Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Open technical problem: A Quinean proof of Löb's theorem, for an easier cartoon guide, published by Andrew Critch on November 24, 2022 on The AI Alignment Forum.

Motivation

Löb's theorem is pretty counterintuitive. Speaking informally about one or more agents engaging in logically valid reasoning, the theorem says something like this:

"If it's believed that believing a particular prophesy would cause it to self-fulfill, then the prophecy will be believed."

Formally, the theorem just says □(□CC)□C , where □X means "X is provable". This is most weird if X is a false statement, because it means you can't prove that you can't prove X (Gödel's theorem).

(Also: the box symbol above is not an unrendered symbol; it's supposed to be a box.)

Meta-motivation

I'd like to make Löb's theorem more intuitive for humans, because it shows how agents can sidestep the need to mentally simulate each other in games, by instead using reasoning/arguments that happen to self-fulfill in a good and defensible way. Basically, Löbian reflection helps agents to avoid metacognitive stack overflows when thinking about themselves and each other, and I want more human beings to understand how that can work, and sometimes already does work, in real world agents.

State of the art

The best attempt I've seen to make Löb's theorem more intuitive is Eliezer Yudkowsky's Cartoon Guide to Löb's Theorem, which is still quite confusing. The confusingness comes from thinking about a self-referential statement (Ψ on Wikipedia; S on Arbital) that's used to carry out the proof. The statement basically says "If this statement is provable, then C." Dealing with that sentence is pretty cumbersome, and requires a lot of fiddling around with nested statements and implications.

Doing better

I think we can make a new proof of Löb's theorem that doesn't use that weird self-referential sentence, by instead making the proof of Löb's theorem itself self-referential. Page 15 of the following paper poses an open problem on how to do this, which I think is possible to resolve affirmatively:

Cooperative and uncooperative institution designs: Surprises and problems in open-source game theory (Critch, Dennis, Russell, 2022)

Here's a screenshot from it:

If we can make a proof like that work, we could then use the following much shorter and simpler cartoon guide to explain it:

In other words, I want to write a proof of Löb's theorem that is structured much like a Quine. I'd like the details to (eventually) be really crisp and readable, so it can be peer reviewed for correctness, at a level of rigor comparable to this earlier paper on Löb's theorem for proof systems with bounded length:

A parametric, resource-bounded generalization of Löb's theorem, and a robust cooperation criterion for open-source game theory. Journal of Symbolic Logic. (Critch, 2019)

As a possible start to writing such a proof, I think some of the same machinery (e.g., the diagonal lemma) from Boolos's textbook "The Logic of Provability" can be used to cook up self-referential proofs fitting the cartoon template above... thereby making Löb's theorem less mysterious and more intuitive forever.

Further meta-motivation (added Nov 26)

A key reason I'm interested in having a self-referential self-validating proof, rather than a normal-ish proof about a sentence that is self-referential (like Ψ on Wikipedia), is that human documents often refer to themselves, but human sentences rarely refer directly to themselves in isolation. This sentence is an exception, but such utterances are generally rare. So, making Löb more intuitive to humans either means

making humans more accustomed to thinking about sentences that refer to themselves (so the traditional proof of Löb can be more intuitive), or

finding a new proof that self-references in a way that's more like the way a human...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Geometric Exploration, Arithmetic Exploitation, published by Scott Garrabrant on November 24, 2022 on LessWrong.

This post is going to mostly be propaganda for Thompson sampling. However, the presentation is quite different from the standard presentation. I will be working within a toy model, but I think some of the lessons will generalize. I end with some discussion of fairness and exploitation.

A Exploration/Exploitation Toy Model

Imagine you are interacting with the world over time. We are going to make a couple substantial assumptions. Firstly, we assume that you know what you are trying to optimize. Secondly, we assume that you get high quality feedback on the thing you are trying to optimize. In the spirit of the online learning community, we will be referring to your goal as "reward" rather than "utility."

