Hide table of contents

These are two things I posted on the EA Facebook group that I thought should be saved somewhere for future reference and discussion. I link to the Facebook posts so you can see the responses people have made.

--

Post I

Alyssa Vance​ thought this was one of the biggest problems with effective altruism as it exists in practice today (my comments below):

"Most critically, a lot of EA (though notably not OPP) discourages donors from taking risks, by advocating interventions whose effects are short-term, concrete, certain, and easy to evaluate. You need lots of risk to get very high payoffs, but since donors won't directly benefit from those payoffs, they tend to play it very safe, the same way that bureaucrats do. The problem is, playing it safe lops off the long tail of very good outcomes. Paul Graham explains the general principle (http://paulgraham.com/hiring.html):

"Risk and reward [in wealth investing] are always proportionate. For example, stocks are riskier than bonds, and over time always have greater returns. So why does anyone invest in bonds? The catch is that phrase "over time." Stocks will generate greater returns over thirty years, but they might lose value from year to year. So what you should invest in depends on how soon you need the money. If you're young, you should take the riskiest investments you can find.

All this talk about investing may seem very theoretical. Most undergrads probably have more debts than assets. They may feel they have nothing to invest. But that's not true: they have their time to invest, and the same rule about risk applies there. Your early twenties are exactly the time to take insane career risks.

The reason risk is always proportionate to reward is that market forces make it so. People will pay extra for stability. So if you choose stability-- by buying bonds, or by going to work for a big company-- it's going to cost you.""

--

Clearly as a society we should do a combination of:

i) scaling things with well-understood payoffs;

ii) trying things with uncertain returns, but where at least either the benefit or likelihood of success can be roughly estimated;

iii) trying new things with uncertain returns where the benefit and likelihood of success are very hard to estimate.

There will be some outstanding opportunities in all of these categories - which approach this community should be disproportionately focussed on depends on which one is relatively neglected by the rest of society.

So, is the rest of the world overall risk-loving, risk-neutral or risk-averse when it comes to their (intentional and accidental) social impact?

Or to put it another way, are we being left mostly neglected safe opportunities, or mostly neglected high-risk/high-leverage opportunities?

---

Post 2

Is 'highly skeptical EA' a self-undermining position? Here's my line of thought:

* 'Highly skeptical EAs' think you should demand strong empirical evidence before believing that something works / is true / will happen.

* As a result they typically think we should work on scaling up 'proven' interventions, as opposed to doing 'speculative/unproven' things which can't be demonstrated to a high standard of evidence to be better.

* But the claim that it's higher expected value to do the best 'proven' thing rather than a speculative/unproven thing that on its face looks important, neglected and tractable, is itself unproven to a high standard of evidence. Indeed, it's a very hard claim to substantiate and would require a very large project involving lots of people over a long period of time investigating the average long-term return on, e.g. basic science research. As a result I think we don't really know at the moment and should be pretty agnostic on this question.

* So if we need strong evidence to accept positions, should we in fact believe with confidence that we really need strong evidence to think something has a high expected social impact?

Philosophers may note the similarity to logical positivism undermining its own core claim, though in this case it's more probabilistic than a matter of contradictory formal logic.

The fact that an idea partially undermines itself isn't a decisive argument against it, but it does suggest we should tread with care.

6

0
0

Reactions

0
0

More posts like this

Comments11
Sorted by Click to highlight new comments since: Today at 8:17 PM

I think you attribute an overly strong position to the highly skeptical EAs in post 2, as discussed in my longer comment on this. To illustrate this briefly, I'd distinguish between two positions:

(1) We should never fund activities which are [much?] more speculative than GiveWell recommendations.

(2) We should sometimes fund activities which are [much?] more speculative than GiveWell recommendations, but less often than the average EA would accept.

I'd only defend (2), not (1).

Agreed, I would also defend 2, not 1 - good point!

[My mini defense of "Charity Science"-style empirical EA]

Earlier I said I was skeptical of uncertain causes. I still think this approach is largely correct (though with some revisions) and I have hoped for quite some time to elaborate again in length.

Even if it is self-defeating to request evidence in some respect (even if you have evidence for evidence, do you have evidence for evidence for evidence?), the opposite position "we need absolutely no evidence" is also flatly ridiculous (though, admittedly, not self-defeating) and we end up chasing after Pascal's Mugging.

So we need some evidence, clearly -- the question is, how much?

This is a matter of epistemology and priors, but I think the "evidence crowd" has done quite well. Remember that EA was founded on the "evidence crowd". The crowd that didn't request evidence gave us Play Pumps, whereas the crowd that did request evidence has saved thousands of lives through AMF.

But what about moonshots? How much evidence went into eradicating smallpox or producing vaccines? Is GiveWell underinvesting in these sorts of things? I don't really know personally and haven't thought about it much, but it's possible.

I still think, as I've argued before, that it really comes down to what the optimal strategy is under uncertainty and I believe that game theory has established the optimal strategy is to do an explore-exploit pattern, where you start out by exploring the space widely and then slowly transition into exploiting the best things found so far.

I think this explore-exploit is done very well by Good Ventures and GiveWell with OpenPhil doing a lot of the exploring but still exploiting the best so far by investing in AMF.

However, it's pointless to explore unless you actually learn -- otherwise you're not really exploring, you're just exploiting randomly, which is even worse. And the way you learn is by collecting evidence -- evidence that is high-quality enough that you can use it to improve your future actions.

Does this need to be 22 RCTs, as which backs the idea of distributing malaria nets? No -- notably, GiveWell still gives millions to orgs that have much less RCTs and there still are several outstanding questions about AMF's impact.

But on the other hand I see EAs acting without thinking statistically about their activities, running surveys, and showing any signs of skepticism in their work.

