Thanks! Yes, definitely in scope. There was a lot of discussion of this paper when it came out, and we had Raymond Douglas speak at a seminar.
Opinions vary within the team on how valuable it is to work on this; I believe Fin and Tom are pretty worried about this sort of scenario (I don’t know about others).. I feel a bit less convinced on the value of working on it (relative to other things), and I’ll just say why briefly: - I feel less convinced that people wouldn’t foresee the bad gradual disempowerment scenarios and act to stop them from happening, esp with advanced AI assistance - In the cases that feel more likely, I feel less convinced that gradual disempowerment is particularly bad (rather than just “alien”). - Insofar as there are bad outcomes here, it seems particularly hard to steer the course of history away from them.
The biggest upshot I see is that, the more you buy these sorts of scenarios, the more it increases the value of AGI being developed by a single e.g. multilateral project rather than being developed by multiply companies and countries. That’s something I’m really unsure about, so reasoning around this could easily switch my views.
quick thougths RE your reasons for working on it or not:
1) It seems like many people are not seeing them coming (e.g. AI safety community seems surprisingly unreceptive and to have made many predictable mistakes by ignoring structural causes of risk, e.g. being overly optimistic about companies prioritizing safety over competitiveness) 1) It seems like seeing them coming is predictably insufficient to stopping them happening, because they are the result of social dilemmas. 1) The structure of the argument appears to be the (fallacious): “if it is a real problem, other people will address it, so we don’t need to” (cf https://www.explainxkcd.com/wiki/index.php/2278:_Scientific_Briefing)
2) Interesting. Seems potentially cruxy.
3) I guess we might agree here… combined with (1), I guess your argument is: “won’t be neglected (1) and is not tractable (3)”, whereas I might say: “currently neglected, could require a lot of work to become tractable, seems important enough to warrant that effort”
The main upshots I see are: - higher P(doom) due to stories that are easier for many people to swallow --> greater ability and potential for public awareness and political will if messaging includes this. - more attention needed to questions of social organization post-AGI.
Thanks! Yes, definitely in scope. There was a lot of discussion of this paper when it came out, and we had Raymond Douglas speak at a seminar.
Opinions vary within the team on how valuable it is to work on this; I believe Fin and Tom are pretty worried about this sort of scenario (I don’t know about others).. I feel a bit less convinced on the value of working on it (relative to other things), and I’ll just say why briefly:
- I feel less convinced that people wouldn’t foresee the bad gradual disempowerment scenarios and act to stop them from happening, esp with advanced AI assistance
- In the cases that feel more likely, I feel less convinced that gradual disempowerment is particularly bad (rather than just “alien”).
- Insofar as there are bad outcomes here, it seems particularly hard to steer the course of history away from them.
The biggest upshot I see is that, the more you buy these sorts of scenarios, the more it increases the value of AGI being developed by a single e.g. multilateral project rather than being developed by multiply companies and countries. That’s something I’m really unsure about, so reasoning around this could easily switch my views.
quick thougths RE your reasons for working on it or not:
1) It seems like many people are not seeing them coming (e.g. AI safety community seems surprisingly unreceptive and to have made many predictable mistakes by ignoring structural causes of risk, e.g. being overly optimistic about companies prioritizing safety over competitiveness)
1) It seems like seeing them coming is predictably insufficient to stopping them happening, because they are the result of social dilemmas.
1) The structure of the argument appears to be the (fallacious): “if it is a real problem, other people will address it, so we don’t need to” (cf https://www.explainxkcd.com/wiki/index.php/2278:_Scientific_Briefing)
2) Interesting. Seems potentially cruxy.
3) I guess we might agree here… combined with (1), I guess your argument is: “won’t be neglected (1) and is not tractable (3)”, whereas I might say: “currently neglected, could require a lot of work to become tractable, seems important enough to warrant that effort”
The main upshots I see are:
- higher P(doom) due to stories that are easier for many people to swallow --> greater ability and potential for public awareness and political will if messaging includes this.
- more attention needed to questions of social organization post-AGI.