TL;DR: I'm releasing a website that ranks people according to the expected value that they've produced via donations, and allows users to re-rank the list using their own assumptions. I'd like feedback and help making it better. This has been crossposted to the EA forum.
I recently built Impact List (impactlist.xyz), a site which ranks people by the expected value of their donations.
The goal is to make the list popular enough that people care about their ranking on it, so that it influences their decisions about where and how much to donate. A secondary goal is influencing people (whether or not they appear on the list) by making them more aware of the large differences in expected value depending on where money is donated.
Of particular interest to LessWrong readers:
'AI existential risk' is currently the the most cost effective cause area according to the default assumptions. If the site becomes popular and if this remains the case as the research improves, it likely will cause more money to flow into AI safety.
I see the hardest problem here as figuring out how to build a process for combining uncertain evidence via LLM analysis, volunteer researchers, and user feedback in a way that is broadly credible. It feels a bit like trying to build a rationalist version of Wikipedia to handle highly uncertain claims while remaining credible to both rationalists and the broader public.
I'm considering creating a more streamlined AI-xrisk-only version of this site. See below.
The site uses QALYs as the common currency of value (converted to 'lives saved' to be more intuitive), but this doesn't mean it's limited to considering only effects on health/lifespan. The goal is to consider value of any type and convert it to human-QALY-equivalents.
I wrote more about the motivation and theory of this project four years ago in this post.
What Impact List is
The core of the site is a ranking of (currently 73) wealthy philanthropists by their expected impact via donations.
The main list, using the default assumptions
Each donor has their own page showing all their donations (along with the impact of each donation), and aggregate stats by cause area. Users can toggle between 'donations' and 'lives saved' to visualize differences in effectiveness between cause areas. The internal QALY metric is converted into lives saved by defining one life = 80 QALYs.
Expected lives saved by Dustin Moskovitz, by cause area
I've split all donations into 28 cause areas, each with their own 'cost to save a life' value.
Eight of the 28 total cause areas, with the option to edit the cost to save a life for each
There's also a page for every recipient (charitable organization), which can have its own 'cost per life' value if it's more or less effective than the cause area average. Each cause area and customized recipient has a page with a detailed justification for how the math was done to arrive at the cost per life, and where assumptions are explicitly listed.
Three of the nine assumptions for the Global Health cause area
The estimates are of course very approximate, and will be controversial. I make heavy use of LLMs to do this research, encouraging them to synthesize existing analysis. Users who disagree with any of the default parameter values can edit them.
Editing the three parameters for the Global Health cause area
Users can also edit global parameters specifying their time horizon, assumptions about population growth, and the discount rate.
After entering their own assumptions, users can create shareable links. This allows others to see what Impact List would look like given person X's worldview.
I'd eventually like to add worldviews from notable researchers or organizations, but I'm starting with just a handful of worldviews based on simple tweaks to the defaults:
Users can select between suggested worldviews, or create their own
There's also a calculator feature where people can enter their past donations and/or donations they're considering and see where they'd rank on the list, given those donations.
What I'd like help with
If you see something below that you want to help with, comment on this post, DM me, or join the Discord.
Improving the process of effectiveness estimation
Few people are going to dive into the details of the effectiveness estimates and verify for themselves that the research is high quality. So it's not enough that the estimates be excellent. People need to be able to easily understand and trust the process that leads to the estimates.
The current process is LLM-driven, opaque (users can't see the prompts that lead to the estimates and they can't see how feedback is processed), and dependent on the judgments of someone (me) without any reputation as a researcher. I've tried to partly compensate for this by making the justifications explicit about which assumptions are being made, and tried to make the reasoning as clear as possible, but I expect it'll be very hard to generate enough trust using the existing process.
A few things that could help:
Incorporating something like X's Community Notes on each individual assumption[1], or otherwise making user feedback more of a first-class thing.
Getting people/orgs who have already built up trust to publish their worldviews, and doing some blending between them.
Making the estimation process more auditable, including giving details on currently-behind-the-scenes LLM interactions.
Coming up with some council-of-LLMs framework that minimizes special privilege given to any person or group's judgment. Even if LLMs aren't good enough for people to trust something like this now, maybe they will be in N months.
If you have ideas in this area, please share in the comments. I see this as the most important obstacle to the site becoming popular.
Improving the individual effectiveness estimates
The effectiveness estimates for all cause areas are mostly based on research from LLMs (frontier models from Anthropic and OpenAI). I've also done some manual review and back and forth with the LLMs to refine the estimates. The LLMs rely a lot on existing EA research, but the quality of the analysis is pretty uneven and could probably be improved a lot by putting expert human researchers in the loop.
