I am extremely impressed by this, and this is a great example of the kind of ambitious projects I would love to see more of in the EA community. I have added it to the list on my post Even More Ambitious Altruistic Tech Efforts.
Best of luck!
I am extremely impressed by this, and this is a great example of the kind of ambitious projects I would love to see more of in the EA community. I have added it to the list on my post Even More Ambitious Altruistic Tech Efforts.
Best of luck!
Not answering the question, but I would like to quickly mention a few of the benefits of having confidence/credible intervals or otherwise quantifying uncertainty. All of these comments are fairly general, and are not specific criticisms of GiveWell’s work.
Decision making under risk aversion—Donors (large or small) may have different levels of risk aversion. In particular, some donors might prefer having higher certainty of actually making an impact at the cost of having a lower expected value. Moreover, (mostly large) donors could build a portfolio of different donations in order to achieve a better risk profile. To that end, one needs to know more about the distribution rather than a point-estimate.
Point-estimates are many times done badly—It is fairly easy to make many kinds of mistakes when doing point-estimates, some of which are more noticeable when quantifying uncertainties. To name one example, point-estimates of cost-effectiveness typically try to estimate the expected value, and is many times calculated as a product of different factors. While it is true that expected value is multiplicative (assuming that the factors are uncorrelated or, more generally, independent, which is also sometimes not the case but that’s another problem), this is not true for other statistics, such as the median. I think it is a common mistake to use an estimate of the median for the mean, or something in between, which in many cases are wildly different.
Sensitivity analysis—Quantifying uncertainty allows for sensitivity analysis, which serves many purposes, one of which is to get more accurate (point-)estimate and reduce uncertainty. One example is by understanding which parameters are the most uncertain, and focus further (internal and external) research on improving their certainty.
In direct response to Hazelfire’s comment, I think that even if the uncertainty spans only one order of magnitude (he mentioned 2-3, which seems reasonable to me), this could have a really larger effect on resource allocation. The bar for funding is currently 8x relative to GiveDirectly IIRC, which is one order of magnitude, so gaining a better understanding of the certainty could be really important. For instance, we could learn that some interventions which are currently above the bar, are not very clearly so, whereas other interventions which seem to be under the bar but very close to it, could turn out to be fairly certain and thus perhaps a very safe bet.
I think that all of these effects could have a large influence on GiveWell’s recommendations and donors choices, future research, and directly on getting more accurate point-estimates (which could potentially be fairly big).
Hi Jason, thank you for giving a quick response. Both points are very reasonable.
The contest announcement post outlined “several ways an essay could substantively inform the thinking of a panelist”, namely, changing the central estimate or shape of the probability distribution of AGI / AGI catastrophe, or clarifying a concept or identifying a crux.
It would be very interesting to hear if any of the submissions did change any of the panelists’ (or other Open Phil employees’) mind in these ways, and how so. If not, whether because you learned an unanticipated kind of a thing, or because the contest turned out to be less useful than you initially hoped, I think that might also be very valuable for the community to know.
Thanks!
I am not sure that there is actually a disagreement between you and Guy.
If I understand correctly, Guy says that in so far as the funder wants research to be conducted to deepen our understanding of a specific topic, the funders should not judge researchers based on their conclusions about the topic, but based on the quality and rigor of their work in the field and their contributions to the relevant research community.
This does not seem to conflict what you said, as the focus is still on work on that specific topic.
I just want to add, on top of Haydn’s comment to your comment, that:
You don’t need the treatment and the control group to be of the same size, so you could, for instance, randomize among the top 300 candidates.
In my experience, when there isn’t a clear metric for ordering, it is extremely hard to make clear judgements. Therefore, I think that in practice, it is very likely that let’s say places 100-200 in their ranking seem very similar.
I think that these two factors, combined with Haydn’s suggestion to take the top candidates and exclude them from the study, make it very reasonable, and of very low cost.
Thanks for publishing negative results. I think that it is important to do so in general, and especially given that many other group may have relied on your previous recommendations.
If possible, I think you should edit the previous post to reflect your new findings and link to this post.
Thank you for posting this, Paul. I have questions about two different aspects.
In the beginning of your post you suggest that this is “the real thing” and that these systems “could pose an existential risk if scaled up”.
I personally, and I believe other members of the community, would like to learn more about your reasoning.