I think the "evidence crowd" has done quite well. Remember that EA was founded on the "evidence crowd". The crowd that didn't request evidence gave us Play Pumps, whereas the crowd that did request evidence has saved thousands of lives through AMF.

This seems really hard to settle.

Arguably, all of medical research is in the non evidence crowd in the sense that they can't have empirical evidence that a research program is going to work ahead of time. You can work out whether a medical intervention works later, but you have to invest hugely up front. Medical research has done an absolutely huge amount of good, far more than people focused on scaling up evidence-backed charity and government programs (so far).

Who's done the most good doesn't settle it. We want to know something more like who's done the most good per unit of input. But even then the average biomedical research does pretty well.

https://80000hours.org/career-guide/top-careers/profiles/biomedical-research/

You could also put all social movements in the non evidence crowd, including the end of slavery, expansion of the vote, civil rights etc. Likewise, all technological and political innovation.

Perhaps the odd playpump is worth it? And playpump isn't a good example because sensible non-evidence people would rule it out as well. Looking at facts like how they cost 4x more than regular water pumps and the recipients didn't want them should make anyone cautious about scaling them up, whether or not you've got an RCT.

In the meantime, I agree explore-exploit is a good approach. I'd also say being modest about which causes and methods are best. Expert common sense as a prior seems to be significant weight on things like research, politics, innovation, and social advocacy being major ways to make the world a better place. Finally, there's looking for other arguments, such as considerations around neglectedness.

I figure I fall into the "skeptical EA" camp, so let me try defending it. :) It'd be good to make progress on this issue so I appreciate your engaging with it! Here I'll consider your two key steps in post 2.


To take your first step first:

"Highly skeptical EAs think you need strong evidence to prove that something works / is true / will happen."

This would be self-refuting, as you say. But I don't think most have quite as strong a position as that. After all, 'provenness' is a matter of degree. It's more that we're relatively negative about speculative activities.

It's also worth distinguishing between:

(1) charities which definitely have a 1% chance of doing an enormous of good. For example, a charity which'd definitely do 101 times as much good as AMF if a 100-sided die were rolled and came up 100.

(2) charities which may have a 1% chance of doing an enormous amount of good, but lack robust evidence for this. E.g. they have no track record, no precedent, no broad expert/academic endorsement. But a back of the envelope calculation including some very rough guesses suggests that they have a 1% chance of doing 101 times as much good as AMF.

I'd give to (1) over AMF, but not (2).


To consider your third step:

"But the claim that it's higher expected value to do the best 'proven' thing rather than a speculative/unproven thing (that on its face looks important, neglected and tractable) is itself unproven to a high standard of evidence."

True. This would be a reductio ad absurdum of the claim that we should only ever believe 'proven' propositions (which we could perhaps define as those which an ideally rational agent in our position would have >90% credence in). But 'skeptical EAs' rarely claim anything so implausible.

The best epistemic approach on these issues is clearly far from proven, so we have no choice but to pick our best bet (adjusting for our uncertainty). It could still be the case, without inconsistency, that the best epistemic approach is to rate relatively speculative activities lower than the average EA does.

"Indeed, it's a very hard claim to ever prove and would require a very large project involving lots of people over a long period of time looking at the average return on e.g. basic science research."

This sort of look at the historical track record of different epistemic approaches does indeed seem the best approach. You're right that the correct answer is far from 90% proven.

"(In my view, we don't really know at the moment and should be agnostic on this issue.)"

If by 'agnostic' you mean 'completely neutral' (which you may very well not?) then I disagree. Some approaches seem better than others, and we should take our best bet.

they have no track record, no precedent, no broad expert/academic endorsement.

One problem is that interventions which do have these things are usually not neglected. i.e. you get higher tractability but lower neglectedness.

Since both matter, it becomes unclear where on the spectrum to focus.

It seems unlikely you either want 100% neglectedness or 100% tractability (because then you'll be ignoring the opposite factor), so my guess is somewhere in the middle.

I think this speaks in favor of looking at areas on the edge of consensus, where there's emerging but not fully decisive evidence.

As you note, the next step is "what's the robust evidence that AMF beats 2)"?

I think your response though is the right one overall.

It's a difficult judgement call made with little evidence and we have to make our best bet.

I think people should be modest about their confidence in focusing on projects that look like 1), AMF, or 2). It wouldn't take a lot of evidence to convince me one way or another, and I would advocate a mixed strategy between them among the community today.

By agnostic, I just mean thinking there's a decent chance (10%+) any of these could be the best approach for someone, and so not using this difference as the key issue on which you judge projects.

By agnostic, I just mean thinking there's a decent chance (10%+) any of these could be the best approach for someone, and so not using this difference as the key issue on which you judge projects.

That's much more plausible than total neutrality! I agree that there's no theoretical argument (that I know of) for thinking that (2) is very likely to be worse than AMF. So it all depends on what the best available candidates for (2) are. Perhaps people could make progress by discussing a real world example of (2). (Ideally this wouldn't be an org anyone has ties to, to allow for especially neutral discussion of it.)

I'm worried there's a motte and bailey argument where people start out by saying "we need to be speculative because impact will happen through taking risks and focusing on growth!" (motte) and then go to "therefore we don't need to bother considering things like counterfactuals or spending that much time on impact reports" (bailey).

I agree this is a danger, but I'm not sure many people are currently making this mistake. When do you think it might have happened?

On post 1, this is a good observation, and it certainly speaks in favour of riskier donations. But it's not clear that it does so very strongly, and I'd expect it to be a relatively weak consideration once you start directly considering candidate charities (e.g. AMF vs. an outfit that lobbies for a US carbon tax).