The Nonlinear Library

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 Germany's Mid-Year Update 2023, published by Sarah Tegeler on September 19, 2023 on The Effective Altruism Forum.

This post outlines our current status regarding the planned focus areas in our strategy for this year.

Background

EA Germany (EAD)

In the 2022 EA survey, 9.9% of all participants were from Germany, up from 7.4% in 2020, remaining the third-largest community behind the US and the UK. We are a membership association with 108 members, six board members, five employed team members (four FTEs), and ten contractors and volunteers (for EAGxBerlin 2023, intro program and newsletter).

In 2023, the association gained 35 new members and two new employees.

Local Groups

There are currently 27 active local groups in Germany, some of which are university-based, but most refer to themselves as local groups.

Group size ranges from 5-50 active members. Overall, there are ~50-70 community builders and a total of >300 people regularly involved in the groups.

Impact Estimation

We have gathered data points about the outcomes of our programmes, which we will share in the following sections. Since, however, we are uncertain about the interpretation of this data, we cannot be sure about the overall impact of our programs.

In the future, we will focus on finding better evaluation criteria in order to estimate our impact.

Focus Areas

Foundational Programs

Foundational programs are either the continuation of existing programs or those that seem broadly beneficial to growing a sustainable community. We have established OKRs for each program and are reviewing them monthly.

Communications

We have been running and are regularly updating effektiveraltruismus.de, the German website about effective altruism, since Effektiv Spenden handed it over to us in late 2022. They also handed us the newsletter, which enabled us to send out our existing monthly German-language newsletter to more than 4,000 subscribers compared to the 350 we had before.

We are in regular exchange with Open Philanthropy grantees to coordinate the translation of articles from English into German. They have published some of their content on our website, and we have promoted their podcast with narrations of articles.

Additionally, we have been helping to coordinate the publication of EA-relevant books in German, including the German launch of What We Owe The Future on 30th August 2023.

EAGx Conferences

We have applied for and received funding to run EAGxBerlin on September 8-10, 2023 and hired a team of six people in order to do so.

Additionally, we have organised meetups for German EAs at 4 EAG(x) conferences with ~5-50 attendees each.

Intro Program [formerly Intro Fellowship]

The Intro Program, which used to be called "EA Intro Fellowship", was held in the winter of 22/23 and summer of '23. During the last round, we received a peak of more than 100 applications. Around 60 % of the participants completed the program successfully and more than 90 % of program participants described at least one relevant event outcome, e.g. making an important professional connection, discovering a new opportunity, or getting a deeper understanding of a key idea.

Career 1-1s

We had ~160 calls and meetings related to career paths and decisions between January and June: ~60 at conferences, ~30 at retreats. The others were career calls or office hours. Recommendations came through our programs, 80,000 hours, and the form on our website.

Community Health

We appointed our team member Milena Canzler as the German Community Health contact person, listing her contact details on our website while also including a German and English contact form. In several cases, we have already been able to provide support. Additionally, we provide materials and training for awareness teams at EA-related events in order to avoid negative experiences for and har...

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: Ticking Clock: The Rapid Rise of Farmed Animals in Africa, published by AnimalAdvocacyAfrica on September 19, 2023 on The Effective Altruism Forum.

As humanity continues its era of rapid population growth and rising economic prosperity, the demand for animal protein is anticipated to reach unparalleled heights. This surge in consumption is set to drastically impact the lives of farmed animals worldwide. Nowhere is this growth more pronounced than in Africa.

The evidence

Previously, our anticipation of Africa's sharp increase in livestock numbers was primarily grounded in the historical global expansion of farmed animal populations over the past decades, coupled with human population growth trends across the African continent. This post, however, delves into the specific projections of farmed animal numbers and animal farming intensification from 2012 to 2050, as outlined by the Food and Agriculture Organization of the United Nations (FAO), which is based on many more factors than just historical changes in animal agriculture and human population growth, providing us with more detailed estimates than previously available. If not stated otherwise, all figures named in this blog post refer to these projections, which can be found here.

