Against EA-Community-Received-Wisdom on Practical Sociological Questions

In my view, there is a rot in the EA community that is so consequential that it inclines me to discourage effective altruists from putting much, if any, trust in EA community members, EA “leaders”, the EA Forum, or LessWrong. But I think that it can be fixed, and the EA movement would become very good.

In my view, this rot comes from incorrect answers to certain practical sociological questions, like:

  1. How important for success is having experience or having been apprenticed to someone experienced?

  2. Is the EA Forum a good tool for collaborative truth-seeking?

  3. How helpful is peer review for collaborative truth-seeking?

...

Meta-1. Is “Defer to a consensus among EA community members” a good strategy for answering practical sociological questions?

Meta-2. How accurate are conventional answers to practical sociological questions that many people want to get right?

I’ll spend a few sentences attempting to persuade EA readers that my position is not easily explained away by certain things they might call mistakes. Most of my recent friends are in the EA community. (I don’t think EAs are cringe). I assign >10% probability to AI killing everyone, so I’m doing technical AI Safety research as a PhD student at FHI. (I don’t think longtermism or sci-fi has corrupted the EA community). I’ve read the sequences, and I thought they were mostly good. (I’m not “inferentially distant”). I think quite highly, for the most part, of the philosophical and economic reasoning of Toby Ord, Will MacAskill, Nick Bostrom, Rob Wiblin, Holden Karnofsky, and Eliezer Yudkowsky. (I’m “value-aligned”, although I object to this term).

Let me begin with an observation about Amazon’s organizational structure. From what I’ve heard, Team A at Amazon does not have to use the tool that Team B made for them. Team A is encouraged to look for alternatives elsewhere. And Team B is encouraged to make the tool into something that they can sell to other organizations. This is apparently how Amazon Web Services became a product. The lesson I want to draw from this is that wherever possible, Amazon outsources quality control to the market (external people) rather than having internal “value-aligned” people attempt to assess quality and issue a pass/​fail verdict. This is an instance of the principle: “if there is a large group of people trying to answer a question correctly (like ‘Is Amazon’s tool X the best option available?’), and they are trying (almost) as hard as you to answer it correctly, defer to their answer.”

That is my claim; now let me defend it, not just by pointing at Amazon, and claiming that they agree with me.

High-Level Claims

Claim 1: If there is a large group of people trying to answer a question correctly, and they are trying (almost) as hard as you to answer it correctly, any consensus of theirs is more likely to be correct than you.

There is extensive evidence (Surowiecki, 2004) that aggregating the estimates of many people produces a more accurate estimate as the number of people grows. It may matter in many cases that people are actually trying rather than just professing to try. If you have extensive and unique technical expertise, you might be able to say no one is trying as hard as you, because properly trying to answer the question correctly involves seeking to understand the implications of certain technical arguments, which only you have bothered to do. There is potentially plenty of gray area here, but hopefully, all of my applications of Claim 1 steer well clear of it.

Let’s now turn to Meta-2 from above.

Claim 2: For practical sociological questions that many people want to get right, if there is a conventional answer, you should go with the conventional answer.

Claim 2 can be justified from Claim 1. Note that it is similar, but different. It is weakened by its restriction to sociological questions, and it is strengthened by moving from “people trying hard to get it right, using their expertise as much as you” to merely “wanting to get it right”.

The question “when and to what extent should I trust conventional wisdom?” is one that a large group of people are trying very hard to answer correctly. The consensus of most people is that conventional wisdom is pretty good when it comes to designing institutions; at least compared to what a first-principles-reasoner could come up with. Conventional wisdom has incorporated the intuitions, trials, and errors of millions of people. Practical sociological questions are empirical questions, where logical aptitude can only take you so far, and they are one of the domains within which our intuitions are most thoroughly honed, so the ability to productively correct a consensus is likely to be fairly evenly distributed over a big population.

EA Trust-Networks

Claim 3: “Defer to a consensus among EA community members/​EA leadership” is a bad strategy for answering practical sociological questions.

This follows from Claim 2. This is a practical sociological question where there exists a conventional answer, so Claim 2 applies. The conventional thing to do when a small group of people proposes new norms for relating to others, new norms for structuring organizations, and new norms for aggregating individuals’ beliefs, is to reject those proposals without even bothering to refute them. In general, logic will not get you far when trying to address practical sociological questions, so knockdown refutations are as rare as knockdown defenses.