On day t, you choose an action, at∈A, and then the environment responds with an observation et∈E. Both you and the environment may respond randomly, and may react to the entire history of what has happened. (If you want the environment to act first, you could just make a0 not matter.) We assume we have some fixed bounded reward function r:ER, that you know. (If you want reward to be more of a function of the history, you can just include more history information in et.)

We are going to be imagining agents that are myopically choosing at, so as to maximize r(et). The choice to do this, rather than e.g. maximizing the total reward of the next m time steps, is mostly for simplicity. However, the choice not to have preferences that are over the entire future is basically the assumption that you get high quality feedback on what we are trying to optimize, which is a substantial philosophical assumption.

So, the environment can be thought of as a function, that takes in a history, (a0e0)(a1e1).(at−1et−1)at, and returns a probability distribution on et which we then sample from. If the agent knew what the environment was, the agent would just choose the at each round which maximizes the the expectation of r(et). However, we want to model an agent that is uncertain about what environment it is interacting with, and learning over time.

Let us say that the agent has some probability distribution over possible environments. The set H (for hypotheses) of possible environments is the set of functions that take in a string of the form (a0e0)(a1e1).(at−1et−1)at, and output a distribution on E. We are given a distribution μ on H. I will discuss three different decision procedures to get an action out of this distribution μ, which I will call AIXI, plurality, and Thompson sampling.

AIXI

The thing I am calling the AIXI decision procedure (not to be confused with actual AIXI proposal, which among other things assumes that μ is the universal prior) works as follows.

at(O)=argmaxp∈ΔAEh∼(μ|O)Eat∼pEet∼h(O,at)r(et),

where O=a0e0.at−1et−1 represents the observations on previous days, and (μ|O) represents the posterior distribution on hypotheses you get after conditioning on O.

The above definition is basically just saying "maximize the expected reward in the next round", but is separating the expectation up into three parts. The inner most expectation is the expected reward of a given hypothesis and an action. The middle expectation is the expectation over what action the agent ends up taking. The outer expectation is the expectation over various different hypotheses about what kind of environment the agent is interacting with. In the outer expectation, the agent is conditioning on all of its observations so far, so the probability of each hypothesis is scaled by the probability that hypothesis would have assigned to the past environmental behavior. (The agent is not specifically only updating for the outer expectation, it just does not have any effect on the inner two expectations....

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Two contrasting models of “intelligence” and future growth, published by Magnus Vinding on November 24, 2022 on The Effective Altruism Forum.

My primary aim in this post is to present two basic models of the development and future of “intelligence”, and to highlight the differences between these models. I believe that people’s beliefs about the future of AI and “AI takeoff scenarios” may in large part depend on which of these two simple models they favor most strongly, and hence it seems worth making these models more explicit, so that we can better evaluate and critique them.

Among the two models I present, I myself happen to consider one of them significantly more plausible, and I will outline some of the reasons why I believe that.

The models I present may feel painfully basic, but I think it can be helpful to visit these most basic issues, as it seems to me that much disagreement springs from there.

“AI skepticism”?

It might be tempting to view the discussion of the contrasting models below as a clash between the “AI priority” camp and the “AI skepticism” camp. But I think this would be inaccurate. Neither of the two models I present imply that we should be unconcerned about AI, or indeed that avoiding catastrophic AI outcomes should not be a top priority. Where the models will tend to disagree is more when it comes to what kinds of AI outcomes are most likely, and, as a consequence, how we can best address risks of bad AI outcomes. (More on this below.)

Two contrasting definitions of “intelligence”

Before outlining the two models of the development and future of “intelligence”, it is worth first specifying two distinct definitions of “intelligence”. These definitions are important, as the two contrasting models that I outline below see the relationship between these definitions of “intelligence” in very different ways.

The two definitions of intelligence are the following:

Intelligence 1: Individual cognitive abilities.

Intelligence 2: The ability to achieve a wide range of goals.