I'd like these estimates to become the best place to look for a synthesis of all effectiveness research that EAs have done (or found).
If you have expertise in cause area effectiveness research, and especially if you're an expert in some particular cause area, it would be great if you'd be willing to help improve the estimates.
Making the site more usable and beautiful
I think the site looks OK now, but not great (especially on mobile). If you have UI skills/taste and want to make the site look better, let me know.
Getting notable people and organizations to publish their own worldviews
As mentioned above, the site allows anyone to make their own customized set of assumptions and create a shareable link to the rankings using these assumptions.
For notable people/orgs, I'd like to add these links as curated options for all users, so people could see Impact List according to Carl Shulman, Rethink Priorities, etc.
If you're a notable researcher or work in a well-respected EA org, and you also want this to happen, please get in touch.
Go-to-market advice
How should we actually make the site popular? In my post introducing this project I talked about some ideas for how to do this, but I'm pretty uncertain about the right path here.
My sense is that the quality of the site should be significantly higher before I try to popularize this beyond EA/rationalist audiences, but it'd be nice to have a distribution plan soon to guide the other site-improvement decisions.
Feedback on the idea of an AI-xrisk-only version of this site
I’m considering making a version of Impact List focused entirely on AI existential risk. This would avoid many of the hardest comparisons between different cause areas and would make the site much simpler. With a narrower scope, I could possibly expand beyond donations sooner and include impact through research, companies, policy, advocacy, etc.
Given that AI and to some extent AI x-risk is dominating more and more of the public discourse and that this trend will likely continue, it seems more plausible to take it for granted that the audience cares about AI safety.
I’m curious whether people think this would be useful, and whether it should be a separate site or if I should just focus on making good cause-specific features within Impact List.
Suggestions for high-impact people currently not on the list
Everyone currently on the list is well known, but that's not a requirement.Feel free to nominate yourself (if your donations can be proven to the satisfaction of a skeptical reader) or someone else I missed.
Improvements to the underlying data model and estimation framework
See Appendix A below for details on how this works. I've tried to strike a balance between simplicity and expressive power, but I'm not sure I've picked the right tradeoff or whether I'm on the Pareto frontier.
Funding may be available
A couple of funders have offered to give me grants for this work. I haven't accepted any yet, but if you want to work on Impact List and you want to be paid, these grantors may be willing to make that happen.
Getting involved
The project has a Discord. You can also submit pull requests via GitHub. Feel free to DM me on this forum.
Above I've listed what I think I most need, but I may be wrong. I'm happy for all sorts of help or feedback.
Thanks to Austin Chen, Ryan Kidd, Kim Korte, Nathan Young, plex, and Laszlo Treszkai for discussion and feedback about the site.
Appendix A: How effectiveness is modeled
The details are on GitHub. This is a simplified overview.
Each cause area or charitable organization has one or more 'effects', which define how money translates into QALYs. There are two types of effects, standard effects and population effects, and each effect has several parameters.
Standard Effects
The most intuitive type. The more money you spend, the more impact you get. The parameters are:
Cost per QALY: How much money it costs to produce one QALY
Start time: Years after donation before the effect begins
Duration: How many years the effect lasts (benefits are assumed to be evenly distributed)
Population Effects
Effects where money donated changes the probability of some event. This is used to model x-risks, pandemics, etc. The parameters are:
Cost per microprobability: How much it costs to change the likelihood of an event by 1 in 1,000,000.
Population fraction affected: What fraction of the population is impacted if the event occurs
Welfare change per person per year: estimated welfare change per affected person per year, in QALYs. Positive values mean the funded work makes a bad event less likely or a good event more likely; negative values mean it does the opposite.
Start time / Duration: Same meaning as for the standard effects
Overrides and multipliers
A recipient organization which is part of some cause area can override the parameters of that cause area, to indicate that the recipient is especially effective or ineffective.
Recipient organizations can also have their parameters defined as a multiple of a parameter of the cause area it belongs to. The site's UI currently hides the ability to express parameters in terms of multiples for simplicity, though I'd like to re-add this to the UI at some point.
Time
The model is capable of handling time-based effects, which allows specifying different levels of effectiveness for a cause/recipient depending on when the donation was made. For instance a donation to MIRI today may have a very different impact than a donation to MIRI in 2012. The current version of the site doesn't make use of time-based effects, but they're available and documented on GitHub.
Motivation for these choices
This model gives each effect a 'shape' over time that we can integrate over to get the total value. This allows assumptions about the future population, discount rate, and time horizon to be separated out from cause-specific assumptions.