I really do not want to rest my case on an object-level argument about the virtues of various EA norms. I can make my case, and others can respond with why they don’t see it like that, and the status quo will prevail. My contention is that in the likely event that no one can offer an airtight first-principles argument about the utility of certain norms, we should defer to the conventional wisdom of our society as a whole.

Peer Review

Claim 4: High-prestige peer review is the best institution ever invented for collaborative truth-seeking.

This is the conventional answer to a practical sociological question that many people want to get right.

But Einstein didn’t get his papers peer-reviewed! I mean “peer review” to include ordinary peer review (at high-prestige venues) plus an “alternate track” in which the entire community of people with the technical expertise to understand the sentences you are writing agree it is correct. This is not me awkwardly redrawing of the borders of concepts to make my claim go through; the physics community’s acceptance of general relativity obviously qualifies as high-prestige peer review.

But the pervasiveness of peer review isn’t even a century old! Yes, before that, scientists were not subjected to peer review; rather they presented their work to their … peers … in Royal Societies who … reviewed … their work.

There is much ink spilled on good science that was never peer reviewed, and bad science that passed peer review, and indeed peer review is not a perfect institution, but we also need to consider how much alchemy (and other flawed attempts at truth-seeking) it has spared us from having to attend to. A key challenge of collaborative truth-seeking is that decision-makers do not have time or often the ability to gain expertise in every technical area and correctly evaluate all technical work themselves.

But my main argument for Claim 4 is that many communities have wanted to design the best institution they could for collaborative truth-seeking. And peer review is the institution that has spread over (I think) every academic discipline. It is the institution that governments have chosen as the basis of their decisions, whenever possible.

I’m not saying that everyone involved is perfect at selecting institutions that promote healthy collaborative truth-seeking; far from it. I’m saying they’re about as good as you are, and they’ve been at it for longer.

Claim 5: The comment sections and upvote tallies on LessWrong and the Alignment Forum and the EA Forum do not qualify as high-prestige peer review.

The conventional answer is loud and clear on this topic. I can offer some words of explanation on how this could be, but even if I couldn’t, you should believe it anyway, because we can be absolutely sure what the consensus would be: the consensus among laypeople would be to defer to the consensus of academics, and the consensus among academics would be, “What? No, of course not.” Some words of explanation as to how this could be: the commenters are not experts; the commenters don’t have the ability to prevent publication, so authors don’t take their concerns as seriously; if the author gets into an unresolved debate with a commenter 6 levels deep, the default outcome is for the post to stand and nobody to read the debate, and everyone knows this; the whole community is an echo chamber that takes for granted various propositions which have never been rigorously defended to the satisfaction of outsiders. I don’t know which of these differences between these online forums and standard academic peer review are important, and neither do you.

At least in AI existential safety, the classic reasons not to submit work to high-status peer reviewed venues is that it’s too much work, it’s too slow, the normie reviewers won’t understand because there’s too much “inferential difference”, and peer review is just a status-chasing game anyway. If you are a funder, or a policymaker, or someone reporting on the consensus of a research community, conventional wisdom would suggest that you ignore these excuses, and focus on peer-reviewed work. I do apologize for any offense there, but I think it needs to be said; I do not think people who claim this are necessarily arriving at false conclusions in their research, but conventional wisdom would say that demanding better receipts in the form of peer-reviewed publications is a better strategy than giving them the benefit of the doubt because they belong to your philosophical movement, and because you can’t find any knockdown counterarguments. If you are an expert researcher, mine the work of people who consider peer-review to be not worth their time, but who seem to be making good points, and see what you can turn into published, peer-reviewed work.

One quick comment on peer review being just about status-chasing, because I have too nice a comparison. There was some Yudkowsky tweet I can’t find that refutes the idea that arguments are just for convincing people into doing what you want, not for pursuing the truth. He noted that not only did we evolve to produce convincing arguments; we evolved to believe them. The latter only makes sense if they have some value for pursuing truth. Similarly, human collaborative-truth-seeking norms have evolved not only to produce people who seek status through peer review, but also people who assign status to peer-reviewed work. This only makes sense if peer-reviewers are succeeding at (partially) separating high-quality work from low-quality work.

It is hard for a group to form a consensus around correct claims over incorrect ones; the costs involved are real and cannot be wished away. If you acutely feel the inconvenience of getting your work peer reviewed, and feel that this is evidence of an inefficient system, why are you expecting the existence of a better system? 1) You publish your work, 2) …., 3) its correctness becomes common knowledge. I’m sure anti-peer-review people can come up with creative proposals for 2. Maybe they will distill the correct claims from the incorrect ones more cost-effectively than peer review, and maybe they won’t. My point is that no one should ever have expected the “consensus building” part of collaborative truth-seeking to be costless, so your experience of bearing that cost should not be considered much evidence of inefficiency.

