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: 2021 AI Alignment Literature Review and Charity Comparison, published by Larks on December 23, 2021 on The AI Alignment Forum.

cross-posted to the EA forum here.

Introduction

As in 2016, 2017, 2018, 2019 and 2020 I have attempted to review the research that has been produced by various organisations working on AI safety, to help potential donors gain a better understanding of the landscape. This is a similar role to that which GiveWell performs for global health charities, and somewhat similar to a securities analyst with regards to possible investments.

My aim is basically to judge the output of each organisation in 2021 (technically: 2020-12-01 to 2021-11-30) and compare it to their budget. This should give a sense of the organisations' average cost-effectiveness. We can also compare their financial reserves to their 2021 budgets to get a sense of urgency.

This document aims to be sufficiently broad that someone who has not paid any attention to the space all year could read it (and the linked documents) and be as well-informed to make donation decisions as they could reasonably be without personally interviewing researchers and organisations.

I’d like to apologize in advance to everyone doing useful AI Safety work whose contributions I have overlooked or misconstrued. As ever I am painfully aware of the various corners I have had to cut due to time constraints from my job, as well as being distracted by 1) other projects, 2) the miracle of life and 3) computer games.

This article focuses on AI risk work. If you think other causes are important too, your priorities might differ. This particularly affects GCRI, FHI and CSER, who both do a lot of work on other issues which I attempt to cover but only very cursorily.

How to read this document

This document is fairly extensive, and some parts (particularly the methodology section) are largely the same as last year, so I don’t recommend reading from start to finish. Instead, I recommend navigating to the sections of most interest to you. You should also read the Conflict of Interest Section.

If you are interested in a specific research organisation, you can use the table of contents to navigate to the appropriate section. You might then also want to Ctrl+F for the organisation acronym in case they are mentioned elsewhere as well. Papers listed as ‘X researchers contributed to the following research lead by other organisations’ are included in the section corresponding to their first author and you can Cntrl+F to find them.

If you are interested in a specific topic, I have added a tag to each paper, so you can Ctrl+F for a tag to find associated work. The tags were chosen somewhat informally so you might want to search more than one, especially as a piece might seem to fit in multiple categories.

Here are the un-scientifically-chosen hashtags:

AgentFoundations

Amplification

Capabilities

Community

Ethics

Fiction

Forecasting

GameTheory

Interpretability

IRL

NearTerm

Obstruction

Other

Overview

Policy

Robustness

Scenarios

ShortTerm

Strategy

Translation

ValueLearning

You might also be interested in the ‘Organisation Second Preference’ section, which is new this year.

New to Artificial Intelligence as an existential risk?

If you are new to the idea of General Artificial Intelligence as presenting a major risk to the survival of human value, I recommend this Vox piece by Kelsey Piper, or for a more technical version this by Richard Ngo.

If you are already convinced and are interested in contributing technically, I recommend this piece by Jacob Steinhardt, as unlike this document Jacob covers pre-2019 research and organises by topic, not organisation, or this from Hendrycks et al., or this from Critch & Krueger, or this from Everitt et al, though it is a few years old now.

Conflict of Interest

In the past I have had very demandin...

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 don't governments seem to mind that companies are explicitly trying to make AGIs?, published by Ozzie Gooen on December 23, 2021 on The Effective Altruism Forum.

Epistemic Status: Quickly written, uncertain. I'm fairly sure there's very little in terms of the public or government concerned about AGI claims, but I'm sure there's a lot I'm missing.

This was originally posted to Facebook here, where it had some discussion. Many thanks to all of the people who engaged in the discussion there.

Multiple tech companies now are openly claiming to be working on developing AGI (Artificial General Intelligence).

As written in a lot of work on AGI (See Superintelligence, as an example), if any firm does establish sufficient dominance in AGI, they might have some really powerful capabilities.

