Regulatory type interventions (pre-deployment):
Regulatory restriction (rules on what can be done)
Regulatory oversight (regulators)
Industry (& regulator) peer reviews systems
Senior management regimes
Information sharing regimes
Staff security clearances
Cybersecurity of AI companies
Standardisation (to support ease of oversight etc)
Clarity about liability & legal responsibility
Internal government oversight (all of the above applied internally by government to itself, e.g. internal military safety best practice)
Technical type interventions (pre-deployment):
AI safety research
Defence in depth type interventions (post-deployment):
Windfall clauses etc
Shut-off switches for AI systems
AIs policing other AIs’ behaviours
Internet / technology shut-off systems
A thing I have been thinking about but am not at all an expert on is open-source research into high-yield seed production.Currently high-yield seeds alone can significantly increase outputs by something like 25-55% One of the big challenges with improved seeds and agritech is that the current model doesn’t seem to be working for the ~25% of the worlds population who are subsistence farmers. Maybe there is no profit in it. For example subsistence farmers don’t want to be dependent on having to buy seeds every year but want to be able to save seed from their own crops, but seed developers prefer seeds that need to be purchased annually. If I was suggesting a $6bn megaproject to end world hunger that would appeal to some ambitious tech entrepreneur it would be something like open-source agritech designed for (and by?) the worlds poorest, not like “hey donate to an EA food charity like Fortify Health or Food Fortification Initiative” (although that would be nice too).Just mentioning the idea on the off chance it is helpful to the debate, to anyone’s op-eds, anyone chatting to Musk or whatever.
It it is not the full answer but in my experience active grant making is easier for a committed small donor than a fund.At least two projects I am aware of: Effective Altruism London and the All-Party Parliamentary Group for Future Generations were made significantly more likely to happen by active grant making from small donors. In both cases the projects were being run by founders as volunteers. Donors who knew the founders and could form a view on their value of their work reached out and said: “this project is good, deserves full time staff and I could offer funding if needed”. Both projects hired staff, grew and in later years went on to receive future funding from various official EA funds. I think in this way well-connected EtG’ers can make an huge impact that grant-makers cannot make.
Your push back here seems fair. These orgs certainly do some good work across this whole spectrum. My shoddy diagrams were supposed to be more illustrative of a high level point than accurate. But perhaps they are somewhat over-exaggerated and critical. I still think the high-level point about expectations and reality is worth making (like the point about people’s expectations about GovAI).
Hi Jack, some semi-unstructured thoughts for you on this topic. And as the old quote goes “sorry, if I had more time, I would have written you a shorter letter”:
What are we aiming to do here with this discussion? It seems like we are trying to work out what are the best thinking tools for various kinds of questions an altruist might want answered. And we are deciding between two somewhat distinct and also overlapping and also not well defined ‘ways of thinking’ whilst still acknowledging that both are going to be at least somewhat useful across the spectrum and that everyone will have a preference for the tools they are used to using (and speaking at least for myself I have limited expertise in judging the value of academic type work) !!! …. Well sure why not. Let’s do that. Happy to throw some uniformed opinions out into the void and see where it goes ….
How should or could we even make such a judgement call? I believe we have evidence from domains such as global health and policy design that is you are not evaluating and testing you are likely not having the impact you expect. I don’t see why this would not apply to research. In my humble view of you wanted to know what research was most valuable you would want monitoring and evaluation and iteration. Clear theories of change and justification for research topics. Research could be judged by evidence of real world impact, feedback from users of the research, prediction markets, expert views, etc. Space for more exploratory research would be carved out. This would all be done transparently and maybe there would be competing research organisations* that faced independent evaluation. And so forth. And then you wouldn’t need to guess at what tools were useful, you’d find out.
