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: Why Agent Foundations? An Overly Abstract Explanation, published by johnswentworth on March 25, 2022 on LessWrong.

Let’s say you’re relatively new to the field of AI alignment. You notice a certain cluster of people in the field who claim that no substantive progress is likely to be made on alignment without first solving various foundational questions of agency. These sound like a bunch of weird pseudophilosophical questions, like “what does it mean for some chunk of the world to do optimization?”, or “how does an agent model a world bigger than itself?”, or “how do we ‘point’ at things?”, or in my case “how does abstraction work?”. You feel confused about why otherwise-smart-seeming people expect these weird pseudophilosophical questions to be unavoidable for engineering aligned AI. You go look for an explainer, but all you find is bits and pieces of worldview scattered across many posts, plus one post which does address the question but does so entirely in metaphor.

Nobody seems to have written a straightforward explanation for why foundational questions of agency must be solved in order to significantly move the needle on alignment.

This post is an attempt to fill that gap. In my judgment, it mostly fails; it explains the abstract reasons for foundational agency research, but in order to convey the intuitions, it would need to instead follow the many paths by which researchers actually arrive at foundational questions of agency. But a better post won’t be ready for a while, and maybe this one will prove useful in the meantime.

Note that this post is not an attempt to address people who already have strong opinions that foundational questions of agency don't need to be answered for alignment; it's just intended as an explanation for those who don't understand what's going on.

Starting Point: The Obvious Stupid Idea

Let’s start from the obvious stupid idea for how to produce an aligned AI: have humans label policies/plans/actions/outcomes as good or bad, and then train an AI to optimize for the good things and avoid the bad things. (This is intentionally general enough to cover a broad range of setups; if you want something more specific, picture RL from human feedback.)

Assuming that this strategy could be efficiently implemented at scale, why would it not produce an aligned AI?

I see two main classes of problems:

In cases where humans label bad things as “good”, the trained system will also be selected to label bad things as “good”. In other words, the trained AI will optimize for things which look “good'' to humans, even when those things are not very good.

The trained system will likely end up implementing strategies which do “good”-labeled things in the training environment, but those strategies will not necessarily continue to do the things humans would consider “good” in other environments. The canonical analogy here is to human evolution: humans use condoms, even though evolution selected us to maximize reproductive fitness.

Note that both of these classes of problems are very pernicious: in both cases, the trained system’s results will look good at first glance.

Neither of these problems is obviously all that bad. In both cases, the system is behaving at least approximately well, at least within contexts not-too-different-from-training. These problems don’t become really bad until we apply optimization pressure, and Goodhart kicks in.

Goodhart’s Law

There’s a story about a Soviet nail factory. The factory was instructed to produce as many nails as possible, with rewards for high numbers and punishments for low numbers. Within a few years, the factory was producing huge numbers of nails - tiny useless nails, more like thumbtacks really. They were not very useful for nailing things.

So the planners changed the incentives: they decided to reward the factory f...

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: Some observations about ELK, published by leogao on March 26, 2022 on The AI Alignment Forum.

This document outlines some of my current thinking about ELK, in the form of a series of observations I have made that inform my thinking about ELK.

Editor’s note: I’m experimenting with having a lower quality threshold for just posting things even while I’m still confused and unconfident about my conclusions, but with this disclaimer at the top. Thanks to AI_WAIFU and Peter Barnett for discussions.

One way we can think of ELK is we have some set of all possible reporters, each of which takes in the latent states of a world model and outputs, for the sake of concreteness, the answer to some particular fixed question. So essentially we can think of a reporter+model pair as assigning some answer to every point in the action space. Specifically, collapsing possible sequences of actions into just a set of possible actions and only worrying about one photo doesn’t result in any loss of generality but makes things easier to talk about.

We pick some prior over the set of possible reporters. We can collect training data which looks like pairs (action, answer). This can, however, only cover a small part of the action space, specifically limited by how well we can “do science.”