Write bots that could convince (some) people to do almost anything

Hack into government weapons systems

Dominate vital parts of the economy

Find ways to interrupt other efforts to make AGI

And yet, from what I can tell, almost no one seems to really mind? Governments, in particular, seem really chill with it. Companies working on AGI get treated similarly to other exciting AI companies.

If some company were to make a claim like,

"We're building advanced capabilities that can hack and modify any computer on the planet"

or,

"We're building a private nuclear arsenal",

I'd expect that to draw attention.

But with AGI, crickets.

I assume that governments dismiss corporate claims of AGI development as overconfident marketing-speak or something.

You might think,

"But concerns about AGI are really remote and niche. State actors wouldn't have come across them."

That argument probably applied 10 years ago. But at this point, the conversation has spread a whole lot. Superintelligence was released in 2014 and was an NYT bestseller. There are hundreds of books out now about concerns about increasing AI capabilities. Elon Musk and Bill Gates both talked about it publicly. This should be one of the easiest social issues at this point for someone technically savvy to find.

The risks and dangers (of a large power-grab, not of alignment failures, though those too) are really straightforward and have been public for a long time.

Responses

In the comments to my post, a few points were made, some of which I was roughly expecting. Points include:1. Companies saying they are making AGI are ridiculously overconfident2. Governments are dramatically incompetent3. AGI will roll out gradually and not give one company a dominant advantageMy quick responses would be:1. I think many longtermist effective altruists believe these companies might have a legitimate chance in the next 10 to 50 years, in large part because of a lot of significant research (see everything on AI and forecasting on LessWrong and the EA Forum). At the same time, my impression is that most of the rest of the world is indeed incredibly skeptical of serious AGI transformation.2. I think this is true to an extent. My impression is that government nonattention can change dramatically and quickly, particularly in the United States, so if this is the crux, it might be a temporary situation. 3. I think there's substantial uncertainty here.

But I would be very hesitant to put over a 70% chance that: (a) one, or a few, of these companies will gain a serious advantage, and (b) the general-purpose capabilities of these companies will come with significant global power capabilities. AGI is general-purpose, it seems difficult to be sure that your company can make it without it being an international security issue of some sort or other.

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: Worst-case thinking in AI alignment, published by Buck Shlegeris on December 23, 2021 on The AI Alignment Forum.

Alternative title: “When should you assume that what could go wrong, will go wrong?”

Thanks to Mary Phuong and Ryan Greenblatt for helpful suggestions and discussion, and Akash Wasil for some edits.

In discussions of AI safety, people often propose the assumption that something goes as badly as possible. Eliezer Yudkowsky in particular has argued for the importance of security mindset when thinking about AI alignment.

I think there are several distinct reasons that this might be the right assumption to make in a particular situation. But I think people often conflate these reasons, and I think that this causes confusion and mistaken thinking. So I want to spell out some distinctions.

Throughout this post, I give a bunch of specific arguments about AI alignment, including one argument that I think I was personally getting wrong until I noticed my mistake yesterday (which was my impetus for thinking about this topic more and then writing this post). I think I’m probably still thinking about some of my object level examples wrong, and hope that if so, commenters will point out my mistakes. But I think I’ll stand by the claim that we should be attempting to distinguish between these classes of argument.

My list of reasons to maybe use worst-case thinking

Here’s an attempt at describing some different classes situations where you might want to argue that something goes as badly as it could.

You’re being optimized against

For example, if you’ve built an unaligned AI and you have a team of ten smart humans looking for hidden gotchas in its proposed actions, then the unaligned AI will probably come up with a way of doing something bad that the humans miss. In AI alignment, we most often think about cases where the AI we’re training is optimizing against us, but sometimes we also need to think about cases where other AIs or other humans are optimizing against us or our AIs.

In situations like this, I think Eliezer’s attitude is basically right: we’re being optimized against and so we have to use worst-case thinking and search hard for systems which we can strongly argue are infallible.