The FAO's projections encompass a range of future scenarios, including both a "business as usual" model and a "towards sustainability" model. For the scope of this post, we focus on the "business as usual" projection, which assumes the absence of any significant efforts to reduce the extent of factory farming. This allows us to explore the potential ramifications of the current trajectory of animal agriculture. Notably, even upon considering the projections from the "towards sustainability" model, the prospects for Africa remain largely unaltered. In contrast, for all other continents, we can observe a noticeable reduction in farmed animal numbers in comparison to the "business as usual" model. Although far from certain, this may indicate that the FAO assumes that growth in animal agriculture in Africa is close to unavoidable or that nothing will be done to hinder its growth.

The FAO provides data for farmed land animals, including cattle, pigs, sheep, goats, poultry, and buffaloes, but omits figures for other mammals, such as rabbits, horses, and dogs, as well as insects, fish, and seafood. This exclusion is noteworthy since the protein yield from smaller animals demands significantly greater numbers of individual animals per kilogram in contrast to large mammals like cows and pigs. Thus, while the analysis below offers crucial insights, it only represents a fraction of future developments.

Africa in global comparison

According to the FAO's projections, the number of farmed land animals in Africa is anticipated to experience a remarkable surge in the coming decades. As shown in the graph below, the population of farmed land animals for the entire continent of Africa will rise from around 2.6 billion in 2012 to around 9.4 billion in 2050, an increase of 262%. Consequently, Africa would surpass all other global regions in terms of the total size of farmed animal populations by 2050, except for Asia, and reach roughly twice the total number of animals of other regions like Europe or North America.

At present, Asia leads and will continue to lead all continents in terms of the total number of farmed land animals. This can largely be attributed to factory farming in China, Indonesia, and India. Nevertheless, Africa's livestock numbers are expected to increase by a much larger absolute number and at a higher rate than Asia's projected 26% rise from 2012 to 2050.

Note that these numbers (along with all following numbers in this post) refer to the count of live animals at a given time in a year, which is not to be confused but should be highly c...

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: Technical AI Safety Research Landscape [Slides], published by Magdalena Wache on September 18, 2023 on LessWrong.

I recently gave a technical AI safety research overview talk at EAGx Berlin. Many people told me they found the talk insightful, so I'm sharing the slides here as well. I edited them for clarity and conciseness, and added explanations.

Outline

This presentation contains

An overview of different research directions

Concrete examples for research in each category

Disagreements in the field

Intro

Overview

I'll start with an overview of different categories of technical AI safety research.

The first category of research is what I would just call alignment, which is about making AIs robustly do what we want. Then there are various "meta" research directions such as automating alignment, governance, evaluations, threat modeling and deconfusion. And there is interpretability. Interpretability is probably not enough to build safe AI on its own, but it's really helpful/probably necessary for various alignment proposals. Interpretability also helps with deconfusion.

I'm using clouds because the distinction between the categories often isn't very clear.

Let's take a closer look at the first cloud.

What exactly do I mean by alignment? What do we align with what?

In general, we want to make AIs do what we want, so we want to align "what we want" with "what the AI does". That's why it's called alignment.

We can split this up into intent alignment (make the AI want what we want) and capability robustness (make it able to robustly do what it wants).

And we can split intent alignment up into outer alignment (find a function that captures what we want) and inner alignment (ensure that what the AI ends up wanting is the same as what's specified in the function that we trained it on).

There are a few ways in which this slide is simplified: The outer/inner alignment split is not necessarily the right frame to look at things. Maybe "what the AI wants" isn't even a meaningful concept. And many approaches don't really fit into these categories. Also, this frame looks at making one AI do what we want, but we may end up in a multipolar scenario with many AIs.

Concrete Technical Research

In this section I'll give some examples to give you a flavor of what kinds of research exists in this space. There is of course a lot more research.

Let's start with outer alignment.

Outer alignment is the problem of finding a mathematical function which robustly captures what we want. The difficulty here is specification gaming.