Complaints about the slowness of peer review and the lack of uniformly high-quality reviewers, which are often supposed to strike the reader as a clear indictment of the system, remind me of complaints about high drug prices as an indictment of capitalism. The costs of scrutiny and aggregation for consensus building are real and must be borne somewhere, much like the capital costs of developing drugs. Coming face to face with those costs is just unsurprising. The other social phenomenon I am reminded of is the tendency of revolutionaries to inveigh against the status quo without feeling the need to compare it to an alternative proposal.

I have one more comment on the complaint about supposed “inferential difference” and dumb “normie” reviewers. My entire threat model of AI existential risk, including the contention that our extinction is likely under certain circumstances, has passed peer review at a venue whose brand is backed by AAAI. I suspect that someone worried about AI existential risk whose inclinations are anti-peer review would have been surprised or astonished by that.

How does Rationalist Community Attention/​Consensus compare? I’d like to mention a paper of mine published at the top AI theory conference which proves that when a certain parameter of a certain agent is set sufficiently high, the agent will not aim to kill everyone, while still achieving at least human-level intelligence. This follows from Corollary 14 and Corollary 6. I am quite sure most AI safety researchers would have confidently predicted no such theorems ever appearing in the academic literature. And yet there are no traces of any minds being blown. The associated Alignment Forum post only has 22 upvotes and one comment, and I bet you’ve never heard any of your EA friends discuss it. It hasn’t appeared, to my knowledge, in any AI safety syllabuses. People don’t seem to bother investigating or discussing whether their concerns with the proposal are surmountable. I’m reluctant to bring up this example since it has the air of a personal grievance, but I think the disinterest from the Rationality Community is erroneous enough that it calls for an autopsy. (To be clear, I’m not saying everyone should be hailing this as an answer to AI existential risk, only that it should definitely be of significant interest.) [EDIT: the most upvoted comment disputes some of what is in this paragraph. Please read my response to it, and my response to his response to that. The errors in the comment are so verifiable that I’m mystified by the upvoting behavior. It’s like the upvoters are looking for reasons to write this off, which is kind of exactly my point here.]

But again, if my object-level defenses of peer-review do not satisfy you (which I only add because I expect some people would be interested in an attempted explanation), defer to the conventional wisdom anyway! For potential venues for AI existential safety research, I would say that AAAI, IJCAI, Synthese, and AI Magazine are especially likely to source competent and receptive reviewers, but see the publication records of AI safety researchers for more ideas. (Personal websites often have better publication info than Google scholar).

Conventional Experience

Claim 6: Having experience or having been apprenticed to someone experienced is critically important for success, with rare exceptions.

Conventional wisdom is very unified here. It is a question that lots of people and companies care about getting right. Companies with bad hiring criteria in competitive markets would be outcompeted by companies with better ones. So look around at companies in competitive markets and note how much the companies that have survived until today value experience when they do their hiring, especially for important roles. If you want to be sure that a job done is well, don’t hire an EA fresh out of college. Hire someone with a strong track record that a conventional HR department would judge as demonstrably competent. Companies hire people all the time who are not “aligned” with them, i.e. not philosophically motivated to maximize the company’s profit, and it works out fine.

An EA friend of mine was looking to hire a personal assistant for his boss, and it didn’t even occur to him to look for someone with twenty years of experience and a track record of success! I think it shouldn’t have occurred to him to not do that. He had been more focused on finding someone who shared his boss’s philosophical commitments.

If “grown-ups” had been involved at the officer level at FTX, I claim fraud probably would not have occurred. I can’t say I predicted FTX’s collapse, but I didn’t know it was being run by people with no experience. Once I learned the collective experience-level of the management of the Carrick Flynn campaign, I did predict that he would get very few votes. Maybe that was overdetermined, but I was much more pessimistic than my EA community friends.

Many small organizations are founded by EAs who would not succeed at being hired to run an analogous non-EA organization of similar scope and size. I (tentatively) think that these organizations, which are sometimes given the outrageously unsubstantiated denomination “effective organization”, are mostly ineffective. Regardless of the frequency with which they end up ineffective, I am more confident that the process by which such organizations are deemed “effective” by the EA community is fairly disconnected from reality. It seems to me to be basically predictable from whether the organization in question is run by friends of friends of the speaker. (It is possible that certain individuals use the term more sparingly and more reliably, but the EA community as a whole seems to use it quite liberally). But I do not only object to the term, of course. Conventional wisdom does not think especially highly of the likelihood that these organizations will be of much importance to society.