In particular, do you think that GPT-3 specifically could pose an existential risk (for example if it falls into the wrong hands, or scaled up sufficiently)? If so, why, and what is a plausible mechanism by which it poses an x-risk?
On a different matter, what does aligning GPT-3 (or similar systems) mean for you concretely? What would the optimal result of your team’s work look like?
(This question assumes that GPT-3 is indeed a “prosaic” AI system, and that we will not gain a fundamental understanding of intelligence by this work.)
Thanks again!
I agree with the spirit of this post (and have upvoted it) but I think it kind of obscures the really simple thing going on: the (expected) impact of a project is by definition the cost-effectiveness (also called efficiency) times the cost (or resources).
A 2-fold increase in one, while keeping the other fixed, is literally the same as having the roles reversed.
The question then is what projects we are able to execute, that is, both come up with an efficient idea, and have the resources to execute it. When resources are scarce, you really want to squeeze as much as you can from the efficiency part. Now that we have more resources, we should be more lax, and increase our total impact by pursuing less efficient ideas that still achieve high impact. Right now it starts to look like there’s much more resources ready to be deployed, than projects which are able to absorb them.
Hey Arden, thanks for asking about that. Let me start by also thanking you for all the good work you do at 80,000 Hours, and in particular for the various pieces you wrote that I linked to at 8. General Helpful Resources.
Regarding the key ideas vs old career guide, I have several thoughts which I have written below. Because 80,000 Hours’ content is so central to EA, I think that this discussion is extremely important. I would love to hear your thoughts about this Arden, and I will be glad if others could share their views as well, or even have a separate discussion somewhere else just about this topic.
Content
I think that two important aspects of the old career guide are much less emphasized in the key ideas page: the first is general advice on how to have a successful career, and the second is how to make a plan and get a job. Generally speaking, I felt like the old career guide gave more tools to the reader, rather than only information. Of course, the key ideas page also discusses these issues to some extent, but much less so than the previous career guide. I think that these were very good career advice which could potentially have a large effect on your readers’ careers.
Another important point is that I don’t like, and disagree with the choice of, the emphasis on longtermism and AI safety. Personally, I am not completely persuaded by the arguments for choosing a career by a longtermist view, and even less by the arguments for AI safety. More importantly, I had several conversations with people in the Israeli EA community and with people I gave career consultation to, who were alienated by this emphasis. A minority of them felt like me, and the majority understood it as “all you can meaningfully do in EA is AI safety”, which was very discouraging for them. I understand that this is not your only focus, but people whose first exposure to your website is the key ideas page might get that feeling, if they are not explicitly told otherwise.
Another point is that the “Global priorities” section takes a completely top-to-bottom approach. I do agree that it is sometimes a good approach, but I think that many times it is not. One reason is the tension between opportunities and cause areas which I already wrote about. The other is that some people might already have their career going, or are particularly interested in a specific path. In these situations, while it is true that they can change their careers or realize that they can enjoy a broader collection of careers, it is somewhat irrelevant and discouraging to read about rethinking all of your basic choices. Instead, in these situations it would be much better to help people to optimize their current path towards more important goals. Just to give an example, someone who studies law might get the impression that his choice is wrong and not beneficial, while I believe that if they tried they could find highly impactful opportunities (for example the recently established Legal Priorities Project looks very promising).
I think that these are my major points, but I do have some other smaller reservations about the content (for example I disagree with the principle of maximizing expected value, and definitely don’t think that this is the way it should be phrased as part of the “the big picture”).
Writing Style
I really liked the structure of the previous career guide. It was very straightforward to know what you are about to read and where you can find something, since it was so clearly separated into different pages with clear titles and summaries. Furthermore, its modularity made it very easy to read the parts you are interested in. The key ideas page is much more convoluted, it is very hard to navigate and all of the expandable boxes are not making it easier.
Thanks for spelling out your thoughts, these are good points and questions!
With regards to potentially impactful problems in health. First, you mentioned anti-aging, and I wish to emphasize that I didn’t try to assess it at any point (I am saying this because I recently wrote a post linking to a new Nature journal dedicated to anti-aging). Second, I feel that I am still too new to this domain to really have anything serious to say, and I hope to learn more myself as I progress in my PhD and work at KSM institute. That said, my impression (which is mostly based on conversations with my new advisor) is that there are many areas in health which are much more neglected compared to others, and in particular receive much less attention from the AI and ML community. From my very limited experience, it seems to me that AI and ML techniques are just starting to be applied to problems in public health and related fields, at least in research institutes outside of the for-profit startup scene. I wish I had something more specific to say, and hopefully I will have in a year or two from now.