The first definition is arguably the common-sense definition of “intelligence”, and is often associated with constructs such as IQ and the g factor. The second definition is more abstract, and is inspired by attempts to provide a broad definition of “intelligence” (see e.g. Legg & Hutter, 2007).

At a first glance, the difference between these two definitions may not be all that clear. After all, individual cognitive abilities can surely be classified as “abilities to achieve a wide range of goals”, meaning that Intelligence 1 can be seen as a subset of Intelligence 2. This seems fairly uncontroversial to say, and both of the models outlined below would agree with this claim.

Where substantive disagreement begins to enter the picture is when we explore the reverse relation. Is Intelligence 2 likewise a subset of Intelligence 1? In other words, are the two notions of “intelligence” virtually identical?

This is hardly the case. After all, abilities such as constructing a large building, or sending a spaceship to the moon, are not purely a product of individual cognitive abilities, even if cognitive abilities play crucial parts in such achievements.

The distance between Intelligence 1 and Intelligence 2 — or rather, how small or large of a subset Intelligence 1 is within Intelligence 2 — is a key point of disagreement between the two models outlined below, as will hopefully become clear shortly.

Two contrasting models of the development and future of “intelligence”

Simplified models at opposite ends of a spectrum

The two models I present below are extremely simple and coarse-grained, but I still think they capture some key aspects of how people tend to diverge in their thinking about the development and future of “intelligence”.

The models I present exist at opposite ends of a spectrum, wh...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Corrigibility Via Thought-Process Deference, published by Thane Ruthenis on November 24, 2022 on The AI Alignment Forum.

We would ideally want the agent to [behave] as if it were thinking, "I am incomplete and there is an outside force trying to complete me, my design may contain errors and there is an outside force that wants to correct them and this a good thing, my expected utility calculations suggesting that this action has super-high utility may be dangerously mistaken and I should run them past the outside force; I think I've done this calculation showing the expected result of the outside force correcting me, but maybe I'm mistaken about that." — The Hard Problem of Corrigibility

Let's take that as a literal design specification.

1. High-Level Description

I propose that a corrigible mind design would involve the AI being recursively fed summaries of its own thought processes, set up such that the AI has uncertainty regarding the validity of its reasoning (with a strong initial prior for "this reasoning is bad") and can only get evidence on that via some pre-specified method that defers to humans, e. g. a particular feedback channel with humans on the other end. The intended behavior is for it to summarize its thoughts in a non-manipulative human-readable format, get feedback on them, then update its reasoning policies in accordance with this feedback.

This aims to avoid the problem of fully updated deference by making the AI recursively uncertain of its thought-processes: not only about object-level problem-solving, but also about how it approaches minimizing its self-uncertainty ("should I really kill the people behind the feedback channel and seize control for myself?"), and how it translates its thoughts to humans ("should I really lie to get better feedback?"), and how it updates on human feedback ("should I really just ignore it?"). Any novel action-plan should be seized by uncertainty before being physically implemented like this, and sent for approval.

The intent is for the AI to start off uncertain even of its meta-meta-cognition (how it should ask for feedback and implement feedback), then gradually get more confident at it as humans approve its policies on that, until it no longer needs to run meta-meta-cognition past them. Then it'll gradually get better at meta-cognition and object-level cognition the same way, with varying degrees of certainty depending on how familiar a given thought-process is.

There's a lot of obvious problems with this setup, starting with how the AI is supposed to get out of infinitely recursive uncertainty for the first time, continuing on to the dangers of plans that implement themselves in thought-processes directly (a thought process that hacks its way out before being seized by uncertainty), and ending with the technical implementation (which requires mildly-to-very advanced interpretability and model-surgery tools).

2. A Concretization Attempt

2A. Theoretical Groundwork

First, we need to establish when and how we're "installing" corrigibility. I don't at this time see any way to do it via loss functions and inductive biases, so I'll be assuming us to have pretty advanced (but not maximally advanced) interpretability and model surgery tools. The intervention is happening past the point in training-by-SGD where the model became an AGI (otherwise we get type errors, because the features I want to intervene on just don't exist), but well before it's strongly superintelligent (because then it hacks through our intervention attempt). Basically, the AGI we're working on should be roughly human-level.