This prior has to depend on the world model. If it didn’t, then you could have two different world models with the same behavior on the set where we can do science, but where one of the models understands what the actual value of the latent variable is, and one is only able to predict that the diamond will still appear there but can’t tell whether the diamond is still real. A direct translator for the second will be a human simulator for the first.

More generally, we don’t really have guarantees that the world model will actually understand what’s going on, and so it might genuinely believe things about the latent variable that are wrong. Like, we might have models which think exactly like a human and so they don't understand anything a human wouldn’t understand, so the direct translator for this model would also simultaneously be a human simulator, because the model literally believes exactly what a human would believe.

There are more than just human simulators and direct translators that are consistent with the training data; there are a huge number of ways the remaining data can be labeled. This basically breaks any proposal that starts with penalizing reporters that look like a human simulator.

I basically assume the “human simulator” is actually simulating us and whatever science-doing process we come up with, since this doesn’t change much and we’re assuming that doing science alone isn’t enough because we’re looking at the worst case.

So, for any fixed world model, we need to come up with some prior over all reporters such that after picking all the reporters consistent with our data (action, answer), the maximum of the resulting posterior is the direct translator.

A big assumption I’m making is that the KL distance between our prior and the complexity prior has to be finite. My intuition for why this has to be is that a) universality of the complexity prior, and b) even if our prior doesn’t have to cover all possible reporters, it still seems reasonable that as complexity increases the number of reporters that our prior accepts as nonzero needs to increase exponentially, which gets something similar to universality of complexity on the subset of plausible reporters. I think this point is the main weak point of the argument.

Since the amount of data we can get is bounded by our ability to do science, the amount of information we can get from that is also bounded. However, since we’re assuming worst case, we can make the complexity difference between the direct translator and all the other junk arbitraril...

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: Ukraine Post #7: More Data and Peace Terms, published by Zvi on March 25, 2022 on LessWrong.

I was mostly happy with the Covid-style approach I tried out last time. This is an update of ‘all the things worth noting since then’ but excluding prediction markets. I plan to do a prediction market update next, and keep those two tracks distinct. I do talk a bit about how I would characterize various final settlements, with an aim of getting prediction markets that can better reflect the nature of the final outcome.

Note that there is also an essentially endless list of claims of Russian war crimes and crimes against humanity, and I’m almost entirely not including them. You do not need me to tell you about this aspect of what is going on.

Lots of detail keeps coming in, and there are many incremental changes and updates, but the central facts seem similar to how they seemed as of my last post, so I won’t be re-summarizing them. Things are no longer moving quite so fast.

Potential Donations

From Bloomberg on March 15, an economist talks about fighting for Ukraine.

Here is the money quote, file under Huge If True (but I can’t evaluate it).

We don’t need food or paper towels. What the army needs is munitions and people need medical supplies, specific medical supplies. Most people die from blood loss after a cluster bomb or after some kind of ballistic missile falls. There may be 20 or 200 people wounded. It’s a little bit like when an airplane cabin loses pressure, the masks fall down — and what you need to do is to put the mask on yourself and then help others. So in this case, the mask analogy is a medical kit, which allows you to stop bleeding. So you really have to ensure that you are not bleeding, and then that people next to you are not bleeding.

So you have to have a lot of these medical kits, and they usually cost 10 or 20 bucks. But now of course they cost 100 bucks because it’s surge pricing and no one can deliver. So we are trying to focus on this specifically. Civilian authorities need them and even railroads are asking for these medical kits because evacuation trains get shelled and people die without this specific kit. We will deliver them. We have connections, there’ll be no [extra overhead charges]. It doesn’t get stolen on the way. And [your donation] is also tax deductible in the U.S. So for $100, you’ll save a life and get your taxes back.

There is a catch. It’s not so easy to actually get things where they need be.