But easier said that done? I think people would push back, say they know they don’t need evaluating, creativity will be stifled, evaluation is difficult, we would agree on how to evaluate, that’s not how academia works, we already know we are great, etc. They might be right. I have mostly looked at EA research orgs as a potential donor so this is at least the direction I would want things to move in – but I am sure there are other perspectives. And either way I don’t know if anyone is making an push to move EA research organisations in that direction.
So what can be done? I guess one thing would be if the people who used research to make decisions, would give feedback to researchers about what research they find more useful and less useful, that could be a step in the right direction. As I have done somewhat here. And my feedback from the policy change side of the fence is that the research that looks more like the long-term planning tools (that I am used to using) is more useful. I have found it more useful to date and expect it would be more useful going forward. And I would like to see more of it. I don’t think that feedback from me is sufficient to answer the question, it is a partial answer at best!! There are (I am sure) at least some other folk in the world that use FHI/GPI/EA research who will hopefully have different views.
So with that lengthy caveat about partial answers / plea for more monitoring and evaluation out of the way, you asked for specific examples:
Long-term planning and worldview. I think the basic idea that the long-term future matters is probably true. I think how tractable this is and how much that should effect our decisions is less clear. How to make decisions about influencing the world given uncertainty/cluelessness? How tractable it is to have that influence? Are attractor states the only way to have that long-run influence? Are stopping risks the only easy to reach attractor states? I think the long-term planning tools give us at least one way to approach these kinds of questions (as I set out above in Section B. 4.-5.). To design plans and to influence the future, ensure their robustness, then put them into action and see how they work, and so on. Honestly I doubt you would get radically different answers (although I also doubt other ways of approaching this question would lead to radically different answers either, I just quite uncertain how tractable more worldview research work is).
Long-term planning and cause choice. This seems obvious to me. If you want to know, as per your example, what risks to care about – then mapping out future paths and scenarios the world the world might take, doing estimates of risks on 5, 10, 25, 50, 100 year timescales, explicitly evaluating your assumptions, doing forecasting, using risk management tools, identifying the signs to watch-out for that would warrant a change in priorities, and so on – all seems to be super useful.Also I think there might be a misunderstanding but the whole point of all the tools listed above is that they are for use in situations where you are dealing with the “unprecedented” and the unknown black swans. If something is easy to predict then you can just predict it and do a simple expected value calculation and you are done (and EA folk are already good at that). Overall I doubt you would get drastically different answers to which risks matter, although I would expect there may be a greater focus on building a world that is robust to other “unforeseen anthropogenic risks”, not just AI and bio. I also think in specific circumstances that people might be in, say a meeting with a politician or writing a strategy, they would hopefully have a better sense of which risks to focus on.
* P.S. Not sure anyone will have read this far but if anyone is reading this and actually thinks it could be a good (or v bad) idea to start an research org with a focus on demonstrating impact, policy research, and planning type tools – then do get in touch.
I guess my main point is that I’d like to see some applications of this framework (and some of the other frameworks you mention too) to important longtermist problems, before I accept it as useful.
I 100% fully agree. Totally. I think a key point I want to make is that we should be testing all our frameworks against the real world and seeing what is useful. I would love to sit down with CLR or FHI or other organisations and see how this framework can be applied. (Also expect a future post with details of some of the policy work I have been doing that uses some of this).
(I would also love people who have alternative frameworks to be similarly testing them in terms of how much they lead to real world outputs or changes in decisions.)
I’m still unsure if this would be the best approach to reducing existential risk
The aim here is to provide a tool kit that folk can use when needed.
For example these tools they are not that useful where solutions are technical and fairly obvious. I don’t think you need to go through all these steps to conclude that we should be doing interpretability research on AI systems. But if you want to make plans to ensure the incentives of future researches who develop a transformative AI are aligned to the global good then you have a complex high-uncertainty long-term problem and I expect these kinds of tools become the sort of think you would want to use.
Also as I say in the post more bespoke tools beat more general tools. Even in specific cases there will be other toolboxes to use. Organisational design methods for aligning future actors incentives, vulnerability assessments for identifying risks, etc. The tools above are the most general form for anyone to pick up and use.