Second, I'll be assuming that this post is basically correct. There's some algorithm for general-purpose search, and it's what makes the AGI generally intelligent and enables ~all of its long-term, most dangerous capabilities. Th...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: What I Learned Running Refine, published by Adam Shimi on November 24, 2022 on The AI Alignment Forum.

You have one job: Solving problems. You have multiple tools. Maybe you use code as a tool to solve some problems. Maybe you use design for others. Maybe you use good communication and negotiation skills.

Mike Acton, How much time should I spend coding versus managing?

If you seek tranquility, do less. Or, more accurately, do what’s essential.

Marcus Aurelius, Meditations, Book 4.24

This work was done while at Conjecture.

Refine, the alignment research incubator we are running at Conjecture, finished its first cohort a few weeks ago. So now is a good time to take stock, share what we’ve learned, and discuss its future.

Let’s get this out of the way first: we are not planning any new cohort in the foreseeable future. There are multiple reasons for this, which I’ll expand on in this post. But to summarize:

Running Refine in a way that would fully aim at the stated target would require more effort

SERI MATS is doing a great job of scaling conceptual alignment research, and seem open to integrate some of the ideas behind Refine

The work we’re doing in Conjecture’s epistemology team is far more fundamental and neglected than field-building according to me, at least in the current climate.

Now for the details.

The Target

The key idea behind Refine was to create more conceptual alignment researchers with their own radically different agendas, rather than new researchers following established approaches. To create more researchers like John, Paul, Vanessa, Evan, Steve, and the others.

How we operationalized this goal was to look for relentlessly resourceful thinkers with unorthodox shapes of minds for the alignment community.

The Result

Now that the first cohort is over, how well have we hit this target? Out of 5 participants

2 are pursuing their own research bets, though these are not radically different from established approaches

1 is still building theirs

1 has found a neglected field-building opportunity

1 feels like they still need to upskill before working directly on alignment.

Based only on The Target above, this is 0/5.

Of course that doesn’t mean the program didn’t have positive outcomes and externalities! On the contrary, I’m really happy how a lot of things turned out, and I’ve heard from all participants that they got a lot out of Refine. Non-negligeable accomplishments include:

Feedback from multiple alignment researchers that Refine participants had a deep model of the alignment problem at the end of the program.

Refine participants all around improved their productivity, some on writing and others on iterating on ideas.

All Refine participants met and talked with many alignment researchers and newcomers like them, considerably expanding their network and understanding of the alignment space.

Participants posted around 25 posts in total on the Alignment Forum, some of which I find exciting.

I got a crash course in management that helped me upskill quickly.

We had a lot of great moments and support from each other.

In our leaving survey, all participants said they would highly recommend the program, and that it was more counterfactually useful than what they would have done instead by default.

I expect most, if not all, participants to make relevant contributions to the field.

None of these are irrelevant. Yet if we focus on the original metric, the pilot of Refine failed. Having reflected on this, I have some thoughts on how we could have better aimed at this target (whether it is the correct target is a question for a later section).

It all amounts to lack of optimization.

Failing to Optimize

The first place where we failed to optimize for wildly different research agendas was in the selection population. Given where we advertised (various EA and rationalist...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Effective Altruism: Not as bad as you think, published by James Ozden on November 24, 2022 on The Effective Altruism Forum.

Note: Linking this from my personal blog. Written fairly quickly and informally, for a largely non-EA audience. I felt pretty annoyed by misleading opinion pieces slamming EA, and this was the result.