And you need to get suppliers. In war, there are so many intermediaries and fees. So you have to establish a procurement department to figure out who is serious. Once you have suppliers, you have to figure out all the [wartime] export licenses, all these government regulations. And of course things get stolen on the way. For example, one large, well-known American charity sends 95 pallets of medical supplies to Ukraine. When it arrives, it’s only two pallets because 93 of them got stolen somewhere. Not in Ukraine — either in Poland or even in New Jersey. So you have to watch for this. You actually have to put your own people in Warsaw or in New Jersey, in Israel, in Sweden, to check what has been loaded at every point so that it doesn’t get stolen. It’s a logistical nightmare. And when you finally get it to Ukraine, people are shooting at you at checkpoints.

One specific thing: We need 307,000 medical kits. I have the specification. Let’s say Israel can only supply 30,000 and Canada probably can supply 20 or 30,000. But we have suppliers who can provide the medical kits. We give this specification to [Ukraine’s] Ministry of Health, and our charitable foundation will pay. So tag me or email me or ping me on Twitter — and then donate, please donate.All the fundraising goes directly to logistics. I have a website at the university of the charitable...

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: Effectiveness is a Conjunction of Multipliers, published by Thomas Kwa on March 25, 2022 on The Effective Altruism Forum.

Epistemic status: Not new material, but hopefully points more directly at key EA intuitions.

Ana is a hypothetical junior software engineer in Silicon Valley making $150k/year. Every year, she spends 10% of her income to anonymously buy socks for her colleagues. Most people would agree that Ana is being altruistic, but not being particularly efficient about it. If utility is logarithmic in income, Ana can 40x her impact by giving the socks to a local homeless person instead who has an income of $5000. But in the EA community, we've noticed further multipliers:

40x: giving socks to local homeless people instead of her colleagues

10x more: giving socks to the poorest people in the world (income $500) instead of homeless people

2x more: giving cash (GiveDirectly) instead of socks

8x more: giving malaria nets rather than cash

10x more: farmed animal welfare rather than human welfare.

4x more: working in a more lucrative industry like quant research, working longer hours, and doing salary negotiation to raise her salary to $600k

8x more: donating 80% instead of 10%

10x more: taking on risk to shoot for charity entrepreneurship or billionairedom, producing $6M of expected value yearly

Total multiplier: about 20,480,000x

I think that many people new to EA have heard that multipliers like these exist, but don't really internalize that all of these multipliers stack multiplicatively. If Ana hits all of these bonuses, she will have a direct impact 20,480,000 times larger than giving her socks to random colleagues. If she misses one of these multipliers, say the last one, Ana will still have a direct impact 2,048,000 times larger than with the initial socks plan. This sounds good until you realize that Ana is losing out on 90% of her potential impact, consigning literally millions of chickens to an existence worse than death. To get more than 50% of her maximum possible impact, Ana must hit every single multiplier. This is one way that reality is unforgiving.

Multipliers result from judgment, ambition, and risk

Good judgment: responsible for multipliers (1) through (4), making the impact 80,000 times larger, and is implicit in (8) too, because going through with a bad for-profit or charity startup idea could be of zero or even negative value.

Ambition: responsible for multipliers (6) through (8), making her expected impact 320x larger.

Willingness to take on risk is mostly relevant in (8), though you could think of (5) as having risk from moral uncertainty.

This example is neartermist to make the numbers more concrete, but the same principles apply within longtermism. For a longtermist, good judgment and ambition are even more critical. It's difficult to tell the difference between a project that reduces existential risk by 0.02%, a project that reduces x-risk 0.002%, and a worthless project, so you need excellent judgment to get within 50% of your maximum impact. Ambition is in some sense what longtermism is all about-- it would be silly to claim that the value of the future is orders of magnitude larger than most things in the present and not aim for that value. And taking on risk allows hits-based giving whether in neartermism or longtermism.

More generally, actions, especially complicated actions and research directions, are an extremely high-dimensional space. If actions are vectors and the goodness of an action is its cosine similarity to the best action, and your action is 90% as good as the optimum (25° off the best path) in each of 50 orthogonal directions, the amount of good you do is capped at 0.9^50 = 0.005x the maximum.