I’m also sceptical about the claim that we can’t affect probabilities of lock-in events that may happen beyond the next few decades. As I also say here, what about growing the Effective Altruism/longtermist community, or saving/investing money for the future, or improving values?
I think this is a misunderstanding. You totally can affect those events. (I gave the example of patient philanthropy that has non-negligible expected value even in 300 years.) But in most cases a good way of having an impact in more than a few decades is to map out high level goals on shorter decade long timelines. On climate change we are trying to prevent disaster in 2100 but we do it by stetting targets for 2050. The forestry commission might plant oak tress that will grow for 100s of years but they will make planting plans on 10 year cycles. Etc
What would the 30 year vision be? What would intermediate targets be?
Some examples here if helpful.
Part 2 – Also, a note on expectations
On a slightly separate point maybe some of the challenge I feel here comes from me having misplaced expectations. I think that before I dived into the longtermist academic research I was hoping that the world looked like this:
GPI and FHI, etc
and I could find the answers I needed and get on with driving change – YAY.
But maybe the world actually looks like more this:
GPI and FHI
and there is so much more to do – Awww.
(Reminds me of talking to GovAI about policy and they said GovAI does not do applied policy research but people often think that they do it. )
I know it is not going to be top of anyone’s to do list but I would love at some point to see an FHI post like this one from 80K setting out what is in scope and what is out of scope that could be great for others in the ecosystem to do.(* diagrams oversimplified again but hopefully they make the point)
Thank you for the thoughts and for engaging. I think this is a really good point. I mostly agree.
To check what you are saying. It seems like the idea here is that there are different aims of altruistic research. In fact we could imagine a spectrum, something like:
At the top end, for people thinking about ethics etc, traditional longtermist ways of thinking are best and at the lower end for someone thinking about plans etc, long-term planning tools are the best.
I think this is roughly correct.
My hope is to provide a set of long-term planning tools that people might find useful, not to rubbish the old tools.
That said, reading between the lines a bit, it feels like there is still some disconnect about the usefulness and importance of different kinds of research. I go into each of these a little bit below.
A useful analogy would be the difference between using cost-effectiveness as a tool for selecting a top cause or intervention to work on, vs using it to work out the most cost-effective way to do what you are already doing.
I am speculating a bit (so correct me if I am wrong) but I get the impression from that analogy that you would see the best tools to use a bit like this
(Diagram is an over-simplification as both ways of thinking will be good across the spectrum so the cut of would be vague, but this chart is the best I can do on this forum.)
However I would see it a bit more like this:
And the analogy I would use would be something like:
A useful analogy would be the difference between using philosophical “would you rather?” thought experiments as a tool for selecting an ethical view , vs using thought experiments to work out the most important causes to work on .
Deciding what the best ways of thinking are best suited for different intellectual challenges is a huge debate. I could give views but not sure we are going to solve it here. And it makes sense that we each prefer to rely on the ways of thinking that we are experienced in using.
One of my aims of writing this post is to give feedback to researchers, as a practitioner about what kind of work I find useful. Basically trying to shorten the feedback loop as much as I can to help guide future research.
So what I can do is provide my own experiences. I do have to on a semi-regular basis make decisions about causes and interventions to focus on (do we engage politicians about AI or bio or improving institutions, etc). And in making these high level decisions there is some good research and some less useful research (of the type I discuss in my post) and my general view is that more work like: scenarios, short term estimates of x-risk, vulnerability assessments, etc – would be particularly useful for me making even very high-level cause decisions.
Maybe that is useful for you or others (I hope so).
Perhaps we also have different views on what work is valuable. I guess I already think that the future matters and see less value in more work on is longtermism true and more value on work into what are the risks we face now and how can we address them.
[Long term planning] is not always required in order to get large gains
Let’s flesh out what we mean by “gains”.