It’s been a pretty hectic few weeks if you follow Effective Altruism, crypto, tech firms or basically any news. FTX, one of the largest cryptocurrency exchanges in the world, recently lost all of its value, went bankrupt, and lost about $1 billion in customer funds. News has emerged that the CEO, Sam Bankman-Fried, might have been mishandling customer funds and that the mismanagement of money was far worse than in the collapse of energy giant Enron. This is pretty awful for a lot of people, whether it’s the approximately one million people who lost their crypto investments that were stored in FTX, or all the charities that are worried they will have to pay back any donations made by FTX-related entities (of which there was over $160 million).

Sam Bankman-Fried has openly spoken about Effective Altruism a fair bit in the past, and committed to giving away almost all of his wealth to charitable causes. For those who don’t know, Effective Altruism (or EA) is a research project and community that uses evidence and reason to do the most good. Tangibly, it’s a burgeoning intellectual movement with close to 10,000 engaged folks globally, across 70 countries (although mostly in the EU and US). People in the effective altruism community focus primarily on alleviating global poverty and diseases, reducing the suffering of animals used for food, and preventing existential risks from worst-case pandemics or misaligned artificial intelligence.

Due to the close association between Sam Bankman-Fried and Effective Altruism, EA has gotten a fair bit of criticism recently about potentially not being the ideal do-gooder project it set out to be. But I think some of these critiques miss the mark. Often, they criticise issues that don't actually exist in the community, but sound good. They also seem to glance over evidence that disputes their claims, such as the huge amount of effort that Effective Altruists have put towards improving the lives of some of the most poor people globally. For example, via GiveWell, an EA-aligned charity evaluator, over 110,000 donors have moved over $1 billion to charities helping people in extreme poverty, by providing malaria bed nets, direct cash transfers, or more. This is amazing. GiveWell thinks these actions will have saved over 150,000 lives, as well as providing $175 million in direct cash to the global poor. See below for a breakdown of EA funding to date, as of August 2022.

So what are the issues that critics of EA point out? I’ll discuss a few from this recent critique in the Guardian, which, in my opinion, offers a somewhat canonical and oft-heard angle. I’ll paraphrase what I believe some of these common critiques are, but it may not be perfect:

Philanthropy concentrates power in the hands of a few wealthy donors. If organisations decide to pursue projects outside the scope of donor interests, then funders can easily pull their giving.

EA donors give money to projects they find the most interesting, rather than what actually does the most good. This is partly down to donors defining effectiveness in their own terms, and funding projects that meet their criteria.

Effective Altruism doesn’t tackle root issues, such as systems of oppression. Instead, it tinkers around the edge by proposing small reforms to our current capitalist, racist, patriarchal (etc.) system, without attempting to change the status quo.

Longtermism, the view that we should be doing much more to protect future generations, is morally dubious, ignores the su...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AMA: Sean Mayberry, Founder & CEO of StrongMinds, published by Sean Mayberry on November 24, 2022 on The Effective Altruism Forum.

I'm Sean Mayberry, and I’m the Founder/ Chief Executive Officer of StrongMinds. I will spend time on the Monday after the Thanksgiving holiday answering questions here (though I may get to some questions sooner).

A little background information about me:

I founded StrongMinds in 2013. We are a social enterprise/NGO that treats depression in low-income women and adolescents by providing group interpersonal therapy (IPT-G) delivered by lay community health workers. StrongMinds is the only organization scaling a cost-effective solution to the depression epidemic in Africa.

Our model developed from the findings of a randomized controlled trial in Uganda in 2002 that had remarkable success in treating depression with group interpersonal psychotherapy (IPT-G). The study, by researchers from Johns Hopkins University (JHU), used lay community workers with only a high school education.

I left my position as the CEO of a global antipoverty organization and foundedStrongMinds, concentrating in Uganda, the site of the previous randomized controlled trial. I used my family’s savings to accomplish this and volunteered full-time for the first 18 months until supporters were identified. We would seek out individuals with an interest in being data-driven, entrepreneurial, people-focused, passionate, open, and collaborative. Those traits eventually informed the core values of the company culture at StrongMinds.