I could have made these shapes more complex by having options for exponential or linear decay of effects, but I think it's not worth the extra complexity. Every effect-shape in this model is a simple rectangle.
Appendix B: Current limitations and uncertainties
I currently only consider impact via donations. I'd eventually like to include all of a person's positive and negative externalities, including via businesses they've created, via politics, etc.
The site's estimates are very uncertain, and this could probably be quantified and communicated better to non-EAs, possibly along with some explanation of why uncertain estimates are still valuable.
As seen in Appendix A, the underlying model is fairly simple.
Although the site infrastructure supports it, I don't currently distinguish average effectiveness from marginal effectiveness or account for effectiveness changes over time.
I don't try to measure fungibility or crowding-out/in effects.
I use LLMs to gather donation data for each person using this skill, and it seems to work fairly well, but not perfectly.
Not all donations are public, so the data is missing some donations. However because it's usually in the interest of both the donor and the recipient to publicize donations (the donor gets status and to raise the profile of a cause they care about, the recipient gets attention from other potential donors), my guess is that this effect isn't huge.
I don't count pledges, but I fully count donations that a person makes to a charitable foundation or fund that they control. In some cases this seems like what we want, but in others it probably overstates the donor's impact. This could be addressed by giving these foundations customized lower effectiveness estimates, but I haven't done that yet.
The cause area breakdown for multi-cause recipients like The Gates Foundation is based on estimates from LLMs, not on classifying all of the grants that they've made.
I'm unsure whether the 28 cause areas that I've picked are the best way to carve up the space.
TL;DR: I'm releasing a website that ranks people according to the expected value that they've produced via donations, and allows users to re-rank the list using their own assumptions. I'd like feedback and help making it better. This has been crossposted to the EA forum.
I recently built Impact List (impactlist.xyz), a site which ranks people by the expected value of their donations.
The goal is to make the list popular enough that people care about their ranking on it, so that it influences their decisions about where and how much to donate. A secondary goal is influencing people (whether or not they appear on the list) by making them more aware of the large differences in expected value depending on where money is donated.
Of particular interest to LessWrong readers:
The site uses QALYs as the common currency of value (converted to 'lives saved' to be more intuitive), but this doesn't mean it's limited to considering only effects on health/lifespan. The goal is to consider value of any type and convert it to human-QALY-equivalents.
I wrote more about the motivation and theory of this project four years ago in this post.
What Impact List is
The core of the site is a ranking of (currently 73) wealthy philanthropists by their expected impact via donations.
Each donor has their own page showing all their donations (along with the impact of each donation), and aggregate stats by cause area. Users can toggle between 'donations' and 'lives saved' to visualize differences in effectiveness between cause areas. The internal QALY metric is converted into lives saved by defining one life = 80 QALYs.
I've split all donations into 28 cause areas, each with their own 'cost to save a life' value.
There's also a page for every recipient (charitable organization), which can have its own 'cost per life' value if it's more or less effective than the cause area average. Each cause area and customized recipient has a page with a detailed justification for how the math was done to arrive at the cost per life, and where assumptions are explicitly listed.
The estimates are of course very approximate, and will be controversial. I make heavy use of LLMs to do this research, encouraging them to synthesize existing analysis. Users who disagree with any of the default parameter values can edit them.
Users can also edit global parameters specifying their time horizon, assumptions about population growth, and the discount rate.
After entering their own assumptions, users can create shareable links. This allows others to see what Impact List would look like given person X's worldview.
I'd eventually like to add worldviews from notable researchers or organizations, but I'm starting with just a handful of worldviews based on simple tweaks to the defaults:
There's also a calculator feature where people can enter their past donations and/or donations they're considering and see where they'd rank on the list, given those donations.
What I'd like help with
If you see something below that you want to help with, comment on this post, DM me, or join the Discord.
Improving the process of effectiveness estimation
Few people are going to dive into the details of the effectiveness estimates and verify for themselves that the research is high quality. So it's not enough that the estimates be excellent. People need to be able to easily understand and trust the process that leads to the estimates.
The current process is LLM-driven, opaque (users can't see the prompts that lead to the estimates and they can't see how feedback is processed), and dependent on the judgments of someone (me) without any reputation as a researcher. I've tried to partly compensate for this by making the justifications explicit about which assumptions are being made, and tried to make the reasoning as clear as possible, but I expect it'll be very hard to generate enough trust using the existing process.
A few things that could help:
If you have ideas in this area, please share in the comments. I see this as the most important obstacle to the site becoming popular.