Implications

It's very difficult to take an arbitrary project that you're excited about for other reasons, and tweak it to "mak...

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: Questions That Lead to Impactful Conversations, published by Akash on March 24, 2022 on The Effective Altruism Forum.

Context

We often know the answers to questions that can help us dramatically improve our impact/productivity/happiness/health/etc.

This became particularly salient to us after we learned techniques and saw how helpful they were. Examples include:

Resolve Cycles (think of a problem in your life, set a five minute timer, and solve your problem, or set a five minute timer to think of a plan to solve your problem, and then set another five minute timer and pick a harder problem)

Murphyjitsu (ask yourself “How surprised would I be if my current plan were unsuccessful after the deadline passed?”, and “How can I make myself more surprised?”, and then repeatedly iterating).

Hamming Questions (asking "what are the most important problems in your field? Why aren't you working on them?")

We (Akash and Kuhan) also often ask questions to try to improve others’ impact/productivity/happiness/health, form close relationships, and produce valuable conversations. Questions can help us understand new concepts and paradigms, generate novel ideas and connections, notice our confusion and holes in our understanding, and make friends.

We wanted to compile a list of the best questions to ask in different scenarios - one-off questions, questions that should be asked regularly, both to ourselves and others, questions to make meetings better, and more. We’d be excited for readers to also suggest questions they’ve gotten mileage out of/have enjoyed asking and answering in the comments.

The NYT published a list of questions that lead to love. Here's a list of questions that (often) lead to impactful conversations.

List of Questions

Life Optimization Call Questions

credit to Neel Nanda for bringing many of these questions to our attention

What do you think are the most important problems? Why aren’t you working on them? (Hamming questions)

What is most holding you back? What are your biggest bottlenecks?

Suppose that this meeting increases your lifetime productivity/impact/happiness/motivation by 2x. What happened?

What are the projects you’ve spent >=10 hours on in the last month? Rate each out of 10 for impact, and write a brief theory of impact. Do you endorse this?

What is your personal theory of impact? How do your actions concretely reduce existential risk?

Do you consistently feel motivated about what you work on? What goes wrong? Give a concrete example? How frequently do you use obligations to motivate yourself? Do you feel guilty and anxious all of the time?

What probability would you put that you have any of insomnia, ADHD, depression, anxiety, bipolar? How are your physical health, sleep, exercise, diet, hydration? What niggling health issues are you ignoring? Are there other important health issues I’m missing? Have you spent at least a full-time week solving this? Why not?

Please fill out this depression questionnaire and anxiety questionnaire and consider next steps depending on the results.

What is the biggest mistake you’ve made in the last 6 months? Have you successfully changed the default path of the world such that such a mistake is very unlikely to happen again?

Suppose you leave this conversation feeling super inspired and confident your life has changed, and in practice nothing changes. How surprised are you?

Bonus: Has this happened before? What went wrong?

Suppose you leave this conversation with some concrete action points that are obviously good ideas that will majorly improve your life. What are they? Make them concrete and give deadlines.

What are the biggest and sharpest trade-offs in your life right now? What are the costs and benefits of each option? Set a 5 minute timer and brainstorm third alternatives, that capture some of the benefits of each a...

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: Pre-announcing a contest for critiques and red teaming, published by Lizka on March 25, 2022 on The Effective Altruism Forum.

We’re writing this post to say that we’re interested in running a contest on the Forum for writing that critically engages with theory or work from the EA community.

We are: Lizka (the Content Specialist at CEA), Joshua, and Fin.

Consider this a pre-announcement: we’re still figuring out the details, but we want to commit right now to supporting high-quality critical work within EA — including critiques, questioning, red teaming, and minimal-trust investigations. We think a contest could make a good start.