In this experiment the virtual robot learned to turn the red lego block upside down instead of the intended outcome of stacking it on top of the blue block. This might not seem like a big problem - the AI did what we told it to do. We just need to find a better specification and then it does what we want. But this toy example is indicative of a real and important problem. It is extremely hard to capture everything that we want in a specification. And if the specification is missing something, then the AI will do what is specified rather than what we meant to specify.

A well-known technique in reward specification is called Reinforcement Learning from Human Feedback (RLHF). In the Deep reinforcement learning from human preferences paper they were able to make a virtual leg perform a backflip, despite "backflip" being very hard to specify mathematically.

(Links: blogpost, paper)

Let's continue with inner alignment.

Inner alignment is about making sure that the AI actually ends up wanting the thing which it is trained on. The failure mode here is goal misgeneralization:

(Links: forum post, paper)

One way to train in more diverse environments is adversarial training:

(Links: paper, takeaways post, deceptive alignment)

As I mentioned above, for many approaches it doesn't really...

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: Microdooms averted by working on AI Safety, published by nikola on September 18, 2023 on LessWrong.

Disclaimer: the models presented are extremely naive and simple, and assume that existential risk from AI is higher than 20%. Play around with the models using this (mostly GPT-4 generated) jupyter notebook.

1 microdoom = 1/1,000,000 probability of existential risk

Diminishing returns model

The model has the following assumptions:

Absolute Risk Reduction: There exists an absolute decrease in existential risk that could be achieved if the AI safety workforce were at an "ideal size." This absolute risk reduction is a parameter in the model.

Note that this is absolute reduction, not relative reduction. So, a 10% absolute reduction means going from 20% x-risk to 10% x-risk, or from 70% x-risk to 60% x-risk.

Current and Ideal Workforce Size: The model also takes into account the current size of the workforce and an "ideal" size (some size that would lead to a much higher decrease in existential risk than the current size), which is larger than the current size. These are both parameters in the model.

Diminishing Returns: The model assumes diminishing returns on adding more people to the AI safety effort. Specifically, the returns are modeled to increase logarithmically with the size of the workforce.

The goal is to estimate the expected decrease in existential risk that would result from adding one more person to the current AI safety workforce. By inputting the current size of the workforce, the ideal size, and the potential absolute risk reduction, the model gives the expected decrease.

If we run this with:

Current size = 350

Ideal size = 100,000

Absolute decrease (between 0 and ideal size) = 20%

we get that one additional career averts 49 microdooms. Because of diminishing returns, the impact from an additional career is very sensitive to how big the workforce currently is.

Pareto distribution model

We assume that the impact of professionals in the field follows a Pareto distribution, where 10% of the people account for 90% of the impact.

Model Parameters

Workforce Size: The total number of people currently working in AI safety.

Total Risk Reduction: The absolute decrease in existential risk that the AI safety workforce is currently achieving.

If we run this with:

Current size = 350

Absolute risk reduction (from current size) = 10%

We get that, if you're a typical current AIS professional (between 10th and 90th percentile), you reduce somewhere between 10 and 270 microdooms. Because of how skewed the distribution is, the mean is at 286 microdooms, which is higher than the 90th percentile.

A 10th percentile AI Safety professional reduces x-risk by 14 microdooms

A 20th percentile AI Safety professional reduces x-risk by 16 microdooms

A 30th percentile AI Safety professional reduces x-risk by 20 microdooms

A 40th percentile AI Safety professional reduces x-risk by 24 microdooms

A 50th percentile AI Safety professional reduces x-risk by 31 microdooms

A 60th percentile AI Safety professional reduces x-risk by 41 microdooms

A 70th percentile AI Safety professional reduces x-risk by 61 microdooms

A 80th percentile AI Safety professional reduces x-risk by 106 microdooms

A 90th percentile AI Safety professional reduces x-risk by 269 microdooms

Linear growth model

If we just assume that going from 350 current people to 10,000 people would decrease x-risk by 10% linearly, we get that one additional career averts 10 microdooms.

One microdoom is A Lot Of Impact

Every model points at the conclusion that one additional AI safety professional decreases existential risks from AI by one microdoom at the very least.

