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: Epistemic Legibility, published by Elizabeth on February 9, 2022 on LessWrong.

Tl;dr: being easy to argue with is a virtue, separate from being correct.

Introduction

Regular readers of my blog know of my epistemic spot check series, where I take claims (evidential or logical) from a work of nonfiction and check to see if they’re well supported. It’s not a total check of correctness: the goal is to rule out things that are obviously wrong/badly formed before investing much time in a work, and to build up my familiarity with its subject.

Before I did epistemic spot checks, I defined an easy-to-read book as, roughly, imparting an understanding of its claims with as little work from me as possible. After epistemic spot checks, I started defining easy to read as “easy to epistemic spot check”. It should be as easy as possible (but no easier) to identify what claims are load-bearing to a work’s conclusions, and figure out how to check them. This is separate from correctness: things can be extremely legibly wrong. The difference is that when something is legibly wrong someone can tell you why, often quite simply. Illegible things just sit there at an unknown level of correctness, giving the audience no way to engage.

There will be more detailed examples later, but real quick: “The English GDP in 1700 was $890324890. I base this on $TECHNIQUE interpretation of tax records, as recorded in $REFERENCE” is very legible (although probably wrong, since I generated the number by banging on my keyboard). “Historically, England was rich” is not. “Historically, England was richer than France” is somewhere in-between.

“It was easy to apply this blog post format I made up to this book” is not a good name, so I’ve taken to calling the collection of traits that make things easy to check “epistemic legibility”, in the James C. Scott sense of the word legible. Legible works are (comparatively) easy to understand, they require less external context, their explanations scale instead of needing to be tailored for each person. They’re easier to productively disagree with, easier to partially agree with instead of forcing a yes or no, and overall easier to integrate into your own models.

[Like everything in life, epistemic legibility is a spectrum, but I’ll talk about it mostly as a binary for readability’s sake]

When people talk about “legible” in the Scott sense they often mean it as a criticism, because pushing processes to be more legible cuts out illegible sources of value. One of the reasons I chose the term here is that I want to be very clear about the costs of legibility and the harms of demanding it in excess. But I also think epistemic legibility leads people to learn more correct things faster and is typically underprovided in discussion.

If I hear an epistemically legible argument, I have a lot of options. I can point out places I think the author missed data that impacts their conclusion, or made an illogical leap. I can notice when I know of evidence supporting their conclusions that they didn’t mention. I can see implications of their conclusions that they didn’t spell out. I can synthesize with other things I know, that the author didn’t include.

If I hear an illegible argument, I have very few options. Perhaps the best case scenario is that it unlocks something I already knew subconsciously but was unable to articulate, or needed permission to admit. This is a huge service! But if I disagree with the argument, or even just find it suspicious, my options are kind of crap. I write a response of equally low legibility, which is unlikely to improve understanding for anyone. Or I could write up a legible case for why I disagree, but that is much more work than responding to a legible original, and often more work than went into the argument I’m responding to, because it’s not obviou...

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: [Updated] EA conferences in 2022: save the dates, published by Lizka on February 9, 2022 on The Effective Altruism Forum.

This post was updated on February 9 to reflect new dates. We changed the post date to reflect this.

CEA will be running and supporting conferences for the EA community all over the world in 2022.

We are currently organizing the following events. All dates are provisional and may change in response to local COVID-19 restrictions.

EA Global conferences

EA Global: London (15 - 17 April)[1]

EA Global: San Francisco (29 - 31 July)

EA Global: Washington, D.C. (23 - 25 September)

EAGx conferences

EAGx Oxford (26 - 27 March)

EAGx Boston (1 - 3 April)

EAGx Prague (13 - 15 May)

EAGx Australia (8 - 10 July)

EAGx Singapore (2 - 4 September)

EAGx Berlin (September/October)

Applications for the EA Global: London and EAGx Oxford conferences will open soon—keep an eye out! If you want to make sure that you will get an email reminder, sign up here.

Some notes on these conferences:

EA Global conferences are for people who are knowledgeable about the core ideas of effective altruism and are taking significant actions (e.g. work or study) based on these ideas.

