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: Against LLM Reductionism, published by Erich Grunewald on March 8, 2023 on LessWrong.

Summary

Large language models (henceforth, LLMs) are sometimes said to be "just" shallow pattern matchers, "just" massive look-up tables or "just" autocomplete engines. These comparisons amount to a form of (methodological) reductionism. While there's some truth to them, I think they smuggle in corollaries that are either false or at least not obviously true.

For example, they seem to imply that what LLMs are doing amounts merely to rote memorisation and/or clever parlour tricks, and that they cannot generalise to out-of-distribution data. In fact, there's empirical evidence that suggests that LLMs can learn general algorithms and can contain and use representations of the world similar to those we use.

They also seem to suggest that LLMs merely optimise for success on next-token prediction. It's true that LLMs are (mostly) trained on next-token prediction, and it's true that this profoundly shapes their output, but we don't know whether this is how they actually function. We also don't know what sorts of advanced capabilities can or cannot arise when you train on next-token prediction.

So there's reason to be cautious when thinking about LLMs. In particular, I think, caution should be exercised (1) when making predictions about what LLMs will or will not in future be capable of and (2) when assuming that such-and-such a thing must or cannot possibly happen inside an LLM.

Pattern Matchers, Look-up Tables, Stochastic Parrots

My understanding of what goes on inside machine learning (henceforth, ML) models, and LLMs in particular, is still in many ways rudimentary, but it seems clear enough that, however tempting that is to imagine, it's little like what goes on in the minds of humans; it's weirder than that, more alien, more eldritch. As LLMs have been scaled up, and more compute and data have been poured into models with more parameters, they have undergone qualitative shifts, and are now capable of a range of tasks their predecessors couldn't even grasp, let alone fail at, even as they have retained essentially the same architecture and training process.[1] How do you square their awesome, if erratic, brilliance with the awareness that their inner workings are so ordinary?

One route would be to directly deny the brilliance. Gary Marcus does this, pointing out, and relishing in, the myriad ways that LLMs misfire. Their main limits are, he says, that they are unreliable and untruthful. (See the footnote for my thoughts on that.[2])

That's one route, but it's not the one I want to discuss here. The route I want to discuss here is to dispel the magic, so to speak: to argue that what goes on inside LLMs is "shallow", and that LLMs lack "understanding". This often takes the form of asserting that LLMs are just doing pattern matching[3], or just rephrasing material from the web[4], amounting to mere stochastic parrots[5], or just retrieving things from a massive look-up table. Gary Marcus describes the underlying problem as one of "a lack of cognitive models of the world":

The improvements, such as they are, come primarily because the newer models have larger and larger sets of data about how human beings use word sequences, and bigger word sequences are certainly helpful for pattern matching machines. But they still don't convey genuine comprehension, and so they are still very easy [...] to break.

Well -- in a certain light and for the sake of fairness -- this view is not entirely wrong:

LLMs are, in a sense, pattern matching. They likely have a great deal of attention heads and neurons and whatever that detect certain patterns in the input, which then help determine the model's output.

LLMs are, in a sense, merely rephrasing material from the web. All, or nearly all, of the data that 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: 80,000 Hours two-year review: 2021–2022, published by 80000 Hours on March 8, 2023 on The Effective Altruism Forum.

80,000 Hours has released a review of our programmes for the years 2021 and 2022. The full document is available for the public, and we’re sharing the summary below.

You can find our previous evaluations here. We have also updated our mistakes page.

80,000 Hours delivers four programmes: website, job board, podcast, and one-on-one. We also have a marketing team that attracts users to these programmes, primarily by getting them to visit the website.

Over the past two years, three of four programmes grew their engagement 2-3x:

Podcast listening time in 2022 was 2x higher than in 2020

Job board vacancy clicks in 2022 were 3x higher than in 2020

The number of one-on-one team calls in 2022 was 3x higher than in 2020

Web engagement hours fell by 20% in 2021, then grew by 38% in 2022 after we increased investment in our marketing.

From December 2020 to December 2022, the core team grew by 78% from 14 FTEs to 25 FTEs.

Ben Todd stepped down as CEO in May 2022 and was replaced by Howie Lempel.

The collapse of FTX in November 2022 caused significant disruption. As a result, Howie went on leave from 80,000 Hours to be Interim CEO of Effective Ventures Foundation (UK). Brenton Mayer took over as Interim CEO of 80,000 Hours. We are also spending substantially more time liaising with management across the Effective Ventures group, as we are a project of the group.