I completely agree with your view on AI for good being “a robustly good career path in many ways”. I would like mention once more that in order to have a really large impact in it though, one needs to really optimize for that and avoid the trap of lower counterfactual impact (at least in later stages of the career, after they have enough experience and credentials).
It is very hard for me to say where the highest impact position are, and this is somewhat related to the view that I express at the subsection Opportunities and Cause Areas. I imagine that the best opportunities for someone in this field highly depend on their location, connections and experience. For example, in my case it seemed that joining the floods predictions efforts at Google, and the computational healthcare PhD, are significantly better options than the next options in the AI and ML world.
With regards to entering the field, I am super new to this, so I can’t really answer. In any case, I think that entering to the fields of AI, ML and data science is no different for people in EA than others, so I would follow the general recommendations. In my situation, I had enough other credentials (background in math and in programming/cyber-security) to make people believe that I can become productive in ML after a relatively short time (though at least one place did reject me for not having background in ML), so I jumped right in to working on real-world problems rather than dedicating time to studying.
As to estimating impact of a specific role or project, I think it is sometimes fairly straightforward (when the problem is well-defined and the probabilities are fairly high, you can “just do the math” [don’t forget to account for counterfactuals!]), while in other cases it might be difficult (for example more basic research or things having more indirect effects). In the latter case, I think it is helpful to have a rough estimate—understand how large the scope is (how many people have a certain disease or die from it every year?), figure out who is working on the problem and which techniques they use, try to estimate how much of the problem you imagine you can solve (e.g. can we eliminate the disease? [probably not.] how many people can we realistically reach? how expensive is the solution going to be?). All of this together can help you in figuring out the orders of magnitudes you are talking about. Let me give a very rough example of an outcomes of these estimates: A project will take roughly 1-3 years, seems likely to succeed, and if successful, will significantly improve the lives of 200-800 people suffering from some disease every year, and there’s only one other team working on the exact same problem. This sounds great! Changing the variables a little might make it seem much less attractive, for example if only 4 people will be able to pay for the solution (or suffer from it to being with), or if there are 15 other teams working on exactly the same problem, in which case your impact will probably be much lower. One can also imagine projects with lower chances of success, which if successful will have a much larger effect. I tend to be cautious in these cases, because I think that it is much easier to be wrong about small probabilities (I can say more about this).
Let me also mention that it possible to work on multiple projects at the same time, or over a few years, especially if each one consist of several steps in which gain more information and you can re-evaluate them along the way. In such cases, you’d expect some of the projects to succeed, and learn how to calibrate your estimates over time.
Lastly, with regards to your description of my views, that’s almost right, except that I also see opportunities for high impact not only on particularly important problems but also on smaller problems which are neglected for some reason (e.g. things that are less prestigious or don’t have economic incentives). I’d also add that at least in my case in computational healthcare I also intend to apply other techniques from computer science besides AI and ML (but that’s really a different story than AI for good).
This comment already becomes way too long, so I will stop here. I hope that it is somewhat useful, and, again, if someone wants me to write more about a specific aspect, I will gladly do so.
Thanks for writing this! I really like the way you write, which I found both fun and light and, at the same time, highlighting the important parts vividly. I too was surprised to learn that this is the version of utilitarianism Bentham had in his mind, and I find the views expressed in your summary (Ergo) lovely too.
I completely agree, and I too was troubled by this analysis. For me, the bottom line is:
The fact that something is of little-to-no cost, does not mean that its moral value is also little.
Furthermore, in cases like reducing animal suffering, one can both avoid being harmful himself (i.e. become vegan) AND donate to relevant charities, rather than OR.
As someone in the intersection of these subjects I tend to agree with your conclusion, and with your next comment to Arden describing the design-implementation relationship.
Edit 19 Feb 2022: I want to clarify my position, namely, that I don’t see formal verification as a promising career path. As for what I write below, I both don’t believe it is a very practical suggestions, and I am not at all sold on AI safety.
However, while thinking about this, I did come up with a (very rough) idea for AI alignment , where formal verification could play a significant role.