Why do we think this matters? In short, we think there are some reasons to expect good criticism to be undersupplied relative to its real value. And that matters: as EA grows, it’s going to become increasingly important that we scrutinize the ideas and assumptions behind key decisions — and that we welcome outside experts to do the same.

So we want to run a contest to incentivize and reward the most thoughtful and most action-relevant critical work (cross-)posted to the Forum. The Creative Writing Contest last year would be a good point of comparison, but we’re hopeful the prize pool will be significantly larger this time — perhaps up to around $100k if enough submissions really meet the bar. The judging panel is taking shape already, and we’re looking forward to sharing it.

We’re excited to read about errors in widely-cited analyses, gaps in EA’s philosophical toolset, missing crucial considerations, independent reassessments of key claims, evaluations of EA organizations, suggestions for how to further improve community norms and institutions, and efforts to communicate and ‘steelman’ existing criticisms for an EA audience.

The judging criteria aren’t finalized, but we’re imagining an ideal submission would be:

Critical — the piece takes a critical or questioning stance towards some aspect of EA, theory or practice;

Important — the issues discussed really matter for our ability to do the most good as a movement;

Novel — the piece presents new arguments, or otherwise presents familiar ideas in a new way;

Constructive — we can see how the work is decision-relevant (directly or indirectly). Bonus points for concrete solutions.

We don't expect that every winning piece needs to do well at every one of these criteria, but we do value them all.

We also want to reward reasoning transparency / ‘epistemic legibility’, clarity of writing, ‘punching up’, awareness of context, and a scout mindset. On the other hand, we don’t want to encourage personal attacks or diatribes that are likely to produce much more heat than light. And we hope that subject-matter experts who don’t typically associate with EA find out about this and share insights we haven’t yet heard.

Should you hold off on posting until the contest is properly announced in order to ensure eligibility? You don’t have to: if we go ahead, we’ll accept pieces posted from this week (beginning March 21st, 2022) onwards.

Another reason we’re ‘pre-announcing’ this idea is to hear some initial thoughts from Forum users. We think it's especially important to get the details right here, so we’d be interested to hear what considerations we should bear in mind, that we might otherwise miss. Thanks!

Joshua Teperowski Monrad elaborates in his post ‘Rowing and Steering the Effective Altruism Movement’. Fin Moorhouse describes the rationale for a contest in his ‘EA Project Ideas’ post. We also recommend Linch’s shortform post on ‘red teaming papers as an EA training exercise’.

This should complement, rather than duplicate, the ‘Red Team Challenge’ recently announced by Training for Good.

Of course, it might make sense to wait to read the full announcement and the finalized judging criteria.

Thanks for listen...

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: Rational and irrational infinite integers, published by Viliam on March 23, 2022 on LessWrong.

Epistemic status: a fine line between genius and madness

Let's start with the observation that it is much easier to use digits before the decimal point than the digits after it. A six years old child could calculate e.g. 11 + 2, but calculating reliably 0.11 + 0.2 will take a few more years of education.

From this perspective, it is quite ironic that we typically use numbers that only have a few digits before the decimal point, and many, sometimes infinitely many digits after the decimal point, such as 1/3 = 0.333333333... = 0.¯¯¯3.

Obviously, it would be better the other way round.

Technically, we already have infinitely many digits in front of every number. But they are all zeroes; and by convention, we do not write them. Writing a zero or multiple zeroes in front of a number does not change its value; the number 007 mathematically means 7. Therefore, technically, we could write the number 7 with infinitely many digits in front of it, like this: ¯¯¯07.

But it seems like a waste of space to only ever fill those infinitely many pre-decimal places with zeroes. What about using some other digit instead?

Consider the number ¯¯¯9, which consists of infinitely many nines. Written in full, it would be like this: ¯¯¯9 = ...999999999.

The value of this number is -1.

How can we know that? Simply, just calculate ...999999999 + 1 = ...000000000, which is zero. (Zero with an arbitrary number of zeroes in front of it is zero: 0 = 00 = 000 = ¯¯¯0.)