If gains at philosophy / ethics / deciding if longtermism is true, then yes this would apply.
If gains at actually reducing the chance of an existential catastrophe (other than in cases where the solution is predominately technical) then I don’t think this would be true.
I expect we agree on that. So maybe the question is less about the best way of thinking about the world and and more about what the aims of additional research should be? Should we be pushing resources to more philosophy or towards more actionable plans to affect the long-term and/or reduce risks?
(Also worth considering the extent to which demonstrating practical actionable plans is useful for the philosophy, either for learning how to deal with uncertainty or for making the case that the long-term is a thing people can act on).
I don’t think I said the the US military was good at risk management I think I said that a) the DMDU community (RAND, US military and others) was good at making plans that manage uncertainty and b) that industry was good at risk management
It feels wrong to use reference classes of X to implicitly say that actions the reference class does is good and we ought to emulate them, without ever an explicit argument that the reference classes’ actions or decision procedures are good!
I do think where reference class X is large and dominant enough it does make sense to assume some trust in their approach or that it is worth some investigation of their approach before dismissing it. For example most (large) businesses have a board and a CEO and a hierarchical management structure so unless I had a good reason to do otherwise that sets a reasonable prior for how I think it is best to run a business.For more on this see Common sense as a prior.So even if I had zero evidence I think it would make sense for someone in the EA community to spend time looking into the topic of what tools had worked well in the past to deal with uncertainty and that the US military would be a good place to look for ideas.
Answering: is the US military good at making plans that manage uncertainty:
Historical evidence – no.I have zero empirical historical evidence that DMDU tools have worked well for the US military.
Theoretical evidence – yes.I think the theoretical case for these tools is strong, see the case here and here.
Interpersonal evidence – yes.I believe Taleb in Black Swan describes that the people he met in the US military had very good empirical ways of thinking about risk and uncertainty (I don’t have the book here so cannot double check). Similarly to Taleb I have been much impressed by folk in the UK working on counter-terrorism etc, compared to other policy folk who work on risks.
Evidence from trust – mixed.I mostly expect the US military have the right incentives in place to aim to do this well and the ability to test ideas in the field but also would not be surprised if there were a bunch of perverse incentives that corrupted this.
So all in all pretty weak evidence.
My views are probably somewhat moved on from when I wrote this post a year ago. I should revisit it at some point
Thank you Luke – great to hear this work is happening but still surprised by the lack of progress and would be keen to see more such work out in public!(FWIW Minor point but I am not sure I would phrase a goal as “make government generically smarter about AI policy” just being “smart” is not good. Ideally want a combination of smart + has good incentives + has space to take action. To be more precise when planning I often use COM-B models, as used in international development governance reform work, to ensure all three factors are captured and balanced.)
Also Ben, is there a Jews and EA Facebook group – any plans to set one up? Or if I set one up do you think you could email / share it?
Thank you Luke for sharing your views. I just want to pick up one thing you said where your experience of the longtermist space seems sharply contrary to mine.
You said: “We lack the strategic clarity … [about] intermediate goals”. Which is a great point and I fully agree. Also I am super pleased to hear you have been working on this. You then said:
I caution that several people have tried this … such work is very hard
This surprised me when I read it. In fact my intuition is that such work is highly neglected, almost no one has done any of this and I expect it is reasonably tractable. Upon reflection I came up with three reasons for my intuition on this.1. Reading longtermist research and not seeing much work of this type.I have seem some really impressive forecasting and trend analysis focused but if anyone had worked on setting intermediate goals I would expect to see some evidence of basic steps such as listing out a range of plausible intermediate goals or consensus building exercises to set viable short and mid term visions of what AI governance progress looks like (maybe it’s there and I’ve just not seen it). If anyone had made a serious stab at this I would expect to have seen thorough exploration exercises to map out and describe possible near-term futures, assumption based planning, scenario based planning, strategic analysis of a variety of options, tabletop exercises, etc. I have seen very little of this.2. Talking to key people in the longtermist space and being told this research is not happening.For a policy research project I was considering recently I went and talked to a bunch of longtermists about research gaps (eg at GovAI, CSET, FLI, CSER, etc). I was told time and time again that policy research (which I would see as a combination of setting intermediate goals and working out what policies are needed to get there) was not happening, was a task for another organisation, was a key bottleneck that no-one was working on, etc.