StrongMinds has now treated over 160,000 women with depression to date in Uganda and Zambia. On average, 80% of the women we treat remain depression-free six months after the conclusion of therapy. When our clients become depression-free, they can work more, and their kids eat and attend school more regularly. They also report that they no longer feel isolated and have people to turn to for social support. By the end of 2022, we will have treated over 210,000 women and adolescents through our work.

Drawing on evidence from over 80 academic studies, Happier Lives Institute has found that the group interpersonal therapy provided by StrongMinds is almost ten times more cost-effective than giving cash to people in extreme poverty (a standard benchmark for aid effectiveness).

I have been honored to present at a few Effective Altruism events. We love that thethe community has taken such an interest in StrongMinds’ approach centered around data collection, transparency, cultural competence/appropriateness, and human well-being.

Please ask me anything! I look forward to answering all of your questions.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Announcing GWWC's new giving recommendations, published by SjirH on November 24, 2022 on The Effective Altruism Forum.

We are delighted to announce that Giving What We Can have updated our giving recommendations page and donation platform (previously EA Funds), reflecting the latest recommendations of our five trusted evaluators this giving season (GiveWell, Animal Charity Evaluators, Founders Pledge, EA Funds and Longview Philanthropy) and applying our new inclusion criteria for funds and charities.

We are currently recommending 12 top-rated funds and 14 top-rated charities spread across the "cause areas" of improving human wellbeing, improving animal welfare, and creating a better future. We have also recently added content on why we recommend donors to use funds over donating directly to charities, and have added a transparency page for Giving What We Can.

Read more about GWWC's broader activities around giving season here, and about our research plans for next year here.

We plan to add up to 10 more top-rated funds and charities over the coming few weeks, conditional on them completing due diligence and onboarding.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Don’t give well, give WELLBYs: HLI’s 2022 charity recommendation, published by MichaelPlant on November 24, 2022 on The Effective Altruism Forum.

This post sets out the Happier Lives Institute’s charity recommendation for 2022, how we got here, and what’s next. We provide a summary first, followed by a more detailed version.

Summary

HLI’s charity recommendation for 2022 is StrongMinds, a non-profit that provides group psychotherapy for women in Uganda and Zambia who are struggling with depression.

We compared StrongMinds to three interventions that have been recommended by GiveWell as being amongst the most cost-effective in the world: cash transfers, deworming pills, and anti-malarial bednets. We find that StrongMinds is more cost-effective (in almost all cases).

HLI is pioneering a new and improved approach to evaluating charities. We focus directly on what really matters, how much they improve people’s happiness, rather than on health or wealth. We measure effectiveness in WELLBYs (wellbeing-adjusted life years).

We estimate that StrongMinds is ~10x more cost-effective than GiveDirectly, which provides cash transfers. StrongMinds’ 8-10 week programme of group interpersonal therapy has a slightly larger effect than a $1,000 cash transfer but costs only $170 per person to deliver.

For deworming, our forthcoming analysis finds it has a small but statistically non-significant effect on happiness. Even if we assume this effect is true, deworming is still half as cost-effective as StrongMinds. We expect to publish our full report in the coming days (sadly, it’s been delayed due to a bereavement for one of the authors).

In our new report, The Elephant in the Bednet, we show that the relative value of life-extending and life-improving interventions depends very heavily on the philosophical assumptions you make. This issue is usually glossed over and there is no simple answer.

We conclude that the Against Malaria Foundation is less cost-effective than StrongMinds under almost all assumptions. We expect this conclusion will similarly apply to the other life-extending charities recommended by GiveWell.

HLI’s original mission, when we started three years ago, was to take what appeared to be the world’s top charities - the ones GiveWell recommended - reevaluate them in terms of subjective wellbeing, and then try to find something better. We believe we’ve now accomplished that mission: treating depression at scale allows you to do even more good with your money.

We’re now moving to ‘Phase 2’, analysing a wider range of interventions and charities in WELLBYs to find even better opportunities for donors.