Objections

Objection: conventional wisdom is that effective altruism and longtermism are problematic, and AI X-risk is non-existent, but the conventional wisdom is wrong. Conventional wisdom was even more dismissive of them several years ago. But all three have been EA-community-conventional-wisdom for years, and to the extent they have become more mainstream, it is because of people disregarding conventional wisdom while taking EA-community-conventional wisdom seriously.

Yes, these are philosophical positions, not sociological ones, so it is not so outrageous to have a group of philosophers and philosophically-minded college students outperform conventional wisdom by doing first-principles reasoning. This does not contradict my position, which regards sociological questions.

Objection: but if everyone only ever accepted peer-reviewed perspectives on AI X-risk, the field never would have gotten off the ground; initially, everyone would have dismissed it as unsubstantiated. So a conventional answer to the sociological question “Should I trust this non-peer-reviewed first-principles reasoning?” would have interrupted open-minded engagement with AI X-risk.

There is a difference between finding it worthwhile to investigate a claim (“big if true”) and accepting a claim. If the president had come across LessWrong before Superintelligence was published, and he decided he shouldn’t base policy off of such immature research, he would have been making an unfortunate mistake. But he wouldn’t be consigning the field of AI Existential Safety to abandonment, and he would have been following a principle that would usually serve him well. So I claim it is totally acceptable and good for actual decision-makers to only consider peer-reviewed work (or the consensus of experienced academics who take peer review seriously). Meanwhile, researchers should absolutely attend to non-peer-reviewed claims that are “big if true”, and aspiring researchers should seek the technical expertise they need to join their ranks. Researchers can turn these “big if true” claims into papers that can be evaluated by ordinary, high-quality peer reviewers. Research funders can aim to fund such efforts. And if you find you can’t get a certain claim or idea past peer review after having tried (either as a researcher or funder), then be very reluctant to build a world-view around it.

See a recent tweet of mine on this topic within a short conversation https://​​twitter.com/​​Michael05156007/​​status/​​1595755696854810624. “Claim: if you take the claims first articulated on LessWrong that have since been defended in peer-reviewed articles and academic books, you get a set of important claims worth taking seriously. Not as true if you skip the second step.”

Objection: but there are lots of true claims that appear only in non-peer-reviewed, non-academic-book AI safety research. And the AI safety community agrees.

And how many false claims, that are equally well-respected? They may be some claims that are on their way to being defended in a peer-reviewed paper, or could be if someone put in the effort. I would be interested to see prediction markets about whether some claims will go on to be defended in a peer reviewed article at a high-quality venue.

Currently, it seems to me that many AI safety researchers read “intuition-building” blog posts that support a claim, and consider it to be “worth looking into”. And then it seems to everyone that there is lots of talk throughout the AI safety community that this claim is “worth looking into”, so people think it’s probably pretty plausible. And then people start doing work predicated on the claim being true, and now it’s AI-Safety-community-mainstream. This story is not an argument that all intuition-building blog posts introducing a claim are false! I only aim to illustrate how easily the same process can be seeded on a false claim, since you can always discover an intuition that supports a claim.

So I claim the AI safety community does not demand sufficient rigor before internally-generated claims gain the status of commonly-accepted-by-the-community, and I claim it has let false ones through. Much of the AI Safety community has largely abandoned the institution of peer review, where the reviewers are technically experienced and live outside of their echo chamber, especially when it comes to defending their core beliefs. It is just not so surprising to have many matching incorrect views among a community that engages in such insular epistemic practices in violation of conventional approaches to collaborative truth-seeking. I have not argued here that lots of un-peer-reviewed work in AI safety is in fact incorrect, but the reader should not be so convinced that the field is thriving for it to be much evidence when peer review is on trial.

I’ll give an example of a specific claim whose acceptance among AI safety researchers has, in my opinion, significantly outpaced the rigor of any defense. There is peer-reviewed work arguing that sufficiently advanced reinforcement learners would under certain conditions very likely kill everyone (Cohen, 2022). Academic books (Superintellgence and Human-Compatible) have made more modest but similar claims. Where is the peer-reviewed work (or academic book) arguing that sufficiently advanced supervised learners trained to imitate human behavior would probably kill everyone? I have an explanation for why it doesn’t exist: imitation learners trained to imitate humans probably won’t kill everyone, and the AI safety community has never demanded sufficient rigor on this point.