Similarly, ...999999998 = ¯¯¯98 is -2, because ¯¯¯98+2=¯¯¯0. Also ¯¯¯97=−3, ¯¯¯96=−4, and so on.

Seems like we have invented a new syntax for negative numbers. Instead of writing a minus symbol in front of a number, just let the digits underflow. If an odometer in your car showed only nines, you would know what happens if you go the extra mile.

What if we fill the infinite pre-decimal places with digits other than zeroes and nines? For example, how much is ¯¯¯1?

I believe this number is -1/9.

We can easily verify it by multiplying the number by 9 and adding 1.¯¯¯1×9+1=¯¯¯9+1=¯¯¯0.

Okay, this seems strange. How did a fraction smaller than one suddenly become an infinite integer? But as you can clearly see, it did. Remember, our goal was to move the infinite digits behind the decimal point to the front of it. And so far, we are succeeding!

We can similarly get other negative fractions by dividing −1=¯¯¯9 by the denominator, and multiplying the result by the numerator. For example, −1/3=¯¯¯3, −2/3=¯¯¯6, −1/7=¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯142857. Check it, ¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯142857×7+1=¯¯¯9+1=0.

To obtain a positive fraction, we can subtract the negative fraction from zero. For example, 1/9=0−(−1/9)=0−¯¯¯1=¯¯¯89. Check it, ¯¯¯89×9=¯¯¯01. Seems correct.

But there is still problem with fractions like 1/2 or 1/5. There is no integer, not even infinite one, such that if we multiplied it by 2 or by 5, the last digit of the result would be 1. (We can easily check it, because the last digit of the result only depends on the last digit of the original number.) Therefore, some decimal places seem inevitable: 1/2 remains 0.5, and 1/5 remains 0.2. We have not eliminated all decimal places from the fractions, only the infinite sequences thereof.

Using the new syntax, any fraction will have at most finite number of decimal places. To see why it is so, extract the powers of 2 and 5 from the denominator, like this: nd=n2i5jk, then calculate the infinite integer nk and divide it by 2i and 5j. The number of decimal places of the result will be max(i, j). That kinda sucks, but whatever.

By the way, −1/2=¯¯¯9.5, and −1/5=¯¯¯9.8. I assume it is obvious why.

(Having an infinite number of digits both before and after the decimal point would lead to ambiguity. For example, 1/3 could be ...

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: Experimental longtermism: theory needs data, published by Jan Kulveit on March 24, 2022 on LessWrong.

The series explains my part in the response to COVID, my reasons for switching from AI alignment work to the COVID response for a full year, and some new ideas the experience gave me. While it is written from my (Jan Kulveit's) personal perspective, I co-wrote the text with Gavin Leech, with input from many others.

The first post covers my main motivation: experimental longtermism.

Feedback loop

Possibly the main problem with longtermism and x-risk reduction is the weak and slow feedback loop.

(You work on AI safety; at some unknown time in the future, an existential catastrophe happens, or doesn’t happen, as a result of your work, or not as a result of your work.)Most longtermists and existential risk people openly admit that the area doesn't have good feedback loops. Still, I think the community at large underappreciates how epistemically tricky our situation is. Disciplines that lack feedback from reality are exactly the ones that can easily go astray.

But most longtermist work is based on models of how the world works - or doesn’t work. These models try to explain why such large risks are neglected, the ways institutions like government or academia are inadequate, how various biases influence public perception and decision making, how governments work during crises, and so on. Based on these models, we take further steps (e.g. writing posts like this, uncovering true statements in decision theory, founding organisations, working at AI labs, going into policy, or organising conferences where we explain to others why we believe the long-term future is important and x-risk is real).

