good automated HiL testing setup is going to be a lot more important than "preventing reward hacking". If you don't believe me, go ask anyone who's actually built a robot.
So, I think maybe "AI Safety/Robustness broadly construed" could be one or two entries in a list of important problems in CS. But certainly not half the list as they are here.
TBH, one of the highest impact problems in CS is convincing young PhDs to not waste their time applying deep learning to every god damn thing regardless of whether it's the right tool for the job.