We had previously held up Sam Bankman-Fried as a positive example of one of our highly rated career paths, a decision we now regret and feel humbled by. We are updating some aspects of our advice in light of our reflections on the FTX collapse and the lessons the wider community is learning from these events.

In 2023, we will make improving our advice a key focus of our work. As part of this, we’re aiming to hire for a senior research role.

We plan to continue growing our main four programmes and will experiment with additional projects, such as relaunching our headhunting service and creating a new, scripted podcast with a different host. We plan to grow the team by ~45% in 2023, adding an additional 11 people.

Our provisional expansion budgets for 2023 and 2024 (excluding marketing) are $12m and $17m. We’re keen to fundraise for both years and are also interested in extending our runway — though we expect that the amount we raise in practice will be heavily affected by the funding landscape.

The Effective Ventures group is an umbrella term for Effective Ventures Foundation (England and Wales registered charity number 1149828 and registered company number 07962181) and Effective Ventures Foundation USA, Inc. (a section 501(c)(3) tax-exempt organisation in the USA, EIN 47-1988398), two separate legal entities which work together.

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: EA needs Life-Veterans and "Less Smart" people, published by Jeffrey Kursonis on March 8, 2023 on The Effective Altruism Forum.

Healthy communities have all kinds. There is a magic in the plant world when a diversity of plants co-exist. Permaculture has been innovating through realizing how the community of plants help each other by each contributing a different gift to the benefit of the whole. Plants communicate with each other via mushroom like strands underground and work together. Interestingly, they speak French. Just kidding.

I’ve been in a movement that changed the world in a positive way and eventually fell apart, it was very very similar to EA in many ways — a bunch of talented young people trying to do good in the world. We had all the same criticisms people throw at EA and we did listen and learn as much as we had the capacity to. I won’t tell the whole story here, but we didn’t fall apart because of bad things, it was a necessary evolution, but one of the key problems that kept us from surviving was a lack of diversity.

For some reason when I was young, I think it’s because I was smart, I figured out that if older people had already faced all the challenges I face maybe they would be a good source of data and the gritty life wisdom of how to apply the data. So I would go out of my way to befriend them and listen to them. It was mixed results.lots of older people are just bitter, but there were enough that had made it through a life full of thriving and were happy to share it. You just have to find the right one’s.

Most of the world is made up of average people, smart people call them dumb, but they’re really just average. The thing is, if everybody in the room is smart, who is going to see the world as most of the world sees the world? That’s a data poor room.

If we are really smart we’ll make sure to surround ourselves not just with other smart people but with a variety of young and old, different cultures, different life experience levels and some average people. That’s a room rich in data.

Never underestimate the simple wisdom of simple people.and because most of the people in the world are religious, we should have them around too. You just need to find the right one’s.generous and kind and who want everyone to thrive.

Wisdom is learning how to live in reality.when we’re young we are really far from reality.you have a bedroom and a phone and an iPad in a lovely house, all provided for you magically. You have no clue how that all accrued to you. You’re not yet in touch with reality. But as you attend the school of hard knocks year after year, slowly but surely reality drifts in.essentially what happens is that as you are slowly disconnected from your parents and the “magical accrual” fades away, you learn how real life works.

Wise people have had the time it takes to boil it all down to pure essence, filter out the dross and see the pure reality.when you can see it, you can figure out how to negotiate it. It simply takes some years and a person oriented toward thriving rather than increasing bitterness.

If you have a lot of data, what you need more than anything is wisdom to interpret the data and wisdom to creatively imagine real world applications from the data.

You simply cannot be a movement committed to getting more effective at doing good in the world if you do not have some elder wisdom in the room. It’s a glaring deficit. Thank God for Singer, but he’s not around enough.

Especially in this time of post-FTX self examination and reforming, in this time of making efforts concerning the mental health of young people under pressure to save the world—this is the time to round out the community with some village like balance; The young, the old, the strong, the average all making life thrive like plants all mixed up in the jungle.

And artists! My God...

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: The Kids are Not Okay, published by Zvi on March 8, 2023 on LessWrong.

It has been a subject of much recent internet discourse that the kids are not okay. By all reports, the kids very much seem to be not all right.

Suicide attempts are up. Depressive episodes are way up. The general vibes and zeitgeist one gets (or at least that I get) from young people are super negative. From what I can tell, they see a world continuously getting worse along numerous fronts, without an ability to imagine a positive future for the world, and without much hope for a positive future for themselves.