One scenario for AGI takeoff, or for solving AI alignment, is to do it inductively—that is, each generation of agents designs the next generation, which should be more sophisticated (and hopefully still aligned). Perhaps one plan to do achieve this is as follows (I’m not claiming that any step is easy or even plausible):
Formally define what it means for an agent to be aligned, in such a way that subsequent agents designed by this agent are also aligned.
Build your first generation of AI agents (which should be lean and simple as possible, to make the next step easier).
Let a (perhaps computer assisted) human prove that the first generation of AI is aligned in the formal sense of 1.
Then, once you deploy the first generation of agents, it is their job to formally prove that further agents designed by them are aligned as well. Hopefully, since they are very intelligent, and plausibly good at manipulating the previous formal proofs, they can find such proofs. Since the proof is formal, humans can trust and verify it (for example using traditional formal proof checkers), despite not being able to come up with the proof themselves.
This plan has many pitfalls (for example, each step may turn out to be extremely hard to carry out, or maybe your definition of alignment will be so strict that the agents won’t be able to construct any new and interesting aligned agents), however it is a possible way to be certain about having aligned AI.
I meant the difference between using the two, I don’t doubt that you understand the difference between autism and (lack of) leadership. In any case, this was not main point, which is that the word autistic in the title does not help your post in any way, and spreads misinformation.
I do find the rest of the post insightful, and I don’t think you are intentionally trying to start a controversy. If you really believe that this helps your post, please explain why (you haven’t so far).
I don’t understand how you can seriously not understand that difference between the two. Autism is a developmental disorder, which manifests itself in many ways, most of which are completely irrelevant to your post. Whereas being a “terrible leader”, as you call them, is a personal trait which does not resemble autism in almost any way.
Furthermore, the word autistic in the title is not only completely speculative, but also does not help your case at all.
I think that by using that term so explicitly in your title, you spread misinformation, and with no good reason. I ask you to change the title, or let the forum moderators handle this situation.
I would like to reiterate Edo’s answer, and add my perspective.
First and foremost, I believe that one can follow EA perspectives (e.g. donate effectively) AND be kind and helpful to strangers, rather than OR (repeating an argument I made before in another context).
In particular, I personally don’t write giving a couple of dollars in my donation sheet, and it does not affect my EA-related giving (at least not intentionally).
Additionally, they constitute such a little fraction of my other spending, that I don’t notice them financially.
Despite that, I truly believe that being kind to strangers, giving a few coins, or trying to help in other ways, can meaningfully help the other person (even if not as cost-effectively as donating to, say, GiveWell).
I don’t view this and my other donations as means to achieve the exact same goal, but rather as two distinct and non-competing ways to achieve the purpose of making the world better.
Thanks for the response.
I believe this answers the first part, why GPT-3 poses an x-risk specifically.
Did you or anyone else ever write what aligning a system like GPT-3 looks like? I have to admit that it’s hard for me to even have a definition of being (intent) aligned for a system GPT-3, which is not really an agent on its own. How do you define or measure something like this?
Thanks for sharing your computation. This highly resonates with a (very rough) back of the envelope estimate I ran for the cost-effectiveness of the Good Food Institute, the guesstimate model is here https://www.getguesstimate.com/models/16617. The result (which shouldn’t be taken to literally) is $1.4 per ton CO2e (and $0.05-$5.42 for 90% CI).
I can give more details on how my model works, but very roughly I try to estimate the amount of CO2e saved by clean meat in general, and then try to estimate how much earlier will that happen because of GFI. Again, this is very rough, and I’d love any input, or comparison to other models.
Thank you for writing this summary (and conducting this research project)!
I have a question. I am not sure what the standard terminology is, but there are (at least) two different kinds of mental processes: reflexes/automatic response and thoughts or experiences which span longer times. I am not certain which are more related to capacity for welfare, but I guess it is the latter. Additionally I imagine that the experience of time is more relevant for the former. This suggests that maybe the two are not really correlated. Have you thought about this? Is my view of the situation flawed?
Thanks again!
I, for one, think that it is good that climate change was not mentioned. Not necessarily because there are no analogies and lessons to be drawn, but rather because it can more easily be misinterpreted. I think that the kind of actions and risks are much more similar to bio and nuclear, in that there are way less actors and, at least for now, it is much less integrated to day-to-day life. Moreover, in many scenarios, the risk itself is of more abrupt and binary nature (though of course not completely so), rather than a very long and gradual process. I’d be worried that comparing AI safety to climate change would be easily misinterpreted or dismissed by irrelevant claims.