To attend an EAGx conference, you should at least be familiar with the core ideas of effective altruism.

Please apply to all conferences you wish to attend once applications open — we would rather get too many applications for some conferences and recommend that applicants attend a different one, than miss out on potential applicants to a conference.

Applicants can request financial aid to cover the costs of travel, accommodation and tickets.

Find more info on our website.

As always, please feel free to email [email protected] with any questions, or comment below.

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: Prediction Markets are for Outcomes Beyond Our Control, published by lsusr on February 9, 2022 on LessWrong.

Betting markets are the gold standard of expert predictions because bets are the ultimate test of what people truly believe.

The best betting markets are highly liquid. A liquid market is one where you can place a large bet without moving the price very much. Liquid prediction markets work when no individual person can influence the outcome. Bettering markets are a great way to find out if "it will rain tomorrow" or whether "candidate x will be elected president next year".

But what if a single person can influence the outcome? For example, what would happen if I created a betting market for "Lsusr will publish a blog post tomorrow"?

Suppose I am ambivalent about whether I will publish a blog post tomorrow. If the price of "Lsusr will publish a blog post tomorrow" drops below 1.00 then I will buy shares of "Lsusr will publish a blog post tomorrow" and then pocket a risk-free profit by posting a blog post tomorrow. If the price of "Lsusr will publish a blog post tomorrow rises above 0.00 then I will buy shares of "Lsusr will not publish a blog post tomorrow" and then pocket a risk-free profit by not posting a blog post tomorrow.

The market equilibrium occurs even if I am unaware that the prediction market exists. Suppose the price drops to 0.99. A trader could buy shares and then pay me a small fee to influence the outcome.

I am not truly ambivalent. Suppose I'm willing to influence the outcome in exchange for $500. What happens? If the market liquidity is less than $500 then we have a functional prediction market. If the market liquidity is more than $500 then we have a regular market.

Prediction markets function best when liquidity is high, but they break completely if the liquidity exceeds the price of influencing the outcome. Prediction markets function only in situations where outcomes are expensive to influence.

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: [Intro to brain-like-AGI safety] 3. Two subsystems: Learning & Steering, published by Steve Byrnes on February 9, 2022 on The AI Alignment Forum.

3.1 Post summary / Table of contents

Part of the “Intro to brain-like-AGI safety” post series.

In the previous post I defined the notion of “learning from scratch” algorithms—a broad category that includes, among other things, any randomly-initialized machine learning algorithm (no matter how complicated), and any memory system that starts out empty. I then proposed a division of the brain into two parts based on whether or not they learn from scratch. Now I’m giving them names:

The Learning Subsystem is the 96% of the brain that “learns from scratch”—basically the telencephalon and cerebellum.

The Steering Subsystem is the 4% of the brain that doesn’t “learn from scratch”—basically the hypothalamus and brainstem.

(See previous post for a more detailed anatomical breakdown.)

This post will be a discussion of this two-subsystems picture in general, and of the Steering Subsystem in particular.

In Section 3.2, I’ll talk about the big picture of what these subsystems do and how they interact. As an example, I’ll explain why each subsystem needs its own sensory-processing circuitry—for example, why visual inputs get processed by both the visual cortex in the Learning Subsystem, and the superior colliculus in the Steering Subsystem.

In Section 3.3, I’ll acknowledge that this two-subsystem picture has some echoes of the discredited “triune brain theory”. But I’ll argue that the various problems with triune brain theory do not apply to my two-subsystem picture.

In Section 3.4, I’ll discuss three categories of ingredients that could go into a Steering Subsystem:

Category A: Things that are plausibly essential for general intelligence (e.g. an innate drive for curiosity),

Category B: Everything else in the human steering subsystem (e.g. an innate drive to be kind to your friends),

Category C: Any other possibility that an AGI programmer might dream up, even if it’s radically different from anything in humans or animals (e.g. an innate drive to correctly predict stock prices).

In Section 3.5, I’ll relate those categories to how I expect people to build brain-like AGIs, arguing that “brain-like AGIs with radically non-human (and dangerous) motivations” is not an oxymoron; rather, it’s the default expected outcome, unless we work to prevent it.