Improving the individual effectiveness estimates
The effectiveness estimates for all cause areas are mostly based on research from LLMs (frontier models from Anthropic and OpenAI). I've also done some manual review and back and forth with the LLMs to refine the estimates. The LLMs rely a lot on existing EA research, but the quality of the analysis is pretty uneven and could probably be improved a lot by putting expert human researchers in the loop.
I'd like these estimates to become the best place to look for a synthesis of all effectiveness research that EAs have done (or found).
If you have expertise in cause area effectiveness research, and especially if you're an expert in some particular cause area, it would be great if you'd be willing to help improve the estimates.
Making the site more usable and beautiful
I think the site looks OK now, but not great (especially on mobile). If you have UI skills/taste and want to make the site look better, let me know.
Getting notable people and organizations to publish their own worldviews
As mentioned above, the site allows anyone to make their own customized set of assumptions and create a shareable link to the rankings using these assumptions.
For notable people/orgs, I'd like to add these links as curated options for all users, so people could see Impact List according to Carl Shulman, Rethink Priorities, etc.
If you're a notable researcher or work in a well-respected EA org, and you also want this to happen, please get in touch.
Go-to-market advice
How should we actually make the site popular? In my post introducing this project I talked about some ideas for how to do this, but I'm pretty uncertain about the right path here.
My sense is that the quality of the site should be significantly higher before I try to popularize this beyond EA/rationalist audiences, but it'd be nice to have a distribution plan soon to guide the other site-improvement decisions.
Feedback on the idea of an AI-xrisk-only version of this site
I’m considering making a version of Impact List focused entirely on AI existential risk. This would avoid many of the hardest comparisons between different cause areas and would make the site much simpler. With a narrower scope, I could possibly expand beyond donations sooner and include impact through research, companies, policy, advocacy, etc.
Given that AI and to some extent AI x-risk is dominating more and more of the public discourse and that this trend will likely continue, it seems more plausible to take it for granted that the audience cares about AI safety.
I’m curious whether people think this would be useful, and whether it should be a separate site or if I should just focus on making good cause-specific features within Impact List.
Suggestions for high-impact people currently not on the list
Everyone currently on the list is well known, but that's not a requirement. Feel free to nominate yourself (if your donations can be proven to the satisfaction of a skeptical reader) or someone else I missed.
Improvements to the underlying data model and estimation framework
See Appendix A below for details on how this works. I've tried to strike a balance between simplicity and expressive power, but I'm not sure I've picked the right tradeoff or whether I'm on the Pareto frontier.
Funding may be available
A couple of funders have offered to give me grants for this work. I haven't accepted any yet, but if you want to work on Impact List and you want to be paid, these grantors may be willing to make that happen.
Getting involved
The project has a Discord. You can also submit pull requests via GitHub. Feel free to DM me on this forum.
You can keep up with the project on Twitter.
Above I've listed what I think I most need, but I may be wrong. I'm happy for all sorts of help or feedback.
Thanks to Austin Chen, Ryan Kidd, Kim Korte, Nathan Young, plex, and Laszlo Treszkai for discussion and feedback about the site.
Appendix A: How effectiveness is modeled
The details are on GitHub. This is a simplified overview.
Each cause area or charitable organization has one or more 'effects', which define how money translates into QALYs. There are two types of effects, standard effects and population effects, and each effect has several parameters.
Standard Effects
The most intuitive type. The more money you spend, the more impact you get. The parameters are:
Population Effects
Effects where money donated changes the probability of some event. This is used to model x-risks, pandemics, etc. The parameters are:
Overrides and multipliers
A recipient organization which is part of some cause area can override the parameters of that cause area, to indicate that the recipient is especially effective or ineffective.
Recipient organizations can also have their parameters defined as a multiple of a parameter of the cause area it belongs to. The site's UI currently hides the ability to express parameters in terms of multiples for simplicity, though I'd like to re-add this to the UI at some point.
Time
The model is capable of handling time-based effects, which allows specifying different levels of effectiveness for a cause/recipient depending on when the donation was made. For instance a donation to MIRI today may have a very different impact than a donation to MIRI in 2012. The current version of the site doesn't make use of time-based effects, but they're available and documented on GitHub.
Motivation for these choices
This model gives each effect a 'shape' over time that we can integrate over to get the total value. This allows assumptions about the future population, discount rate, and time horizon to be separated out from cause-specific assumptions.
I could have made these shapes more complex by having options for exponential or linear decay of effects, but I think it's not worth the extra complexity. Every effect-shape in this model is a simple rectangle.
Appendix B: Current limitations and uncertainties
Thanks to Nathan Young for the suggestion and for discussions around this idea.