Because there are 8 billion people alive today, averting one microdoom roughly corresponds to saving 8 thousand current human lives (especially under short timelines, where the...

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: Show LW: Get a phone call if prediction markets predict nuclear war, published by Lorenzo on September 18, 2023 on LessWrong.

hasrussialaunchednukes.com used to be a website "to help [the author] and a few of [their] friends get notified when Russia launches any nukes towards Ukraine, which would be a sign for many of us to leave the big cities we live in, and chill in the countryside for a few weeks." It used prediction markets for 2022, so it stopped working in 2023.

Some people asked me to make a version for 2023, so here it is at.

I manually curate the list of phone numbers, a bunch of which were sent to me by one of the people requesting this, you can ask for your number to be added using any of the links at the bottom of the website.

You can look at the incredibly hacky script I use to make phone calls here: it runs every 10 minutes on a Hetzner VPS I use for a bunch of things

I don't think this is extremely reliable, I give an ~80% chance that it will work, and I'll do my best to prevent false alarms, but some people found it useful, and I was asked to share this here.

I would be curious to hear if you have any thoughts!

Why do it in mid-september, given that the year is almost over? I actually implemented this in February, but procrastinated until today to write this post. But it will be trivial to update it for 2024, once the year is over.

I use vonage instead of twilio to handle phone calls because someone else was implementing a similar thing with twilio, and we wanted to have redundancy

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: The Talk: a brief explanation of sexual dimorphism, published by Malmesbury on September 18, 2023 on LessWrong.

Cross-posted from substack.

"Everything in the world is about sex, except sex. Sex is about clonal interference."- Oscar Wilde (kind of)

As we all know, sexual reproduction is not about reproduction.

Reproduction is easy. If your goal is to fill the world with copies of your genes, all you need is a good DNA-polymerase to duplicate your genome, and then to divide into two copies of yourself. Asexual reproduction is just better in every way:

SexualAsexualBuild costly DNA-manipulating machinery that chops the DNA into pieces to produce gametesJust copy yourself broScout the perilous wild for a mate, and perform a complicated ceremony so your gametes fuse with each otherJust copy yourself broOnly pass 50% of your genome on to the next generationJust copy yourself broDifferentiate into two types, making it twice as hard to find a matching gameteJust copy yourself broMake all kinds of nonsense ornaments to satisfy the other sex's weird instinctsJust copy yourself bro

It's pretty clear that, on a direct one-v-one cage match, an asexual organism would have much better fitness than a similarly-shaped sexual organism. And yet, all the macroscopic species, including ourselves, do it. What gives?

Here is the secret: yes, sex is indeed bad for reproduction. It does not improve an individual's reproductive fitness. The reason it still took over the macroscopic world is that evolution does not simply select for reproductive fitness.

Instead, the evolution of sexual dimorphism is a long sequence of strange traps, ratchets and outer-world eldritch cosmic forces that made it somehow inevitable. So let's talk about those things your parents never told you about.

The birds, the bees, and the fission yeast

What bugs me is that, not only most people have absolutely no idea why sexual dimorphism exists, but they seem entirely fine with that. Our lives are punctuated with all sorts of frankly weird practices related to it, but the reasons we ended up there remain obscure even to many biologists.

So I figured I would write up a summary of some popular theories. This way, when time comes, you can explain to your children the long evolutionary trajectory that culminated in VR ChatGPT cat-girlfriends.

(Note 1: As always with evolutionary biology, everything in this article is subject to uncertainty, controversy and mystery. Always keep in mind the Golden Rules of biology: all models are wrong; everything has exceptions; don't talk about fungi; mitochondria is the powerhouse of the cell.)

(Note 2: As this is a bottomless topic, I'll have to make some cuts. I know you're burning to learn about Pseudobiceros hancockanus' penis fencing, but I can't cover everything.)

First, let's get something out of the way.

Something something diversity-generation

I often hear vague explanations about sex being a way to generate genetic diversity. I don't find it compelling. If you want genetic diversity, you can do it in much easier ways than turning into a sexually-reproducing dimorphic species. One of them is, just increase the mutation rate, bro.

