[Some of my tentative and uncertain views on AI governance, and different ways of having impact in that area. Excerpts, not in order, from things I wrote in a recent email discussion, so not a coherent text.]
1. In scenarios where OpenAI, DeepMind etc. become key actors because they develop TAI capabilities, our theory of impact will rely on a combination of affecting (a) ‘structure’ and (b) ‘content’. By (a) I roughly mean how the relevant decision-making mechanisms look like irrespective of the specific goals and resources of the actors the mechanism consists of; e.g., whether some key AI lab is a nonprofit or a publicly traded company; who would decide by what rules/voting scheme how Windfall profits would be redistributed; etc. By (b) I mean something like how much the CEO of a key firm, or their advisors, care about the long-term future. -- I can see why relying mostly on (b) is attractive, e.g. it’s arguably more tractable; however, some EA thinking (mostly from the Bay Area / the rationalist community to be honest) strikes me as focusing on (b) for reasons that seem ahistoric or otherwise dubious to me. So I don’t feel convinced that what I perceive to be a very stark focus on (b) is warranted. I think that figuring out if there are viable strategies that rely more on (a) is better done from within institutions that have no ties with key TAI actors, and also might be best done my people that don’t quite match the profile of the typical new EA that got excited about Superintelligence or HPMOR. Overall, I think that making more academic research in broadly “policy relevant” fields happen would be a decent strategy if one ultimately wanted to increase the amount of thinking on type-(a) theories of impact.
2. What’s the theory of impact if TAI happens in more than 20 years? More than 50 years? I think it’s not obvious whether it’s worth spending any current resources on influencing such scenarios (I think they are more likely but we have much less leverage). However, if we wanted to do this, then I think it’s worth bearing in mind that academia is one of few institutions (in a broad sense) that has a strong track record of enabling cumulative intellectual progress over long time scales. I roughly think that, in a modal scenario, no-one in 50 years is going to remember anything that was discussed on the EA Forum or LessWrong, or within the OpenAI policy team, today (except people currently involved); but if AI/TAI was still (or again) a hot topic then, I think it’s likely that academic scholars will read academic papers by Dafoe, his students, the students of his students etc. Similarly, based on track records I think that the norms and structure of academia are much better equipped than EA to enable intellectual progress that is more incremental and distributed (as opposed to progress that happens by way of ‘at least one crisp insight per step’; e.g. the Astronomical Waste argument would count as one crisp insight); so if we needed such progress, it might make sense to seed broadly useful academic research now.
[...]
My view is closer to “~all that matters will be in the specifics, and most of the intuitions and methods for dealing with the specifics are either sort of hard-wired or more generic/have different origins than having thought about race models specifically”. A crux here might be that I expect most of the tasks involved in dealing with the policy issues that would come up if we got TAI within the next 10-20 years to be sufficiently similar to garden-variety tasks involved in familiar policy areas that as a first pass: (i) if theoretical academic research was useful, we’d see more stories of the kind “CEO X / politician Y’s success was due to idea Z developed through theoretical academic research”, and (ii) prior policy/applied strategy experience is the background most useful for TAI policy, with usefulness increasing with the overlap in content and relevant actors; e.g.: working with the OpenAI policy team on pre-TAI issues > working within Facebook on a strategy for how to prevent the government to split up the firm in case a left-wing Democrat wins > business strategy for a tobacco company in the US > business strategy for a company outside of the US that faces little government regulation > academic game theory modeling. That’s probably too pessimistic about the academic path, and of course it’ll depend a lot on the specifics (you could start in academia to then get into Facebook etc.), but you get the idea.
[...]
Overall, the only somewhat open question for me is whether ideally we’d have (A) ~only people working quite directly with key actors or (B) a mix of people working with key actors and more independent ones e.g. in academia. It seems quite clear to me that the optimal allocation will contain a significant share of people working with key actors [...]