Covid as opportunity

Claim: COVID presented an unusually clear opportunity to put some of our models and theory in touch with reality, thus getting more "experimental" data than is usually possible, while at the same time helping to deal with pandemic. The impact of the actions I mentioned above is often unclear even after many years, whereas in the case of COVID impact of similar actions was observable within weeks and months.For me personally, there was one more pull. My background is in physics, and in many ways, I still think like a physicist. Physics - in contrast to most of maths and philosophy - has the advantage of being able to put its models in touch with reality, and to use this signal as an important driver in finding out what's true. In modern maths, (basically) whatever is consistent is true, and a guiding principle for what's important to work on is a sense of beauty. To a large extent, the feedback signal in philosophy is what other philosophers think.

(Except when a philosophy turns into a political movement - then the signal comes from outcomes such as greater happiness, improved governance, large death tolls, etc.) In both maths and philosophy, the core computation mostly happens "in” humans. Physics has the advantage that in its experiments, "reality itself" does the computation for us.

I miss this feedback from reality in my x-risk work. Note that many of the concrete things longtermists do, like posting on the Alignment Forum or explaining things at conferences, actually do have feedback loops. But these are usually more like maths or philosophy: they provide social feedback, including intuitions about what kinds of research are valuable. One may wonder about the problems with these feedback loops, and what kind of blind-spots or biases they entail.

At the beginning of the COVID crisis, it seemed to me that some of our "longtermist" models were making fairly strong predictions about specific things that would fail - particularly about inadequate research support for executive decision-making. After some hesitation, I decided that if I trusted these models for x-risk...

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: Critique of OpenPhil's macroeconomic policy advocacy, published by Hauke Hillebrandt on March 24, 2022 on The Effective Altruism Forum.

Summary

Recently, some defended OpenPhil-funded advocacy for looser (i.e. more expansive) macroeconomic policy (i.e. both fiscal and monetary policy). This is because in the 2010s, and also after covid hit, looser policy and higher inflation were better than lost growth and unemployment.[1]

While this might be true, here I argue that OpenPhil grants in '21 caused looser macroeconomic policy, which contributed to overshooting on inflation.

My model is based on the following claims:

Short-term inflation is too high, volatile, and dispersed

Inflation in poorer countries

Real wages are down

Inflation causes populism

Looser policy is a major cause of inflation

OpenPhil’s advocacy caused looser policy

Strong claim: Risk could have been predicted before the event and loose policy was a mistake ex-ante

Weak claim: Risk was hard to predict and was only a mistake ex-post

To be clear, I'm not arguing that inflation is always worse than unemployment and my ambition here is not to evaluate the program’s overall impact. Rather, I single out grants made in '21, which caused looser policy and might have caused harm on the '21 margin. In other words, given diminishing returns to looser policy in '21, the marginal benefits looser policy was low as it didn’t add many jobs. However, inflation had high costs on the '21 margin like adverse effects in developing countries, reduction of real wages for some populations, which might increase populism and make the US Democrats lose upcoming elections.

I also highlight some general issues with macroeconomic advocacy like reducing central bank independence, problems with lobbying foreign central banks, and having to be very careful and do a lot of analysis before making grants that might have a lot of leverage and are not robustly good (like improving health).

The EA community promotes OpenPhil’s work and funges with their grants,[2] and so it’s important to figure out whether we’re value-aligned.[3] This is also an example of red teaming.

Thanks to Joel Becker, Remmelt Ellen, Fin Moorhouse, Peter McCluskey, Michael Dickens, Glen Weyl, Tyler Cowen, Rob Wiblin, Michael Townsend, Ben Edelman, Andreas Prenner, and Kevin Kuruc for feedback on earlier versions of this manuscript. Errors and opinions are mine.

Short-term inflation is too high, volatile, and dispersed

While average long-term inflation is not too high, it is too high in the short-term. This high inflation volatility and dispersion are causing instability and harm. Even if longer-term inflation is too low, the current spike in inflation is not an effective way of remedying it.