In Section 3.6, I’ll discuss the fact that Jeff Hawkins has a two-subsystems perspective similar to mine, yet argues against AGI catastrophic accidents being a risk. I’ll say where I think he goes wrong.

Sections 3.7 and 3.8 will be the final two parts of my “timelines to brain-like AGI” discussion. The first part was Section 2.8 in the previous post, where I argued that reverse-engineering the Learning Subsystem (at least well enough to enable brain-like AGI) is something that could plausibly happen soon, like within the next decade or two, although it could also take longer. Here, I’ll complete that story by arguing that this same thing is true of reverse-engineering the Steering Subsystem (at least well enough to enable brain-like AGI), and of getting the algorithms cleaned up and scaled up, running model trainings, and so on.

Section 3.9 is a quick non-technical discussion on the wildly divergent attitudes that different people take towards the timeline to AGI, even when they agree on the probabilities. For example, you can have two people agree that the odds are 3:1 against having AGI by 2042, but one might emphasize how low that probability is (“You see? AGI probably isn’t going to arrive for decades”), while the other might emphasize how high that probability is. I’ll talk a bit about the factors that can underlie those attitudes.

3.2 Big picture

In the last post, I claimed that 96% of 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: What (standalone) LessWrong posts would you recommend to most EA community members?, published by vaidehi agarwalla on February 9, 2022 on The Effective Altruism Forum.

Context / Motivation:

I am interested in thinking a lot recently about how we could share ideas from the rationalist community to EA and related subcommunities (perhaps communicating the same ideas in different ways).

I've been diving into LessWrong recently and remembered why I hadn't for a while - it's really overwhelming. Even with the sequences and curation, it's a lot of content, and it's not always obvious to me which posts I'd find most valuable.

I think it's better to read fewer posts in more depth to properly understand them.

I think it's likely that some posts or ideas will be much more relevant to EAs than others, but I'm not sure which ones

My ask:

I'd be interested in recommendations for standlone posts (e.g. All debates are bravery debates), specific concepts (e.g. schelling fence or doublecrux) , or specific sequences (e.g.)

If you have time, I'd love to know why it's valuable to you

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 case for "mental strength", published by VipulNaik on February 7, 2022 on LessWrong.

Recently, I was thinking about the differences between physical health and physical fitness, and wondering if there is an analogous distinction on the mental side. Googling, I found that the term

"mental strength" is often used for the mental counterpart to physical fitness; "mental fitness" is tied too heavily to a narrower conception focused on intelligence and cognition.

I found this Verywell Mind article to be a good overview (and I recommend reading it though I don't assume it as a prerequisite). It identifies three broad components to mental strength: cognitive (thinking), emotional (feeling), and behavioral (acting).

I expect that most of the rest of what I write will be fairly obvious to many LessWrong readers, though I am not absolutely sure. I am still going to say them, for reasons Grognor articulated.

The post makes the most sense when read sequentially, but you can skim it or read whatever parts interest you based on the table of contents. The first section provides a few angles from which to think about mental strength. In the second section, I talk a little more about what I like about the "mental strength" concept and some of its implications. The third section talks about the relationship between mental strength and the LessWrong/rationality cluster.

Relation between mental strength and other concepts

Mental strength versus mental health

The Verywell Mind article includes a good comparison of the two concepts that I recommend reading. I will try to summarize it in my own words, with slightly different emphases.

Mental health is about your current mental state, or your average mental state over time. For instance, do you often feel stress and anxiety? Do you feel uncontrollable anger? Do you have depressive thought patterns? Or are you generally satisfied and content with yourself? Are you able to focus on the tasks that are important to you?

Mental strength is about the inner capacity and resilience you have developed (cognitively, emotionally, and behaviorally) to be mentally healthy across a range of life circumstances. Two people with the same mental strength might have different levels of mental health based on the circumstances they are in. However, on average, for people in similar circumstances, greater mental strength is likely to lead to greater mental health. (This is a bit of a simplification as neither mental health nor mental strength is one-dimensional, but it's still useful directionally).