One minor disagreement: I’m less into hard takeoffs than he is, so I place less weight than he does on situations where your AI becomes superintelligent enough during training that it can exploit some kind of novel physics to jump an airgap or whatever. (Under my model, such a model probably just waits until it’s deployed to the internet–which is one of the first things that AGI developers want to do with it, because that’s how you make money with a powerful AI–and then kills everyone.)

But I fundamentally agree with his rejection of arguments of the form “only a small part of the space of possible AI actions would be devastatingly bad, so things will probably be fine”.

Scott Garrabrant writes about an argument like this here.

The space you’re selecting over happens to mostly contain bad things

When Hubinger et al argue in section 4.4 of Risks from Learned Optimization that “there are more paths to deceptive alignment than to robust alignment,” they aren’t saying that you get a misaligned mesa-optimizer because the base optimizer is trying to produce an agent that is as misaligned as possible, they’re saying that even though the base optimizer isn’t trying to find a misaligned policy, most policies that it can find are misaligned and so you’ll probably get one. But unlike the previous situation, if instead it was the case that 50% of the policies that SGD might find were aligned, then we’d have a 50% chance of surviving, because SGD isn’t optimizing against us.

I think that AI alignment researchers often conflate these two classes of arguments. IMO, when you’re training an AGI:

The 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: Stress - effective ways to reduce it, published by Ben Williamson on December 21, 2021 on The Effective Altruism Forum.

Just in time for the Christmas holidays, often a particularly stressful time of the year, are these practical recommendations for reducing the impact of stress on your wellbeing.

Top recommendation: Training in Cognitive Behavioural Therapy (CBT) had twice the effect of the next most effective recommendations, Sudarshan Kriya Yoga and diaphragmatic breathing, plausibly halving markers of stress after 8 weeks of practice.

Personal recommendation: Diaphragmatic breathing is exceptionally easy to learn, can be beneficially practised at almost any time or location, and produces a large reduction in biological markers of stress.

Introduction

Stress is a leading cause of reduced mental health and wellbeing globally. This article reviews the evidence for 15 main interventions addressing the causes of stress and minimising its impact, both in the moment and long-term. We found that Cognitive Behaviour Therapy (CBT), Sudarshan Kriya Yoga (SKY), diaphragmatic breathing, and Mindfulness-Based Stress Reduction (MBSR) can all reduce indicators of stress by more than 25%.

This research is part of a series looking at the most effective ways individuals can improve their wellbeing. Thanks to the EA Infrastructure Fund for funding this project and to the many people who have reached out to discuss ideas and offer feedback on this and previous posts.

Important note: None of the following constitutes professional medical advice. While stress is often a natural part of the human experience, it can also be a symptom of an underlying mental health issue that should take precedent and may require professional support.

Key recommendations

Efforts to tackle stress roughly target one of three areas: its underlying causes; short-term relief, or long-term resilience. Below are two recommendations for better managing stress across each of these three areas, highlighting the interventions with the greatest effect on stress in the literature we reviewed.

Underlying causes

Cognitive Behavioural Therapy (CBT): take a course in CBT, either online or through an app.

Reducing key stressors: money, work, and relationships are the three most common causes of stress in the US. Building a financial runway, task management systems, and adopting better strategies for managing difficulties in relationships could all be highly impactful depending on your personal circumstances.

Short-term relief

Diaphragmatic breathing: practice slow, diaphragmatic breathing at regular intervals and at the onset of stressful situations.

Listen to music: put away any distractions and listen to some music to relax, either classical or just songs you particularly enjoy.

Long-term resilience

Mindfulness: try a mindfulness-based stress reduction programme (MBSR), either online or through an app.

Rhythmic breathing: learn a rhythmic breathing practice, such as Sudarshan Kriya Yoga (SKY), and do this regularly (e.g. twice a week).