Bacteria are good at this. E. coli comes with a whole toolkit of DNA-polymerases with various degrees of accuracy. When everything is going fine, they use the most accurate one to faithfully replicate their genomes. But, in case of particularly bad stress, the bacteria start expressing error-prone polymerases, which increase their mutation rate. Who knows, if the mother cell is going to die anyway, some of the mutant offspring might stumble upon a solution to escape the bad situation.

All that is to say, raw genetic diversity cannot be the whole picture. It has to be a specific kind of genetic diversity.

Part 1: the evolution of sex

Most of the articles I ...

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: Comments on Manheim's "What's in a Pause?", published by RobBensinger on September 18, 2023 on The Effective Altruism Forum.

This post is part of AI Pause Debate Week. Please see this sequence for other posts in the debate.

I agree with David Manheim's post at a high level. I especially agree that a pause on large training runs is needed, that "We absolutely cannot delay responding", and that we should be focusing on a pause mediated by "a multilateral agreement centered on countries and international corporations". I also agree that if we can't respond to the fire today, we should at least be moving fast to get a "sprinkler system".

The basic reason we need a (long) pause, from my perspective, is that we are radically unprepared on a technical level for smarter-than-human AI. We have little notion of how to make such systems reliable or safe, and we'll predictably have very little time to figure this out once smarter-than-human AI is here, before the technology proliferates and causes human extinction.

We need far, far more time to begin building up an alignment field and to develop less opaque approaches to AI, if we're to have a realistic chance of surviving the transition to smarter-than-human AI systems.

My take on AI risk is similar to Eliezer Yudkowsky's, as expressed in his piece in TIME and in the policy agenda he outlined. I think we should be placing more focus on the human extinction and disempowerment risks posed by AGI, and should be putting a heavy focus on the arguments for that position and the reasonably widespread extinction fears among ML professionals.

I have disagreements with some of the specific statements in the post, though in many cases I'm unsure of exactly what Manheim's view is, so the disagreement might turn out to be non-substantive. In the interest of checking my understanding and laying out a few more of my views for discussion, I'll respond to these below.[1]

So the question of whether to stop and how to do so depends on the details of the proposal - but these seem absent from most of the discussion.

This is not apparent to me. I think it would take a pretty unusual proposal in order for me to prefer the status quo over it, assuming the proposal actually pauses progress toward smarter-than-human AI.

It's important to get this right, and the details matter. But if a proposal would actually work then I'm not picky about the additional implementation details, because there's an awful lot at stake, and "actually working" is already an extremely high bar.

An immediate, temporary pause isn't currently possible to monitor, much less enforce, even if it were likely that some or most parties would agree.

A voluntary and temporary moratorium still seems like an obviously good idea to me; it just doesn't go far enough, on its own, to macroscopically increase our odds of surviving AGI. But small movements in the right direction are still worthwhile.

Similarly, a single company, or country announcing a unilateral halt to building advanced models is not credible without assurances,

"Not credible" sounds too strong here, though maybe I'm misunderstanding your claim. Scientists have voluntarily imposed restrictions on their own research in the past (e.g., Asilomar), and I don't think this led to widespread deception. Countries have banned dangerous-but-profitable inventions without pursuing those inventions in secret.

I don't think it would be that hard for many companies or countries to convince me that they're not building advanced models. It might be hard for me to (for example) get to 95% confidence that DeepMind has suspended frontier AI development, merely on DeepMind's say-so; but 75% confidence seems fairly easy to me, if their say-so is concrete and detailed enough.

(Obviously some people will pursue such research in secret, somewhere in t...

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: Where might I direct promising-to-me researchers to apply for alignment jobs/grants?, published by Abram Demski on September 18, 2023 on The AI Alignment Forum.

It sometimes happens that people who I've talked to or worked with ask me where they should go for financial support for their research. I haven't developed a standard list of answers to this question. It seems to me like there are a lot of new orgs recently, and I'm losing track!

If you are looking for such applicants or know someone who is looking, consider replying as an answer (or sending me a PM if that makes more sense for whatever reason).