If there is a disagreement, I’d guess it’s located in the following two points:
(1a) How big are countervailing downsides from working directly with, or at institutions having close ties with, key actors? Here I’m mostly concerned about incentives distorting the content of research and strategic advice. I think the question is broadly similar to: If you’re concerned about the impacts of the British rule on India in the 1800s, is it best to work within the colonial administration? If you want to figure out how to govern externalities from burning fossil fuels, is it best to work in the fossil fuel industry? I think the cliche left-wing answer to these questions is too confident in “no” and is overlooking important upsides, but I’m concerned that some standard EA answers in the AI case are too confident in “yes” and are overlooking risks. Note that I’m most concerned about kind of “benign” or “epistemic” failure modes: I think it’s reasonably easy to tell people with broadly good intentions apart from sadists or even personal-wealth maximizers (at least in principle—if this will get implemented is another question); I think it’s much harder to spot cases like key people incorrectly believing that it’s best if they keep as much control for themselves/their company as possible because after all they are the ones with both good intentions and an epistemic advantage (note that all of this really applies to a colonial administration with little modification, though here in cases such as the “Congo Free State” even the track record of “telling personal-wealth maximizers apart from people with humanitarian intentions” maybe isn’t great—also NB I’m not saying that this argument would necessarily be unsound; i.e. I think that in some situations these people would be correct).
(1b) To what extent to we need (a) novel insights as opposed to (b) an application of known insights or common-sense principles? E.g., I’ve heard claims that the sale of telecommunication licenses by governments is an example where post-1950 research-level economics work in auction theory has had considerable real-world impact, and AFAICT this kind of auction theory strikes me as reasonably abstract and in little need of having worked with either governments or telecommunication firms. Supposing this is true (I haven’t really looked into this), how many opportunities of this kind are there in AI governance? I think the case for (A) is much stronger if we need little to no (a), as I think the upsides from trust networks etc. are mostly (though not exclusively) useful for (b). FWIW, my private view actually is that we probably need very little of (a), but I also feel like I have a poor grasp of this, and I think it will ultimately come down to what high-level heuristics to use in such a situation.
I found this really fascinating to read. Is there any chance that you might turn it into a “coherent text” at some point?
I especially liked the question on possible downsides of working with key actors; orgs in a position to do this are often accused of collaborating in the perpetuation of bad systems (or something like that), but rarely with much evidence to back up those claims. I think your take on the issue would be enlightening.
Thanks for sharing your reaction! There is some chance that I’ll write up these and maybe other thoughts on AI strategy/governance over the coming months, but it depends a lot on my other commitments. My current guess is that it’s maybe only 15% likely that I’ll think this is the best use of my time within the next 6 months.
[Some of my tentative and uncertain views on AI governance, and different ways of having impact in that area. Excerpts, not in order, from things I wrote in a recent email discussion, so not a coherent text.]
1. In scenarios where OpenAI, DeepMind etc. become key actors because they develop TAI capabilities, our theory of impact will rely on a combination of affecting (a) ‘structure’ and (b) ‘content’. By (a) I roughly mean how the relevant decision-making mechanisms look like irrespective of the specific goals and resources of the actors the mechanism consists of; e.g., whether some key AI lab is a nonprofit or a publicly traded company; who would decide by what rules/voting scheme how Windfall profits would be redistributed; etc. By (b) I mean something like how much the CEO of a key firm, or their advisors, care about the long-term future. -- I can see why relying mostly on (b) is attractive, e.g. it’s arguably more tractable; however, some EA thinking (mostly from the Bay Area / the rationalist community to be honest) strikes me as focusing on (b) for reasons that seem ahistoric or otherwise dubious to me. So I don’t feel convinced that what I perceive to be a very stark focus on (b) is warranted. I think that figuring out if there are viable strategies that rely more on (a) is better done from within institutions that have no ties with key TAI actors, and also might be best done my people that don’t quite match the profile of the typical new EA that got excited about Superintelligence or HPMOR. Overall, I think that making more academic research in broadly “policy relevant” fields happen would be a decent strategy if one ultimately wanted to increase the amount of thinking on type-(a) theories of impact.
2. What’s the theory of impact if TAI happens in more than 20 years? More than 50 years? I think it’s not obvious whether it’s worth spending any current resources on influencing such scenarios (I think they are more likely but we have much less leverage). However, if we wanted to do this, then I think it’s worth bearing in mind that academia is one of few institutions (in a broad sense) that has a strong track record of enabling cumulative intellectual progress over long time scales. I roughly think that, in a modal scenario, no-one in 50 years is going to remember anything that was discussed on the EA Forum or LessWrong, or within the OpenAI policy team, today (except people currently involved); but if AI/TAI was still (or again) a hot topic then, I think it’s likely that academic scholars will read academic papers by Dafoe, his students, the students of his students etc. Similarly, based on track records I think that the norms and structure of academia are much better equipped than EA to enable intellectual progress that is more incremental and distributed (as opposed to progress that happens by way of ‘at least one crisp insight per step’; e.g. the Astronomical Waste argument would count as one crisp insight); so if we needed such progress, it might make sense to seed broadly useful academic research now.