A brief(-ish) note on methods

This report is the second in a series researching the most effective ways individuals can improve their own wellbeing. The methods and presentation used are a work in progress, aiming to find a balance between maximising the quality of recommendations, the likelihood people will adopt recommendations made, and the number of topics we can provide recommendations on.

The interventions in this article have been ranked here by their size of effect vs. a control condition in the literature we reviewed. We encourage anyone interested to view our evidence table for a more in-depth understanding of how we have produced the recommendations in this article. Effect size scores are presented with each intervention listed in the main discussion section. All references and ...

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: Transformer Circuits, published by Evan Hubinger on December 22, 2021 on The AI Alignment Forum.

Chris Olah, Neel Nanda, Catherine Olsson, Nelson Elhage, and a bunch of other people at Anthropic just published “Transformer Circuits,” an application of the Circuits-style interpretability paradigm to transformer-based language models. From their very top level summary:

Can we reverse engineer transformer language models into human-understandable computer programs? Inspired by the Distill Circuits Thread, we're going to try.

They've chosen to publish their work in an interestingly novel format, publishing their first paper, “A Mathematical Framework for Transformer Circuits,” alongside a set of YouTube videos that go into even more detail on their findings. I watched the full YouTube playlist, and I found it absolutely fascinating and would highly recommend it as a way to engage with this research.

Some of my high-level takeaways:

Their signature finding, I think, is that of induction heads. I'll let you get the explanation of what they are directly from Chris's video, but essentially they're a very basic mechanism that all transformer-based language models seem to use that drives their ability to “meta-learn”—that is, improve their accuracy over the course of seeing a larger context. Similarly to the exploration of early vision in Circuits, induction heads give us insight into the basic building blocks of large language models. While I think the authors' did a great job uncovering this basic building block, I think their explanation of meta-learning as primarily being driven by induction heads mostly leads us to a big open research question, which is how exactly induction heads can be put together to produce the more complex phenomenon we see in large language models.[1]

Probably their most fascinating finding, in my opinion, was their discovery of the “induction bump.” I won't try to reexplain exactly what the induction bump is, since I think Catherine's video does such a good job of that, but I will say that, to my knowledge, the authors' exploration of the induction bump is the first time there's ever been a detailed, circuits-level analysis of a phase-change that occurs over the course of training in a discrete way. I think this is especially interesting because a lot of stories that people like to tell about how particular safety problems might arise in neural networks often rely on these sorts of phase-change-style transitions (e.g. the development of deception).

It seems to me that the existence of something like the induction bump is suggestive that there might be more of these sorts of phase changes hiding in future large training runs, or even in existing large training runs—as Catherine notes, the bump is pretty easy to miss until you break down the loss into individual tokens.[2]

Overall, I think this is clearly the most exciting progress in transparency and interpretability in general since Circuits—and I'm really happy to see it happening in language models, as I've previously emphasized I think is important for us to focus on. One thing that I think really sets this sort of transparency and interpretability work apart from the rest is the authors' emphasis on understanding the mechanistic building blocks underlying their models—with the hope of eventually being able to reverse-engineer them—rather than just, for example, trying to give humans tools to predict what models are doing (without the output of those tools necessarily having any correspondence to what the models are actually doing).

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: Random facts can come back to bite you, published by tailcalled on December 22, 2021 on LessWrong.

Consider the post Ads Don't Work That Way by Kevin Simler on Melting Asphalt. It contrasts a couple of different mechanisms by which advertising might work, including:

Emotional inception - manipulating your feelings towards a product via conditioning

Raising awareness - making you aware of a product or its features

Cultural imprinting - creating a societal association for a product, to allow people to use the product to associate themselves with this same association

The details are described in the post, and I'm sure that Kevin Simler describes the theory much better than I could. However, they are not very important, as I am not bringing up this post to make a point about advertising. Instead, I want to zoom in on an argument Kevin makes:

UNDERHANDED ADVERTISING

Consider this one for Corona:

Whatever's going on here, it's not about awareness, persuasion, promises, or honest signaling. In fact this image is almost completely devoid of information in the most literal sense. As Steven Pinker defines it, information is "a correlation between two things that is produced by a lawful process (as opposed to coming about by sheer chance)." In this case, the image is so arbitrary that it can't be conveying any information about Corona per se, as distinct from any other beer. Corona wasn't specifically designed for the beach, nor does 'beach-worthiness' emerge from any distinguishing features of Corona. You could swap in a Budweiser or Heineken and no "information" would be lost.