The physical analogy may clarify this: physical exercise builds physical strength. If you take two people, one of whom has built more physical strength, and put them through the same physical ordeal, it's likely that the person with more physical strength will come out of the ordeal in better shape. But if you compare one person with physical strength who's going through a physical ordeal, and another weaker person who's relaxing, the former may be in worse shape!

To put it another way, mental health depends on mental strength and on how stressful the circumstances are. The more stressful the circumstances, the more mental strength is needed to achieve a given level of mental health. For instance, if you're living in a war zone with shooting happening all the time around you, the amount of mental strength you need to stay mentally healthy is probably higher than if you're living a relaxed life. In general, mortal danger, poverty, abuse, and high-stress work can worsen mental health and need more mental strength.

In a later section, I'll talk specifically of the mental strength demands placed in rationalist and effective altruist circles.

Mental strength versus mental toughness

Mental toughness seems to be a subset of mental strength focused on resilience and conf...

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: Hypothesis: gradient descent prefers general circuits, published by Quintin Pope on February 8, 2022 on The AI Alignment Forum.

Summary: I discuss a potential mechanistic explanation for why SGD might prefer general circuits for generating model outputs. I use this preference to explain how models can learn to generalize even after reaching near zero training error through overfitting (i.e., grokking). I also discuss other models of grokking and generalization.

Additionally, I discuss potential experiments to confirm or reject my hypothesis. I suggest that a tendency to unify many shallow patterns into fewer general patterns is a core feature of effective learning systems, potentially including humans and future AI, and briefly address implications to AI alignment.

Epistemic status: I think the hypothesis I present makes a lot of sense and is probably true, but I haven't confirmed things experimentally. Much of my motive for post this is to clarify my own thinking and get feedback on the best ways to experimentally validate this perspective on ML generalization.

Context about circuits: This post assumes the reader is familiar with and accepts the circuits perspective on deep learning. See here for a discussion of circuits for CNN vision models and here for a discussion of circuits for transformer NLP models.

Evidence from grokking

The paper "Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets" uses stochastic gradient descent (SGD) to train self attention based deep learning models on different modular arithmetic expressions (e.g., f(x,y)=x×y (mod p), where p is fixed).

The training data only contain subsets of the function's possible input/output pairs. Initially, the models overfit to their training data and are unable to generalize to the validation input/output pairs. In fact, the models quickly reach near perfect accuracy on their training data. However, training the model for significantly past the point of overfitting causes the model to generalize to the validation data, what the authors call "grokking".

See figure 1a from the paper:

Note the model has near perfect accuracy on the training data. Thanks to a recent replication of this work, we can also look at the loss curves during grokking (though on a different experiment compared to the plot above):

First, the model reaches near-zero loss in training but overfits in validation. However, the validation loss soon starts decreasing until the model correctly classifies both the validation and training data.

This brings up an interesting question: why did the model learn anything at all after reaching near zero loss on the training data? Why not just stick with memorizing the training data? What would prompt SGD to switch over to general circuitry that solves both training and validation data?

I think the answer is surprisingly straightforward: SGD prefers general circuits because general circuits make predictions on a greater fraction of the training data. Thus, general circuits receive more frequent SGD reinforcement for making correct predictions. Think of each data point as "pushing" the model to form circuits that perform well on that data point. General circuits perform well on many data points, so they receive a greater "push" towards forming.

Shallow circuits are easier to form with SGD, but they aren't retained as strongly. Thus, as training progresses, general circuits eventually overtake the shallow circuits.

Toy example of SGD preferring general circuits

Let's consider a toy example of memorizing two input/output pairs. Suppose:

input 1 = (x1,y1)

output 1 = f(x1,y1)

input 2 = (x2,y2)

output 2 = f(x2,y2)

One way the model might memorize the these data points is to use two independent, shallow circuits, one for each data point. I show a diagram of how this might be implemented us...

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: We are giving $10k as forecasting micro-grants, published by Misha Yagudin on February 8, 2022 on The Effective Altruism Forum.

(Cross-posted from the Forecasting Newsletter.)