There is no consensus on how high inflation should be. [4] But few economists argue that inflation should be >4%, and many say that the European Central Bank (ECB) should target 2%.[5]

Yet, the IMF's forecasted ~4% inflation in rich countries and ~6% in poorer countries for 2022 (antebellum).[6]

EU inflation hit 6.2% in Feb ‘22.[7] The ECB projects 3.5% in 2022, 1.7% in 2023.[8] In some countries inflation is higher: ~12% in Estonia and Lithuania, 8.5% in Belgium and Slovakia.[9] Eurozone inflation dispersion is normally at 1 SD, now it’s at 2.7 SD.[10] Such excessive inflation dispersion can be problematic because countries cannot adjust their exchange rate.

US inflation hit 7.9% in Jan ‘22.[11] In some small towns in the Midwest and South, inflation is 9% or more.[12] Even if one excludes more volatile prices from inflation like food and energy, inflation, right now, is too high.[13] However, the median prediction of inflation for 2022 by economists is 3.5%.[14]

Markets predict inflation to go down in the next few years: In the US, average inflation is forecast to stay high at an...

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: Butterfly Ideas, published by Elizabeth on March 23, 2022 on The Effective Altruism Forum.

Note: this post was crossposted from Elizabeth's blog, Aceso Under Glass by Lizka, with the author's permission. The author may not see or respond to comments on this post. You can see the original post here. Or “How I got my hyperanalytical friends to chill out and vibe on ideas for 5 minutes before testing them to destruction”

Sometimes talking with my friends is like intellectual combat, which is great. I am glad I have such strong cognitive warriors on my side. But not all ideas are ready for intellectual combat. If I don’t get my friend on board with this, some of them will crush an idea before it gets a chance to develop, which feels awful and can kill off promising avenues of investigation. It’s like showing a beautiful, fragile butterfly to your friend to demonstrate the power of flight, only to have them grab it and crush it in their hands, then point to the mangled corpse as proof butterflies not only don’t fly, but can’t fly, look how busted their wings are.

You know who you are

When I’m stuck in a conversation like that, it has been really helpful to explicitly label things as butterfly ideas. This has two purposes. First, it’s a shorthand for labeling what I want (nurturance and encouragement). Second, it explicitly labels the idea as not ready for prime time in ways that make it less threatening to my friends. They can support the exploration of my idea without worrying that support of exploration conveys agreement, or agreement conveys a commitment to act.

This is important because very few ideas start out ready for the rigors of combat. If they’re not given a sheltered period, they will die before they become useful. This cuts us off from a lot of goodness in the world. Examples:

A start-up I used to work for had a keyword that meant “I have a vague worried feeling I want to discuss without justifying”. This let people bring up concerns before they had an ironclad case for them and made statements that could otherwise have felt like intense criticism feel more like information sharing (they’re not asserting this will definitely fail, they’re asserting they have a feeling that might lead to some questions). This in turn meant that problems got brought up and addressed earlier, including problems in the classes “this is definitely gonna fail and we need to make major changes” and “this excellent idea but Bob is missing the information that would help him understand why”.

This keyword was “FUD (fear, uncertainty, doubt)”. It is used in exactly the opposite way in cryptocurrency circles, where it means “you are trying to increase our anxiety with unfounded concerns, and that’s bad”. Words are tricky.

Power Buys You Distance From The Crime started out as a much less defensible seed of an idea with a much worse explanation. I know that had I talked about it in public it would have caused a bunch of unproductive yelling that made it harder to think because I did and it did (but later, when it was ready, intellectual combat with John Wentworth improved the idea further).

The entire genre of “Here’s a cool new emotional tool I’m exploring”

The entire genre of “I’m having a feeling about a thing and I don’t know why yet”

I’ve been on the butterfly crushing end of this myself- I’m thinking of a particular case last year where my friend brought up an idea that, if true, would require costly action on my part. I started arguing with the idea, they snapped at me to stop ruining their dreams. I chilled out, we had a long discussion about their goals, how they interpreted some evidence, and why they thought a particular action might further said goals, etc.

A week later all of my objections to the specific idea were substantiated and we agreed not to do the thing- but thanks to...