This seems obviously true. After all, what inherent connection could there possibly be between Corona and a sunny beach? In a footnote in the post, Kevin admits that there is a slight cultural/circumstantial/contextual factor, but that's not enough to change the core point of the post; the Corona/beach connection is obviously fairly arbitrary.

Except... now, years later, I've watched this YouTube video on the channel called "Food Theory":

In it, MatPat explains that beers can get "skunked" when hit by sunlight. However, it turns out that Corona beer is uniquely designed such that skunking improves its taste, unlike other beers which it worsens. Therefore, there actually is a genuine non-arbitrary connection between the Corona brand and drinking beer on the beach.

This was a completely random fact that I hadn't expected to jump out. I guess reality has a surprising amount of detail. But it makes the development and description of these broad social science theories annoying and difficult.

Obviously Kevin Simler didn't know about Corona's skunking properties when writing his post about advertising, and I didn't know about them when reading it. If his readers had said "This example is unconvincing, because you haven't shown there to be no special relationship between Corona and the beach.", there's not much he could have done; there could be any number of connections, and it's hard to know if you've enumerated all the important ones. So it would be really hard to disprove the objection. But at the same time, as we can see now, the objection would have been right.

One solution to this would be to become an expert on everything you talk about, so you know even the most obscure potentially relevant facts. But that's not practical. Another solution would be to accept being wrong - but I often get the impression that these sorts of things lead to people being wrong a lot. A third solution is to talk about narrower things, so you can specialize more into getting them right. Alternatively, maybe we could keep making sweeping generalizations, but only from things we have a lot of familiarity with? I don't know what the best solution is, and it's probably highly context-dependent, but it seems like a problem to co...

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: [Book rec] The War with the Newts as “EA fiction”, published by Lizka on December 22, 2021 on The Effective Altruism Forum.

Tl;dr: Karel Čapek’s 1936 science fiction satire The War with the Newts is brilliant and relevant to various topics in EA, and I’d like more people to hear about it (and to consider reading it). It deals in particular with non-human sentience and welfare, misguided or signal-oriented altruism, and x-risk (among other topics). It’s also just really fun to read, if you’re looking for slightly obscure, slightly experimental science fiction. (Please note that it has a fair bit of violence.) You can find the book here. (I read and endorse Ewald Osers’ translation, but it’s possible others are good, too.)[1]

A note about this post: this is a quick summary and a heartfelt recommendation to give this book a try, and that’s it. Here’s a NYT review (paywalled) in case you’d prefer to read that. There are no real startling revelations in this post, and I don’t think the book will revolutionize your thinking about impact or the long-term future; I simply think it’s great and am jotting down why.

Summary of the post

Historical context and a brief intro + summary of the novel

EA topics.

Non-human sentience and welfare

Misguided or self-congratulatory (rather than goal-oriented) altruism

X-risk

Weirdly AGI-like behavior of the Newts

Concluding thoughts and meta comments

Historical context and a brief intro + summary of the novel

Karel Čapek was a Czech journalist, writer, and early anti-fascist, and is best known for producing the first piece of literature ever to use the word “robot” in his 1920 play, “R.U.R.” The play is a cool (and short) read— especially as a piece of science fiction history— but I prefer the novel.