Should we blame the climate? Should we blame the phones? Or a mind virus turning them to drones? Heck, no! Or at least, not so fast.

Let’s first lay out the evidence and the suspects.1

Then, actually, yes. Spoiler alert, I’m going to blame the phones and social media.

After that, I’ll briefly discuss what might be done about it.

Suicide Rates

The suicide numbers alone would seem at first to make it very very clear how not all right the kids are.

Washington Post reports, in an exercise in bounded distrust:

Nearly 1 in 3 high school girls reported in 2021 that they seriously considered suicide — up nearly 60 percent from a decade ago — according to new findings from the Centers for Disease Control and Prevention. Almost 15 percent of teen girls said they were forced to have sex, an increase of 27 percent over two years and the first increase since the CDC began tracking it.

Thirteen percent [of girls] had attempted suicide during the past year, compared to 7 percent of boys.

One child in ten attempted suicide this past year, and it is steadily increasing? Yikes.

There is a big gender gap here, but as many of you already suspect because the pattern is not new, it is not what you would think from the above.

In the U.S, male adolescents die by suicide at a rate five times greater than that of female adolescents, although suicide attempts by females are three times as frequent as those by males. A possible reason for this is the method of attempted suicide for males is typically that of firearm use, with a 78–90% chance of fatality. Females are more likely to try a different method, such as ingesting poison.[8] Females have more parasuicides. This includes using different methods, such as drug overdose, which are usually less effective.

I am going to go ahead and say that if males die five times as often from suicide, that seems more important than the number of attempts. It is kind of stunning, or at least it should be, to have five boys die for every girl that dies, and for newspapers and experts to make it sound like girls have it worse here. Very big ‘women have always been the primary victims of war. Women lose their husbands, their fathers, their sons in combat’ (actual 1998 quote from Hillary Clinton) energy.

The conflation of suicide rates with forced sex here seems at best highly misleading. The sexual frequency number is rather obviously a reflection of two years where people were doing rather a lot of social distancing. With the end of that, essentially anything social is going to go up in frequency, whether it is good, bad or horrifying – only a 27 percent increase seems well within the range one would expect from that. Given all the other trends in the world, it would be very surprising to me if the rates of girls being subjected to forced sex (for any plausible fixed definition of that) were not continuing to decline.

That implies that in the past, things on such fronts were no-good, horribly terrible, and most of it remained hidden. I do indeed believe exactly this.

Also, can we zoom out a bit? On a historical graph, the suicide rate does not look all that high (scale is suicides per 100,000 children, per year)?

The kids are not okay. The kids in the 1990s were, by some of these g...

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: Suggest new charity ideas for Charity Entrepreneurship, published by CE on March 8, 2023 on The Effective Altruism Forum.

At Charity Entrepreneurship (CE) we launch high-impact nonprofits by connecting entrepreneurs with the effective ideas, training and funding needed to launch and succeed.

Each year our research team collates hundreds and hundreds of ideas for promising new charities. We scour the research, we talk to colleagues in partner organisations and we solicit ideas from academia.

We then vet these ideas and whittle them down to the top ~3. We create detailed reports about these top ideas and then we recruit, train and fund entrepreneurs to take these ideas and make them a reality. In 5 years we've launched over 23 exceptional organisations. You can read more about charities we incubated here.

In 2023 Charity Entrepreneurship will be researching two new cause areas: mass media interventions and preventative animal advocacy

We want your ideas!!!

Prize: If you give us an idea which leads to a new charity then you will win: A copy of the Charity Entrepreneurship handbook, a box of vegan cookies and $100 to a charity of your choice. And more importantly, a new charity will be started!

Notes: If multiple people submit the same idea we will give the award to the first submission. Max 5 prizes will be awarded. If you submit an idea already on our list, you are still eligible for a prize.

Please submit your ideas into this form by end of the day Sunday 12th March.

[SUBMIT]

Mass media

The cause area:

By ‘mass media’ intervention we refer to (1) social and behaviour change communication campaigns delivered through (2) mass media aiming to improve (3) human wellbeing.

Definitions:

1. Social and Behavior Change Communication (SBCC) – The strategic use of communication to promote positive outcomes, based on proven theories and models of behavior change (more here).

2. Mass media – Mass communication modes that reach a very large audience where any targeting of segmented audiences can be mass applied (e.g. would include online advertising that targets relevant audiences but not posters in health centers) (more here). Examples include: TV, Radio, Mobile phones, Newspapers, Outdoor advertising.