After the apparent success of ACX grants (a), we received $10k from an anonymous donor to give out as micro-grants through Nuño's Forecasting newsletter.

Some examples of projects we'd be excited to fund might be:

Open-source software and tooling to automate forecasting. Bonus points if Metaculus users or other forecasters start using it.

Giving a shot at forecasting or making models of a difficult yet useful and decision-relevant area. Think of Rootclaim (a) analyzing the lab-escape story of Omicron.

Pieces similar in quality to the ones mentioned in "best pieces on forecasting from 2021 (a)”, in Forecasting Prize Results (a), or in some possible research areas (a).

Trying to estimate many uncertain parameters, e.g., the quality of all US or UN organizations, quality of academic fields, whether a large list of organizations will fail, enlightened willingness to pay for many products, the accuracy of many public figures, etc.

Create a microcovid (a) or foodimpacts (a) but for other areas, like micro-marriages, micro-insights, micro-dooms, etc. Do this in a way that easily allows the creation of many of these calculators.

Improve metaforecast (a) (which is open source (a)) in some interesting way, e.g., improve the estimates of forecast quality.

The application form is HERE (a). Feel free to apply for more than $10k: we don't anticipate having much difficulty getting more funding for promising applications, and we may refer these to other funders (e.g., the EA Infrastructure Fund (a)) if we can't.

We will be accepting applications until March 15, though we may extend this period if we don't receive enough high-quality submissions. We preliminarily plan to make decisions by April the 1st.

Otherwise, Luke Muehlhauser comments (a) that forecasting related projects might be a good fit for the EA Infrastructure Fund (a). Jonas Vollmer, who runs EA Funds, confirms this (a)

For larger projects, the Survival and Flourishing Fund, backed by philanthropists Jaan Tallinn and Jed McCaleb, is organizing the distribution of around $6M-$10M in grants (a) this June, with applications due on Feb 21. They generally only accept applications from registered charities, but speculation grants (a) might be a good fit for smaller projects (40%).

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: New 80,000 Hours career review: Software engineering, published by Benjamin Hilton on February 8, 2022 on The Effective Altruism Forum.

Summary

Software engineering could be a great option for having a direct impact on the world’s most pressing problems, particularly in AI safety, but also in biosecurity and across other cause areas. This will probably be more impactful than earning to give.

As with operations staff, organisations need exceptional and mission-aligned software engineers. But many still find it difficult to hire.

Some myths:

You need an ML background to work as an engineer on AI safety.

Outside AI safety, the only useful software skill is front-end web development.

Effective organisations will pay far less than top companies.

None of these things are true.

In fact, many organisations have budgets of $10s of millions, and think that software engineers can substantially increase their cost-effectiveness (e.g. in his 80,000 Hours podcast, Chris Olah argues that Anthropic's systems researcher could easily increase their efficiency by at least 10%). So even if you're earning 7 figures, you could be more effective doing direct impact work.

This rest of this post contains an excerpt from my new career review of software engineering for 80,000 Hours, focusing on the parts most relevant to already-engaged EAs.

This review owes a lot to helpful discussions with (and comments from) Andy Jones, Ozzie Gooen, Jeff Kaufman, Sasha Cooper, Ben Kuhn, Nova DasSarma, Kamal Ndousse, Ethan Alley, Ben West, Ben Mann, Tom Conerly, Zac Hatfield-Dodds, and George McGowan. Special thanks go to Roman Duda for our previous review of software engineering, on which this was based.

Why might software engineering be high impact?

Software engineers are in a position to meaningfully contribute directly to solving a wide variety of the world’s most pressing problems.

In particular, there is a shortage of software engineers at the cutting edge of research into AI safety.

We’ve also found that software engineers can contribute greatly to work aiming at preventing pandemics and other global catastrophic biological risks.

Aside from direct work on these crucial problems, while working for startups or larger tech companies you can gain excellent career capital (especially technical skills), and, if you choose, earn and donate substantial amounts to the world’s best charities.

How to do good as a software engineer

Even for skilled engineers who could command high salaries, we think that working directly on a problem will probably be more impactful than earning to give.