Objection: But what if conventional wisdom on a practical sociological question is wrong? Under this view, we (a small community of plucky and smart individuals) could never do better.

Essentially, yes. On what grounds could we have expected to? That said, I do not think conventional wisdom is fixed. I imagine that on the question of peer review, for example, economics journals would be interested in an empirical evaluation of a proposed alternative to peer review, and depending on the results, this could lead to a change in conventional wisdom in academic institutions. So if you intuit that we can do better than peer review, I would recommend getting a PhD in economics under a highly-respected supervisor, and learn how to investigate institutions like peer review (against proposed alternatives!) with a level of rigor that satisfies high-prestige economists. And I would recommend not putting much credence on that intuition in the meantime. Other pieces of conventional wisdom could evolve in other ways.

Objection: You say [X] about [peer review/​importance of conventional experience], but actually [not X] because of [Y].

Recall that I am only offering potential explanations of certain pieces of conventional wisdom on these topics, since I think some readers would be interested in them. But I can’t be sure that I am correctly diagnosing why pieces of conventional wisdom are true. I may be totally missing the point. I am engaging in rationalization of conventional wisdom, not open-minded, rational inquiry. My key claim is that we should trust conventional wisdom on practical sociological questions, regardless of the result of an open-minded, rational, gears-level inquiry, so attempting to work it out for ourselves is inappropriate here. (I understand rationalization is often bad, but not here).

Consequences

I mentioned that I believe the consequences of these problems are so bad that I would discourage effective altruists from taking advice from EA community members or taking seriously the contents of the EA Forum or the Alignment Forum. It looks to me like lots of effective altruists (not all of them) are being encouraged to work for so-called effective organizations, many of which just aim to grow the EA community, or they have inexperienced management and so the blind are leading the blind, or they aim to contribute to the development EA-community-consensus through the publication of unrigorous blog posts to the EA Forum or Alignment Forum, most of which would probably be judged as low-quality by external reviewers. It is not uncommon for poor execution to destroy approximately all the value of a well-considered goal.

It looks to me like the AI safety researchers in the EA community are mostly not producing useful research (in the sense of being helpful for avoiding extinction), which is a topic I’ll have to spend more time on elsewhere. This depends on certain object-level claims that might be controversial in the AI safety community, but I think someone who formed their object-level beliefs only on the basis of peer-reviewed AI existential safety literature would come to the same conclusion as me.

Gratitude to EA

I think the body of written work that constitutes the philosophy of and application of Effective Altruism is excellent. Likewise for longtermism. I think body of written work on AI existential risk and safety has some excellent and life-or-death-critical chapters. And I am extremely grateful to everyone who made this possible. But my gratitude to these authors and to their colleagues who helped them hone their ideas does not compel me to trust their intuitions and deliberations on practical sociological questions. And it certainly doesn’t compel me to trust the intuitions of other readers who find these authors compelling (i.e., the communities that have formed around these ideas).

Alternatives

Here is my alternative advice for people who find the philosophical ideas of effective altruism compelling.

If there are are topics where you have conventionally well-respected experience picking apart validity from nonsense, try your best. Otherwise do what ordinary people would do: on most topics, defer to conventional wisdom. On technical topics, defer to the people conventionally deferred to—those with conventional markers of expertise—naturally, excluding people who produce claims/​arguments you can detect as nonsense.

Then, using that worldview, try to figure out how to do the most good as a professor, politician, civil servant/​bureaucrat/​diplomat, lobbyist for a think tank, or journalist—the conventional professions for changing the world with your ideas. (If there is a pressing engineering problem, maybe you could be an entrepreneur). Of course there will be exceptions here, especially if you try and fail to get one of those jobs. 80,000 Hours will have some good ideas, even if there are some suggestions mixed in that I would disagree with.

Close contact with some sort of “EA community” probably won’t be very helpful in any of these pursuits, although of course it’s nice to make friends. Communities are good for networking, so if you’re looking for a job on Capitol Hill, by all means ask around to see if anyone has any leads. But this is a pretty low-contact use case for a community.

This is an awkward position for most EA community members to espouse. They might worry about offending a friend or a romantic partner. For most effective altruists, inasmuch as they agree with me, they will be less likely to spend as much time conversing with EA community members (in person or online). So the question of whether this position is correct is absolutely one for which the strategy “Defer to a consensus among active EA community members” is inadequate, since there is selection pressure favoring people who find my position unintuitive.