3. Human wellbeing – For our purposes, wellbeing refers broadly to areas of human health, development and poverty.

Headline metric

Our key metrics for the quantitative side of this research will be DALYs averted or % income increases. We will likely compare across these metrics using a moral weight formula. We may set our own moral weights or use recent moral weights by GiveWell (e.g. from here). Note that the use of this as a headline metric does not mean that other factors (autonomy, environmental effects, suffering not captured by DALYs) are excluded, although they may not be explicitly quantified.

Scope limitations

Currently, all interventions that could reasonably be considered mass media are in scope. If in doubt, assume it is in scope.

Example ideas

Note: there is no guarantee that any of the following ideas make it past the initial filter

Promoting healthier diets

Promoting CBT tools for stress

Messaging against tobacco use with resources for quitting

Signposting to available support in cases of abuse, violence, etc.

Anti-suicide campaigns

Changing HIV attitudes

Promotion cancer screening (cervical, breast, bowel, prostate, etc.)

Information campaigns about criminal politicians

Encouraging lower sugar consumption

Preventative animal advocacy

The cause area:

The focus this year is on intervention and policies that prevent future harms done to animals as opposed to solving current problems. We will be looking for interventions that as well as having some short run evidence of impact will prevent future problems i.e. have the biggest impact on farmed animals in the future, say ...

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: AI Safety in a World of Vulnerable Machine Learning Systems, published by AdamGleave on March 8, 2023 on LessWrong.

Even the most advanced contemporary machine learning systems are vulnerable to adversarial attack. The safety community has often assumed adversarial robustness to be a problem that will be solved naturally as machine learning (ML) systems grow more capable and general. However, recent work has shown that superhuman systems in a narrow domain such as AlphaZero are highly vulnerable to adversarial attack, as are general but less capable systems like large language models. This raises the possibility that adversarial (worst-case) robustness will continue to lag behind average-case capabilities. In other words, transformative AI systems are likely to be exploitable.

Exploitability will cause a wide variety of current alignment proposals to fail. Most extant agendas seek to align the main ML system with the assistance of helper ML systems. The main ML system is the primary system that takes actions in the world (e.g. interacting with users), with the helper ML systems acting as scaffolding to train and/or verify the main ML system. These alignment schemes will fail if the helpers are exploited by the main system – and we expect helpers to be vulnerable to exploitation (see Contemporary ML systems are exploitable by default).

In Table 1 we present a subjective risk matrix for a range of popular alignment agendas, evaluating the degree to which main ML systems have the ability and incentive to exploit the helper. We find many alignment agendas have a high risk of exploitation, with all having at least some risk.

Alignment AgendaMain System’s Ability to Exploit HelperMain System’s Incentive to Exploit HelperRisk of ExploitRL on learned reward model (e.g. RLHF, IRL)MediumHighHighScalable oversight (e.g. recursive reward modeling,AI safety via debate)MediumHighHighImitation learning (e.g. behavioral cloning, supervised fine-tuning)MediumLowLow-MediumImitative Iterated Distillation and AmplificationHighLowMediumAuditing Tool (e.g. Adversarial Testing, Transparency)LowMediumLow-Medium

Table 1: Subjective risk matrix for popular alignment agendas (see next section), using a helper ML system to assist with aligning the main ML system that will eventually be deployed.

We are most concerned by vulnerabilities in the helpers as this can impact the alignment of the main system. By contrast, an aligned but adversarially exploitable main system would not necessarily pose a danger, especially if the main system can recursively self-improve to fix itself. However, there is a possibility that even superintelligent systems cannot attain adversarial robustness. This would be a volatile situation, which could conceivably collapse into chaos (systems frequently exploiting each other), an implicit equilibrium (e.g. mutually assured destruction), or an explicit agreement (e.g. all AI systems self-modify to commit to not exploiting one another).

We see two possible approaches to fixing this: improving adversarial robustness, or developing fault tolerant alignment methods that can work even in the presence of vulnerable ML systems. We are most excited by fault tolerant alignment, as it is highly neglected and plausibly tractable, although further work is needed to solidify this approach. By contrast, adversarial robustness is an area that has received significant attention from the ML research community (low neglectedness)[1] but with only modest progress (low to medium tractability).

In the remainder of this document, we will argue that systems are exploitable by default, explore the implications this has for alignment agendas in several different scenarios, and outline several research directions we are excited by.

Alignment agendas need robustness

Most alignment schemes implicitl...

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: Evidence on how cash transfers empower women in poverty, published by GiveDirectly on March 8, 2023 on The Effective Altruism Forum.