[...]
My view is closer to “~all that matters will be in the specifics, and most of the intuitions and methods for dealing with the specifics are either sort of hard-wired or more generic/have different origins than having thought about race models specifically”. A crux here might be that I expect most of the tasks involved in dealing with the policy issues that would come up if we got TAI within the next 10-20 years to be sufficiently similar to garden-variety tasks involved in familiar policy areas that as a first pass: (i) if theoretical academic research was useful, we’d see more stories of the kind “CEO X / politician Y’s success was due to idea Z developed through theoretical academic research”, and (ii) prior policy/applied strategy experience is the background most useful for TAI policy, with usefulness increasing with the overlap in content and relevant actors; e.g.: working with the OpenAI policy team on pre-TAI issues > working within Facebook on a strategy for how to prevent the government to split up the firm in case a left-wing Democrat wins > business strategy for a tobacco company in the US > business strategy for a company outside of the US that faces little government regulation > academic game theory modeling. That’s probably too pessimistic about the academic path, and of course it’ll depend a lot on the specifics (you could start in academia to then get into Facebook etc.), but you get the idea.
[...]
Overall, the only somewhat open question for me is whether ideally we’d have (A) ~only people working quite directly with key actors or (B) a mix of people working with key actors and more independent ones e.g. in academia. It seems quite clear to me that the optimal allocation will contain a significant share of people working with key actors [...]
If there is a disagreement, I’d guess it’s located in the following two points:
(1a) How big are countervailing downsides from working directly with, or at institutions having close ties with, key actors? Here I’m mostly concerned about incentives distorting the content of research and strategic advice. I think the question is broadly similar to: If you’re concerned about the impacts of the British rule on India in the 1800s, is it best to work within the colonial administration? If you want to figure out how to govern externalities from burning fossil fuels, is it best to work in the fossil fuel industry? I think the cliche left-wing answer to these questions is too confident in “no” and is overlooking important upsides, but I’m concerned that some standard EA answers in the AI case are too confident in “yes” and are overlooking risks. Note that I’m most concerned about kind of “benign” or “epistemic” failure modes: I think it’s reasonably easy to tell people with broadly good intentions apart from sadists or even personal-wealth maximizers (at least in principle—if this will get implemented is another question); I think it’s much harder to spot cases like key people incorrectly believing that it’s best if they keep as much control for themselves/their company as possible because after all they are the ones with both good intentions and an epistemic advantage (note that all of this really applies to a colonial administration with little modification, though here in cases such as the “Congo Free State” even the track record of “telling personal-wealth maximizers apart from people with humanitarian intentions” maybe isn’t great—also NB I’m not saying that this argument would necessarily be unsound; i.e. I think that in some situations these people would be correct).
(1b) To what extent to we need (a) novel insights as opposed to (b) an application of known insights or common-sense principles? E.g., I’ve heard claims that the sale of telecommunication licenses by governments is an example where post-1950 research-level economics work in auction theory has had considerable real-world impact, and AFAICT this kind of auction theory strikes me as reasonably abstract and in little need of having worked with either governments or telecommunication firms. Supposing this is true (I haven’t really looked into this), how many opportunities of this kind are there in AI governance? I think the case for (A) is much stronger if we need little to no (a), as I think the upsides from trust networks etc. are mostly (though not exclusively) useful for (b). FWIW, my private view actually is that we probably need very little of (a), but I also feel like I have a poor grasp of this, and I think it will ultimately come down to what high-level heuristics to use in such a situation.
I found this really fascinating to read. Is there any chance that you might turn it into a “coherent text” at some point?
I especially liked the question on possible downsides of working with key actors; orgs in a position to do this are often accused of collaborating in the perpetuation of bad systems (or something like that), but rarely with much evidence to back up those claims. I think your take on the issue would be enlightening.
Thanks for sharing your reaction! There is some chance that I’ll write up these and maybe other thoughts on AI strategy/governance over the coming months, but it depends a lot on my other commitments. My current guess is that it’s maybe only 15% likely that I’ll think this is the best use of my time within the next 6 months.