3. I have found it fairly easy to make progress on identifying intermediate goals and short-term policy goals that seem net-positive for long-run AI governance
I have an intermediate goal of: key actors in positions of influence over AI governance are well equipped to make good decisions if needed (at an AI crunch time). This leads to specific policies such as: Ensuring clear lines of responsibility exist in military procurement of software /AI or, if regulation happens it should be expert driven outcome based regulation or some of the ideas here. I would be surprised if longtermists looking into this (or other intermediate goals I routinely use) would disagree with the above intermediate goal or that the policy suggestions move us towards that goal. I would say this work has not been difficult.– – So why is our experience of the longtermist space so different. One hunch I have is that we are thinking of different things when we consider “strategic clarity on intermediate goals”.In supporting governments to make long-term decisions and has given me a sense of what long-term decision making and “intermediate goal setting” and long-term decision making involves. This colours the things I would expect to see if the longtermist community was really trying to do this kind of work and I compare longtermists’ work to what I understand to be best practice in other long-term fields (from forestry to tech policy to risk management). This approach leaves me thinking that there is almost no longtermist “intermediate goal setting” happening. Yet maybe you have a very different idea of what “intermediate goal setting involves” based on other fields you have worked in.
It might also be that we read different materials and talk to different people. It might be that this work has happened I’ve just missed it or not read the right stuff.– –Does this matter? I guess I would be much more encouraging about someone doing this work than you are and much more positive about how tractable such work is. I would advise that anyone doing this work should have a really good grasp of how wicked problems are addressed and how long-term decision making works in a range of non-EA fields and the various tools that can be used.
I have an idea and though a comment here would be a good place to put it:I wonder if there should be a Jewish run EA charity or Charitable Fund that directs funds to good places (such as assorted EA organisations).
I think lots of Jews want to give to a Jewish run organisation or give within the Jewish community. If a Jewish run EA charity existed it could be helpful for making the case for more global effective giving.It could be run with Jewish grant managers who ensure that funds are used well and in line with Jewish principles (there could be a Pikuach nefesh fund for saving the most lives, or a Maimonides ladder sustainable growth fund, etc).To argue against this idea: one of the nice things about EA is it is not us asking for your money it is us advising on where you should give your money which feels nicer and is maybe an easier pitch. So maybe if there was an EA run Jewish charity or fund it might detract form that or should be separate from the outreach efforts.Happy to help a bit with this if it happens.
Another slightly tangential but very similar question that came up in conversation I had recently is:”How well have EA-funded orgs built on the momentum created by the COVID-motivated global interest in GCRs (global catastrophic risks) to drive policy change or other changes to help prevent GCRs and x-risks”I could have imagined a world where the entire longtermist community pivoted towards this goal and at least for a year or two and focused all available time skill and money on driving GCR related policy change – but this doesn’t seem to have happened much. I could imagine the community looking back at this year and regretting the collective lack of action.The organisation where I work, the APPG for Future Generations pivoted significantly, kickstarted a new Parliament Committee on risks and I wrote a paper on lessons learned from COVID which had significantly government interest and seems to have driven policy change (writeup forthcoming).
But beyond that there has definitely been some exciting stuff happening. I know:
CSER are starting a lessons learned from COVID project, although this is only just getting started.
FHI staff have submitted a some evidence to parliamentary inquiries (example).
The CLTR (funded by the EAIF) has launched a report on risk (I’m unsure if this was a change in direction or always the plan).
No more pandemics (not funded) was started.