Donations to GiveDirectly put power in the hands of recipients, 62% of whom are women. On International Women’s Day, hear directly from women and girls in poverty in Malawi about the unique ways that direct cash empowers them:

This impact is more than anecdotal; research finds that cash aid lets women improve their lives in many ways. Below, we break down the evidence by story.

Maternal & infant health

Lenita - “When I was pregnant, I would fall sick [and] could not afford the fare to go to the hospital.”

Studies find that cash can:

Increase the use of health facilities.

Improve birth weight and infant mortality – one study found GiveDirectly’s program reduced child mortality by ~70% and improved child growth.

Education & domestic violence

Agatha - “My husband was so abusive... so I left him and went back to try to finish school.”

Studies find that cash can:

Reduce incidents of physical abuse by a male partner of a woman – one study found GiveDirectly’s program reduced physical intimate partner violence.

Increase school attendance for girls.

Decision-making power

Beatrice - “My husband and I always argued. about how to spend what little money we had. Now, when we receive the money, we plan together.”

Studies find that cash can:

Increase a woman’s likelihood of being the sole or joint decision-maker.

Entrepreneurship & savings

Anesi - “With the businesses I started, I want to buy land for my children so they will never forget me.”

Studies find that cash can:

Increase entrepreneurship – one study of GiveDirectly’s program found new business creation doubled. For more on female entrepreneurs, watch

Increase the number of families saving and the amount they saved – one study of GiveDirectly’s program found women doubled their savings. To learn about women's savings groups, watch

Elderly support

Faidesi - “Now that I am old, I can’t farm and often sleep hungry. I would have been dead if it wasn’t for these payments.”

Studies find that cash can:

Reduce the likelihood of having had an illness in the last three months – one study in Tanzania found cash reduced the number of doctor visits made by women over 60.

Bastagli et al 2016

Siddiqi et al 2018

McIntosh & Zeitlin 2018

Haushofer et al 2019

McIntosh & Zeitlin 2020

Pega et al 2017

Evans et al 2014

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: Squeezing foundations research assistance out of formal logic narrow AI., published by Donald Hobson on March 8, 2023 on The AI Alignment Forum.

Suppose you have a ML model trained to output formal proofs. Maybe you start with ZFC and then add extra tokens for a range of common concepts. (along with definitions. ). So a human mathematician needs to type in the definition of a gradient in terms of limits, and the definition of limits in terms of epsilon and delta, and the definition of the real numbers in terms of dedekind cuts. All the way back to ZFC. The human needn't type any proofs, just the definitions.

The model could be trained by generating random syntactically correct strings of tokens, and trying to prove or disprove them. (Remember, we have added the notion of a gradient to the token pool, plenty of the random questions will involve gradients) Hopefully it forms intermediate theorems and heuristics useful towards proving a wide class of theorems.

Computer programs can be described as mathematical objects. So the human adds some tokens for lisp programs, and a few definitions about how they behave to the token pool.

"Will program X do Y?" is now a perfectly reasonable question to ask this model.

This is where the magic happens. You give your system a simple toy problem, and ask for short programs that solve the toy problem, and about which many short theorems can be proved. Maybe you do gradient descent on some abstract latent space of mathematical objects. Maybe an inefficient evolutionary algorithm selecting both over the space of programs and the theorems about them. Maybe "replace the last few layers, and fine tune the model to do a new task", like RLHF in ChatGPT.

Now I don't expect this to just work first time. You will want to add conditions like "ignore theorems that are true of trivial programs (eg the identity program)" and perhaps "ignore theorems that only take a few lines to prove" or "ignore theorems so obvious that a copy of you with only 10% the parameters can prove it". For the last one, I am thinking of the programmers actually training a mini version with 10% the parameters, and running some gradients through it. I am not thinking of the AI reasoning about code that is a copy of itself.

The AI model should have a latent space. This can let the programmers say "select programs that are similar to this one" or "choose a program about which theorems close to this theorem in latent space can be proved".

The idea of this is that

Asking questions should be safe. There are a bunch of different things we can optimize, and it should be safe to adjust parameters until it is proving useful results not trivialities. The AI doesn't have much information about human psychology, or about quantum physics or the architecture of the processor it's running on. Gradient descent has been pushing it to be good at answering certain sorts of question. There is little to no advantage to being good at predicting the questions or figuring out what they imply about the people asking them.