Some examples of projects where software engineering is central to their impactful work:

Ought is an AI research lab attempting to build systems that apply machine learning to the task of helping people think.

The Secure DNA Project is attempting to build a secure global screening system for DNA sequences that could be used to engineer a global pandemic.

Momentum is a startup building donation pages that encourage recurring donations to the world’s most effective charities.

Lightcone Infrastructure builds software-based infrastructure for longtermist projects.

Telis Bioscience is a startup attempting to radically accelerate drug development to mitigate risks from future pandemics.

Anthropic is a research company working to build reliable, interpretable, and steerable AI systems.

Redwood Research conducts applied research into the challenge of aligning artificial intelligence.

Wave is a startup building a way for people in developing countries to access financial services.

Most organisations, even ones that don’t focus on developing large software products, need software engineers to manage computer systems, apps, and websites. For example:

Effective Altruism Funds provides expert-managed funds to facilitate donors maximising 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: The Case for Rare Chinese Tofus, published by George Stiffman on February 8, 2022 on The Effective Altruism Forum.

A thank you to Kris Chari and Pranay Mittal for reading an earlier draft.

Purpose

AI alignment seems far more pressing than factory farming, but I think I could have more impact in the latter. This belief seems suspect, given the overwhelming movement of EAs into longtermist roles, so I’d like to test it out. Below is my current case for working in farmed animal welfare (FAW). Please critique.

If this post goes well (meaning it isn’t ripped to shreds), then I’ll write up a comparison piece on longtermist community building.

Besides offering career advice, I hope you can learn a bit about the incredible world of rare Chinese tofus, which over the last three years has given my life so much joy!

The Case for Growing the United States Tofu Market

I think the U.S. tofu market could grow several times over, eventually averting 10s to 100s of millions of factory farmed lives each year.

Problem

To a large degree, American omnivores object to vegan food because it’s reductionary, focused on subtraction (i.e. chicken-less chicken salads) and substitution (i.e. inferior alt proteins). For most consumers, these foods taste chronically worse than the originals. It’s unclear how long it will be till we have high fidelity substitutes across a range of options, but the timeline isn’t years—it’s decades.

Public aversion is exacerbated by chefs. Because they see plant-based ingredients (especially proteins) as less versatile than meat, they promote meat-dominated menus, reinforcing the belief that vegan food is inherently worse.

This is in spite of the fact that chefs, like the general public, care about climate change and, to a lesser extent, animal welfare, and want to cook more plants.

Solution

To go mainstream, vegan cuisine should explore moving beyond reduction–basing itself off meat–towards creation–leveraging culinary strengths of plant-based ingredients to build all new foods.

In 2021, a team of chefs and I researched how to create new foods and found a simple yet effective formula: cross rare Chinese tofus with traditional, western cooking methods.

In the west, tofu is often seen as an ingredient, but it's actually a category of ingredients, like chicken. In China, the birthplace and mecca of tofu, there are over 25 distinct types, which are as dissimilar as chicken feet from chicken breast. The most common types of tofu in the States–firm, soft, and silken–are like chicken feet; while popular in Asia, they are poor fits for western cooking methods and taste preferences. In contrast, there are other varieties that while rare in Asia are great fits–like chicken breast.

If our goal is to create new foods, the best people to do so are restaurant chefs. Fortunately, they have many motivations to help. Chefs love using sexy new ingredients. Many care about sustainability, or at least being perceived as sustainable. They desire positive press, which can be achieved by trying new, bold things. Restaurant goers desire more plant-based options, which at most U.S. restaurants are woefully lacking. The point is that there are many incentives which align with cuisine expansion, and which we can support.

It seems that for many ingredients, uses follow a heavy-tailed distribution. (How often do you see saffron outside of paella? Or peanut butter without bread?) The discovery of scalable use cases will mean the success or failure of tofu growth. If chefs unlock the “acai bowl” or “avocado toast” of rare Chinese tofus, then the market will grow. If we can motivate a decentralized community of chefs to independently create new foods, I'm confident that these use cases will emerge.

Plan

Build consumer hype around rare tofus (chefs seek out sexy ingredients): partner with a chef a...