As far as I understand, Eliezer is abrasive for these reasons:
If a misaligned ASI is created without an aligned counterpart, the ASI will have no reason to keep us alive.
Mankind isn’t on track to solve alignment BEFORE creating the ASI.
Normally, THAT high stakes would prompt anyone to think twice of anything related to the issue. But Eliezer has seen things like many obviously stupid approaches to alignment[1], a high-level official at GDM building a model which predicts that mankind will be left with jobs[2] after the ASI’s rise or OpenPhil (seemingly?) basing its position about AGI timelines on the approach of Ajeya Cotra[3]
As evidenced by him claiming that an approach is “Not obviously stupid on a very quick skim” and congratulating the author on eliciting a THAT positive review. Alas, I also have seen obviously stupid alignment-related ideas make their way at least to LessWrong.
However, it would be possible if the ASIs required OOMs more resources per token than humans. In this case applying the ASIs would be too expensive. Alas, this is unlikely.
IMO Eliezer also believes that the entire approach is totally useless. However, a case against this idea is found in comments mentioning Kokotajlo (e.g. mine)
As far as I understand, Eliezer is abrasive for these reasons:
If a misaligned ASI is created without an aligned counterpart, the ASI will have no reason to keep us alive.
Mankind isn’t on track to solve alignment BEFORE creating the ASI.
Normally, THAT high stakes would prompt anyone to think twice of anything related to the issue. But Eliezer has seen things like many obviously stupid approaches to alignment[1], a high-level official at GDM building a model which predicts that mankind will be left with jobs[2] after the ASI’s rise or OpenPhil (seemingly?) basing its position about AGI timelines on the approach of Ajeya Cotra[3]
As evidenced by him claiming that an approach is “Not obviously stupid on a very quick skim” and congratulating the author on eliciting a THAT positive review. Alas, I also have seen obviously stupid alignment-related ideas make their way at least to LessWrong.
However, it would be possible if the ASIs required OOMs more resources per token than humans. In this case applying the ASIs would be too expensive. Alas, this is unlikely.
IMO Eliezer also believes that the entire approach is totally useless. However, a case against this idea is found in comments mentioning Kokotajlo (e.g. mine)