The War with the Newts is in a whole different class. It was written in Czechoslovakia in 1936, and it responds directly to its scientific and political background. I tend to dislike it when science fiction is judged on “prophetic ability,” as this often leads to survivorship-bias- and overfitting-afflicted discussions that overlook the actual strengths of literature, but it’s impossible to avoid pointing out the eerily insightful portrayal of Europe’s development in the book. (An optional elaboration is in the footnote.[2]) The main strengths of the book have little to do with Čapek’s ability to predict the pathways of WWII, however.

The novel starts with the discovery of a new species of intelligent mild-mannered Newt. The human relationship with the Newts is initially pretty benign—the Newts help human endeavors by fishing pearls from water, and in exchange, humans give them knives to defend themselves from sharks— but the situation soon becomes industrialized and quickly turns sour. The rest of the book explores what follows. Čapek illustrates mental shenanigans humans use to justify treating beings as non-sentient. He creates fictional op-eds and conference notes. Readers learn the strange experience of living in a world that seems to be utterly safe but which moves extremely quickly under the surface. All of this is conveyed through a patchwork of perspectives (and a play of medium) that is great for revealing a bunch of real cognitive biases (satire at its best).

A spoiler-y summary is in this footnote.[3]

Non-human sentience and welfare

The novel painstakingly sets up a core fact of the fictional reality; the Newts are sentient to some vaguely undetermined and fluid degree. It contrasts this truth with descriptions of casual exploitation of the creatures (and the corresponding euphemistic scientific jargon and pseudo-objective journalistic coverage). All of this highlights inconsistencies between the treatment and interpretation of the Newts— and by extension, of real-life non-human beings.[4]

A relevant passage:

“The death-rate during ...

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: Six Specializations Makes You World-Class, published by lsusr on December 22, 2021 on LessWrong.

If you want an average, successful life, it doesn't take much planning. Just stay out of trouble, go to school, and apply for jobs you might like. But if you want something extraordinary, you have two rather: 1) Become the best at one specific thing. 2) become very good (top 25%) at two or more things.

―Scott Adams

You can make a lot of money by being the best at a scalable domain because leverage feeds winner-take-all dynamics.

The problem is there are P=8×109 people on Earth. To be the best, you have to be better than every single one of them. It's not hard to get into the top 0.1 or even 0.01 of people at many skills but it is hard to get into the top 1P=.0000000001.

There are many different skills you can learn e.g. mathematics, sales, drawing, management. Suppose there are S=100 basic skills. You can become the best in the world by being the best at any one of them. But to be the best at one of S=100 different skills you must compete against PS=80,000,000 people. Competing against eighty million people is easier than competing against eight billion but it is still hard.

Many real-world competitions require more than one skill. Suppose you want to be the best author of physics books. You must combine two skills: writing and physics. There are S(S−1)2=100(100−1)2≈5,000 different pairs of skills. Instead of competing against the rest of mankind in 100 domains you are competing in 5,000 different domains. On average, you will compete against 2PS(S−1)≈2PS2=1,600,000 people.

For n randomly-selected skills there will, on average, be only Pn!Sn people competing at the intersection. You are one of those people. At what n is there zero competition (besides yourself)?

Pn!Sn=1P=Snn!lnP=nlnS−ln(n!)≈nlnS−(nlnn−n)≈n(1+lnS−lnn)

lnn≈2 when 5≤n≤10.

lnP≈n(1+lnS−2)≈n(lnS−1)n≈lnPlnS−1≈ln(8×109)ln100−1≈6

Six. If S=100 then combining six skills at random makes you so specialized that literally nobody in the world is competing with you. The equation is weakly dependent on our choice of S. If S=50 then you need seven skills instead.

Six arbitrary skills may not be, in general, synergistic. But if your work combines six or more skills then you should not be surprised if you are the best in the world.

I am not a great artist, and I have never taken a traditional class in writing. I’m not the funniest person in my social circle, and I’m not a great business mind either. But few people in the world have a complete talent stack as valuable as mine. By being good enough in each of those individual talents, I can be a famous syndicated cartoonist and enjoy an ideal career.