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: Relationship between EA Community and AI safety, published by Tom Barnes on September 18, 2023 on The Effective Altruism Forum.

Personal opinion only. Inspired by filling out the Meta coordination forum survey.

Epistemic status: Very uncertain, rough speculation. I'd be keen to see more public discussion on this question

One open question about the EA community is it's relationship to AI safety (see e.g. MacAskill). I think the relationship EA and AI safety (+ GHD & animal welfare) previously looked something like this (up until 2022ish):

With the growth of AI safety, I think the field now looks something like this:

It's an open question whether the EA Community should further grow the AI safety field, or whether the EA Community should become a distinct field from AI safety. I think my preferred approach is something like: EA and AI safety grow into new fields rather than into eachother:

AI safety grows in AI/ML communities

EA grows in other specific causes, as well as an "EA-qua-EA" movement.

As an ideal state, I could imagine the EA community being in a similar state w.r.t AI safety that it currently has in animal welfare or global health and development.

However I'm very uncertain about this, and curious to here what other people's takes are.

I've ommited non-AI longtermism, along with other fields, for simplicity. I strongly encourage not interpreting these diagrams too literally

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: Three ways interpretability could be impactful, published by Arthur Conmy on September 18, 2023 on LessWrong.

Epistemic status: somewhat confident that overall these arguments show interpretability could be impactful, but not confident on any individual point.

Background: I wrote a draft of this post here after reading Charbel's work. Thanks to Charbel and several others' feedback, the post was improved a lot: I would definitely recommend adversarial collaboration so thanks Charbel! I'm hoping to discuss interpretability's impact further so posting now.

Three ways interpretability could be impactful

Just read the test and validation set parts if you have little time.

1. Test set interpretability: Finding alignment failures

What this is. We want to find alignment failures with interpretability that other forms of auditing do not find. Before deploying a model, we find inputs that seem to elicit new capabilities of the model. Either A) the model has correctly generalized to novel capabilities or B) the model is deceiving our evaluation process to produce convincing but false outputs. We then use our future interpretability tools to try to understand some of the computation the AI is doing to reach these new capabilities. We either become more confident that the AI has correctly generalized, or we've uncovered deceptive computation and we can retrain and/or provide robust evidence of misalignment to relevant actors.

Why this isn't achievable with other methods. If a model can deceive our evaluation process it could be very difficult to observe this deception with mere behavorial evals. See here.

In my mind test set interpretability primarily targets a specific set of alignment failures, illustrated here:

Figure 1.

In the appendix I outline my reasoning behind (interpretability's role in) Figure 1.

2. Validation set interpretability: A better science of alignment.

What this is (thanks to Neel). We have little ground truth on whether our models are misaligned now or how far methods such as RLHF will further scale. More generally, we understand little about how machine learning works, which limits our ability to reason about future systems. Interpretability could first and foremost actually provide evidence for what our alignment techniques are doing (e.g interpreting RLHF reward models) and secondly give us a better evidence base for reasoning about deep learning. I think that Progress Measures for Grokking Via Mechanistic Interpretability has already somewhat changed people's perspectives on how ML models select different algorithms (e.g here, here).

This differs from test set interpretability as it is broader and can be applied before testing potentially misaligned models, to steer the field towards better practises for alignment (Russell and Norvig's validation/test distinction here may be helpful for analogy).

Why this isn't achievable with other methods. If we want to understand how models work for safety-relevant end goals, it seems likely to me that interpretability is the best research direction to pursue. Most methods are merely behavorial and so provide limited ground truth, especially when we are uncertain about deception. For example, I think the existing work trying to make chain-of-thought faithful shows that naive prompting is likely insufficient to understand models' reasoning. Non-behavorial methods such as science of deep learning approaches (e.g singular learning theory, scaling laws) by default give high-level descriptions of neural network statistics such as loss or RLCT. I don't think these approaches are as likely to get as close to answering the questions about internal computations in AIs that I think successful interpretability could lead to. I think some other directions are worthwhile bets due to uncertainty and neglectedness, however.

3. Train...