With a bit of fiddling, such a design can spit out interesting designs of AI, and theorems about the designs. This isn't a foolproof solution to alignment, but hopefully such help makes the problem a lot easier.

It is ABSOLUTELY NOT SAFE to throw large amounts of compute at the programs that result. Don't have anything capable of running them installed. The programs and the theorems should be read by humans, in the hope that they are genius insights into the nature of AI. The textbook from the future. Humans can then use the insights to do... something.

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: A Brief Defense of Athleticism, published by Wofsen on March 7, 2023 on LessWrong.

Non-athletic thinkers are myopic. I see at least three very important reasons any rationalist must value exercise.

As someone who has trained extremely hard in distance running, sports in general and athletic conditioning in particular create an intuitive understanding of the fallaciousness of Cartesian Dualism and the accuracy of materialism in the sense that the mind is merely a part of the body.

Physical challenges also force one to understand the limitations of Kahneman's "system two." For example, one may know what it means to not start a race too quickly and then "die" (running jargon for running out of steam prematurely), but repeated failures in actual races teach one to realize the limitations of one's rationality, especially under stress (even if distance running is classified as eustress rather than distress; remember, cortisol levels & arousal are the same in either case).

Conditioning has been shown to increase cognitive performance. Results of a quick google finds: This study notes better reaction time in fit adolescents (no causal link, n = 30). This review shows decreased "system 2" aging in physically fit elderly individuals (causality likely, preponderance of evidence). Acute exercise (<1 hour) has also been shown to cause immediate improvements in decision making. I recommend a deep dive into the mechanistic research into the causes of this for structuralists. But any Bayesian looking to be more rational ought to make an update and start working out.

(Edit: Grammar)

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: Suggestion: A workable romantic non-escalation policy for EA community builders, published by Severin on March 8, 2023 on The Effective Altruism Forum.

Last year, I attended an Authentic Leadership Training with the Authentic Relating org Authentic Revolution, and was course lead mentee for a second iteration.

One thing that struck me about AuthRev's ways is their approach to policing romantic relationships between facilitators and participants, within a community whose personal/professional overlap is even stronger than EA’s.

They have a romantic non-escalation policy that goes roughly like this:

For three months after a retreat, and for one month after an evening event, facilitators are prohibited from engaging romantically, or even hinting at engaging romantically, with attendees. The only exception is when a particular attendee and the facilitator already dated beforehand.

These numbers are drawn from experience: As some people have most of their social life within the community, longer timelines are so unworkable that the default is to just ignore them and do everything in secret. Shorter timelines, however, tend to be insufficiently effective for mitigating the problems this policy tries to address.

Granted, Authentic Relating is a set of activities that is far more emotionally intense than what usually happens at EA events. However, I think there are some reasons for EA community builders to adhere to this policy anyway:

Romance distracts from the cause. Attendees should focus on getting as much ea-related value as possible out of EA events, and we as organizers should focus on generating as much value as possible. Thinking about which hot community builder you can get with later distracts from that. And, thinking about which hot participant you can get with later on can lead to decisions way more costly than just lost opportunities to provide more value.

None of us are as considerate and attuned in our private lives as when doing community building work. Sometimes we don't have the energy to listen well. Sometimes we really need to vent. Sometimes we are just bad at communication when we don't pay particular attention to choosing our words. The personas we put up at work just aren't real people. If people fall in love with the version of me that they see leading groups, they will inevitably be disappointed later.

Power differentials make communication about consent difficult. And the organizer/attendee-separation creates a power differential, whether we like it or not. The more power differential there is, the more important it is to move very slowly and carefully in romance.

Status is sexy. Predatorily-minded people know this. Thus, they are incentivized to climb the social EA ladder for the wrong reasons. If we set norms that make it harder for people to leverage their social status for romantic purposes, we can correct for this. That is, as long as our rules are not so harsh that they will just be ignored by default.

Though a part of me finds this policy inconvenient, I think it would be a concerning sign if I weren’t ready to commit to it after I saw it’s value in practice. However, EA is different from AR, and a milder/different/more specified version might make more sense for us. Accordingly, I’ll let the idea simmer a bit before I fully commit.

Which adjustments would you make for our context? Some specific questions I have:

AR retreats are intensely facilitated experiences. During at least some types of EA retreats, the hierarchies are much flatter, and participants see the organizers "in function" only roughly as much as during an evening-long workshop. Does this justify shortening the three months, e.g. to one month no matter for which type of event?

I'd expect that the same rule should apply for professional 1-on-1s, for example EA career coaching....