―Win Bigly by Scott Adams

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: What’s Up With the CDC Nowcast?, published by Zvi on December 22, 2021 on LessWrong.

How’s it going?

The CDC nowcast last week was 2.7% Omicron. That seemed like a reasonable guess.

The CDC nowcast this week is 73% Omicron, and last week’s nowcast got revised from 2.7% to 12.6%.

That’s two retroactive extra doublings last week, and then four more in the following seven days relative to Delta, for a doubling time of less than two days.

That report came out right after I put out Omicron Post #8, and I quickly noticed I was confused.

One possibility is, hey, large error bars, so I guess there’s that?

This conversation provided potential factors, but did not clear up the confusion.

This provides a clear alternate hypothesis, and I trust the source quite a bit.

Censoring the past ten days is quite a high price to pay given how fast things are moving. That’s a lot of data to give up, and it’s worth noting that 10 days later the difference between 40% and 73% isn’t all that big in context. But if Trevor is right about the speed of submission, and the Nowcast isn’t adjusting, it’s going to give out a nonsense answer.

Which is exactly what it is giving. There are three huge problems with the nowcast’s answer, on top of the revision being rather large and not inspiring confidence.

Where are all the cases?

The regional numbers make even less sense.

Also, it’s averaging over a week so it’s implying even higher rates now.

The third problem I only noticed later, but if it’s 73% for the whole week, and 12% for last week, where did we enter the week, and therefore how high did we have to get to balance out the first few days? Not that this problem is necessary to notice the flaws.

The first issue jumped out at me right away. We know that there wasn’t a dramatic rise in the number of cases overall. We know there wasn’t a rise in the positive test percentage. Yet the claim is there was suddenly, over the whole week, three times as many Omicron cases as Delta. Does that mean that the number of Delta cases was down by more than half inside of a week? Does that seem remotely possible?

Testing capacity is a limiting factor, but if it was having a big effect, we’d presumably be seeing a much bigger jump in the positive test rate. I can imagine a world where that’s not true, but it doesn’t match the data from earlier in the pandemic.

This gets far more extreme if you go to the regional level, and the implications get bonkers.

This was a good visualization of the nowcast by region.

This means that in two regions covering ten states, we had more than 95% Omicron cases, so twenty Omicron cases for every Delta case, whereas a week ago Omicron was a clear minority of cases. Does that possibly live in the same world as our case counts?

This is another good explanation of some of the reasons the data doesn’t add up. This is how one should approach the situation when algorithms produce obvious nonsense. As he notes, this isn’t a knock on the CDC. I can be harsh on the CDC, but this isn’t the time and place for that. It is an example of them doing their best to be helpful, and failing to notice that their algorithm had produced nonsense because they didn’t have a human look at it.

And sure, given how many eyes were on the forecast they should have had a human do a sanity check and put in a warning note. But they also shouldn’t have had to. Everyone else should have also noticed they were confused and that the number didn’t make sense, rather than reporting an obvious nonsense projection with huge error bars as if it were a fact. If I were at the CDC, I’d fix the ‘no human sanity check’ issue but also would be muttering about how this is why we can’t have nice things.

It would be better to not give obvious nonsense as an output, but that’s a lot to ask here. It’s really really hard to create an algor...

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: World's First Octopus Farm - Linkpost, published by CatGoddess on December 22, 2021 on The Effective Altruism Forum.

First post on the forum so be gentle.

TL;DR: The Spanish company Nueva Pescanova is planning to farm octopus and begin selling it on the market in 2023. They intend to sell 3,000 tons of farmed octopus per year, which amounts to several hundred thousand octopuses. This is concerning given that octopuses are highly intelligent animals; there are no released standards for how the octopuses are going to be kept and raised, nor for how they will be slaughtered. There doesn't seem to be much in the way of coordinated action against this yet.

Link to the article:

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