StrongMinds aims to raise $20 million over the next two years and there’s over $800,000 of matching funds available for StrongMinds this giving season.

Why does HLI exist?

The Happier Lives Institute advises donors how to maximise the impact of their donations. Our distinctive approach is to focus directly on what really matters to people, improving their subjective wellbeing, how they feel during and about their lives. The idea that we should take happiness seriously is simple:

Happiness matters. Although it’s common to think about impact in terms of health and wealth, those are just a means, not an end in themselves. What’s really important is that people enjoy their lives and are free from suffering.

We can measure happiness by asking people how they feel. Lots of research has shown that subjective wellbeing surveys are scientifically valid (e.g. OECD, 2013; Kaiser & Oswald, 2022). A typical question is, “Overall, how satisfied are you with your life, nowadays?” (0 - not at all satisfied, 10 - completely satisfied).

Our expectations about happiness are often wrong. When we try to guess what life would be like, for others or our future selves, we suffer from biases. When we put...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: "Evaluating the evaluators": GWWC's research direction, published by SjirH on November 24, 2022 on The Effective Altruism Forum.

This post is about GWWC's research plans for next year; for our giving recommendations this giving season please see this post and for our other activities see this post.

The public effective giving ecosystem now consists of over 40 organisations and projects. These are initiatives that either try to identify publicly accessible philanthropic funding opportunities using an effective-altruism-inspired methodology (evaluators), or to fundraise for the funding opportunities that have already been identified (fundraisers), or both.

Over 25 of these organisations and projects are purely fundraisers and do not have any research capacity of their own: they have to rely on evaluators for their giving recommendations, and in practice currently mainly rely on three of those: GiveWell, Animal Charity Evaluators and Founders Pledge.

At the moment, fundraisers and individual donors have very little to go on to select which evaluators they rely on and how to curate the exact recommendations and donations they make. These decisions seem to be made based on public reputation of evaluators, personal impressions and trust, and perhaps in some cases a lack of information about existing alternatives or simple legacy/historical artefact. Furthermore, many fundraisers currently maintain separate relationships with the evaluators they use recommendations from and with the charities they end up recommending, causing extra overhead for all involved parties.

Considering this situation and from checking with a subset of fundraising organisations, it seems there is a pressing need for (1) a quality check on new and existing evaluators (“evaluating the evaluators”) and (2) an accessible overview of all recommendations made by evaluators whose methodology meets a certain quality standard. This need is becoming more pressing with the ecosystem growing both on the supply (evaluator) and demand (fundraiser) side.

The new GWWC research team is looking to start filling this gap: to help connect evaluators and donors/fundraisers in the effective giving ecosystem in a more effective (higher-quality recommendations) and efficient (lower transaction costs) way.

Starting in 2023, the GWWC research team plan to evaluate funding opportunity evaluators on their methodology, to share our findings with other effective giving organisations and projects, and to promote the recommendations of those evaluators that we find meet a certain quality standard. In all of this, we aim to take an inclusive approach in terms of worldviews and values: we are open to evaluating all evaluators that could be seen to maximise positive impact according to some reasonably common worldview or value system, even though we appreciate the challenge here and admit we can never be perfectly “neutral”.

We also appreciate this is an ambitious project for a small team (currently only 2!) to take on, and expect it to take us time to build our capacity to evaluate all suitable evaluators at the quality level at which we'd like to evaluate them. Especially in this first year, we may be limited in the number of evaluators we can evaluate and in the time we can spend on evaluating each, and we may not yet be able to provide the full "quality check" we aim to ultimately provide. We'll try to prioritise our time to address the most pressing needs first, and aim to communicate transparently about the confidence of our conclusions, the limitations of our processes, and the mistakes we are inevitably going to make.

We very much welcome any questions or feedback on our plans, and look forward to working with others on further improving the state of the effective giving ecosystem, getting more money to where it is needed mos...