This stuff is all great and I am sure there is more happening – but my general sense is that it is much less than and much slower than I would have expected.
I also loosely get the impression (from my own experience and that of 2-3 other orgs that I have talked to) that various EA funders have been disinterested in pivoting to support lessons learned from COVID focused policy work, some of which could scale up quite significantly, and that maybe funding is the main bottleneck for some of this (I think funding for more policy work is a bottleneck for all of the orgs listed above except FHI).
[Disclaimer – I will be bias given that I pivoted my work to focus on COVID lessons learned and policy influencing and looked for funding for this.]
Hello, Thank you for the interesting thoughts. The comments on the GHS index are useful and insightful.
Your analysis of COVID preparation on Twitter is really really interesting. Well done for doing that. I have not yet looked at your analysis spreadsheet but will try to do that soon.To touch on a point you said about preparation, I think we can take a bit more of a nuanced approach to think about when preparation works rather than just saying “effective pandemic response is not about preparation”. Some thoughts from me on this (not just focused on pandemics).
Prevention definitely helps. (It is a semantic question if you want to count prevention as a type of preparation or not). The world is awash with very clear examples of disaster prevention whether it is engineering safe bridges, or flood prevention, or nuclear safety, or preventing pathogens escaping labs, etc.
The idea that preparation (henceforth excluding prevention) helps is conventional wisdom and I think I would want to see good evidence against this to stop believing in this.
Obviously preparation helps in the small cases, talk to a paramedic rushing to treat someone or a fireman. I have not looked into it but I get the impression that it helps in the medium cases, eg rapid response teams responding to terror attacks in the UK / France seem useful, although not an expert. On pandemics specifically the quick containment of SARs seems to be a success story (although I have not looked at how much preparation played a role it does seem to be a part of the story). There are not that many extreme COVID-level cases to look at, but it would be odd if it didn’t help in extreme cases too.
The specific wording of the claim in the linked article headline feels clickbait-y. When you actually read the article it actually says that competence matters more (I agree) and also that we should focus more on designing resilient anti-fragile systems rather than event specific preparation. I agree but I think that designing systems that can make good decisions in a risk scenario is a form of preparation.
I do agree that your analysis provides some evidence that preparation did not help with COVID. I am cautious of the usefulness of this evidence because of the problems with the GHS – e.g. the UK came near top but basically had no plan to deal with any non-influenza pandemic that I have identified.
A confusing factor that might make it hard to tell if preparation helped is that, based on the UK experience (eg discussed here) it appears that having bad plans in place may actually be worse than no plans.
Evidence from COVID does suggest to me that specific preparation does help. Notably countries (E Asia, Australasia) that had SARs and prepared for future SARs type outbreaks managed COIVD better.
So maybe we can say something like:Prevention definitely helps. Both event specific preparation and generally building robust anti-fragile decision systems are useful approaches but the latter of those is more underinvested in. However good leadership is necessary as well as preparation and without good leadership (which maybe rare) preparation can turn out to be useless. Furthermore bad preparation, such as poor planning, can potentially hinder a response more than no preparation. Does that seem like a good summary and sufficiently explain your findings. I am thinking about doing more work to promote preparation so useful to hear if you disagree.
[Edit – moved comment to answer above at suggestion of kbog]
Thank you :-)
think a significant issue is that both of these cost time
I am always amazed at how much you fund managers all do given this isn’t your paid job!
I don’t think it’s obvious whether at the margin the EAIF committee should spend more or less time to get more or fewer benefits in these areas
Fair enough. FWIW my general approach to stuff like this is not to aim for perfection but to aim for each iteration/round to be a little bit better than the last.
… it could be that I’m just bad at getting value out of discussions, or updating my views, or something like that.
That is possible. But also possible that you are particularly smart and have well thought-out views and people learn more from talking to you than you do from talking to them!(And/or just that everyone is different and different ways of learning work for different people)