This essay will require you to watch three short Youtube videos, totaling less than two minutes.

Naming things is hard. Generally speaking, a thing should be named evocatively, such that people find it memorable and sticky, or precisely, such that people can reconstruct the concept just from its title.

(So, "Moloch," or "trigger-action planning.")

This essay is about "cup-stacking skills." It's a noun that I use in phrases like "I think you're exhibiting a cup-stacking skill right now" or "I'm slowly trying to unravel this cup-stacking skill" or "I think we should consult Dave; he has the relevant cup-stacking skill."

Unfortunately, that's not a great name. Most people, encountering the name, will have to memorize both the concept and the label, rather than having to just memorize the concept and have the label stick, or just memorize the label (and being able to rederive the concept from it).

Sorry. I've made a genuine effort for the past couple of years to find a better name, and failed. Since I've failed, I need you to watch three Youtube videos.

Here's the first video.

This is me, in my kitchen, cup-stacking. It's a fun little game-slash-sport in which you stack and unstack cups in a specific pattern, to see how fast you can go. It's extremely rewarding once you get even a tiny bit good at it; you can feel things going almost-right and the pattern loops onto itself and it's very easy to just chase that feeling of smoothness for hours at a time. I've probably put between 50 and 100 hours into cup stacking over the past ten years, though at the time of filming I hadn't pulled them out much at all in the past two.

Here's the second video.

This is my partner Logan, cup-stacking for the very first time, after having watched me demonstrate the pattern exactly twice. They've got a little card for reference on the table with them, so they know each of the three end-states they're shooting for, but otherwise I told them to not worry about process or technique and just generally do their best to imitate what they'd seen in a low-stress sort of way.

The thing about the literal skill of cup-stacking is that (approximately) "anyone can do it." Even as total beginners, most people can follow the directions and hold the pattern in mind and get the cups to stack up in the right shapes.

There's obviously a big difference between someone who's practiced for 50 hours and someone who's practiced for zero. But it feels like a quantitative difference, not a qualitative one. There are little bits of the technique that I am doing that Logan doesn't yet know about, and some basic misunderstandings (they're using their hands symmetrically rather than complementarily), but in-the-absolutely-literal-sense-of-the-word essentially, Logan and I are both attempting, and succeeding at, the same task.

The cups go up, and the cups come down.

This is, in my mind, a pretty solid metaphor for most of what would count as "rationalist skill." Things like checking for truth, recognizing cognitive biases, zeroing in on cruxes, doing intelligent emotional regulation, and employing formalized techniques like TAPs or goal factoring or Gendlin's Focusing.

All of that stuff is wildly popular with a certain class of nerd (in part) because it's accessible. You can pick up the core of the concept in the course of a five-minute lecture, and test it out in the course of a five-minute timer. You can start doing it right away, as a total novice, and see it working, in the way you hoped it would work. And with 50 hours of practice, it goes much more fluidly and reliably and is integrated much better (though still not perfectly).

This is very different from, say, gymnastics, or learning programming from scratch, where many of the learning paths involve spending a lot of time establishing a foundation of background skills and concepts before ever getting to "the good stuff."

Here is the third video.

I have paused for emphasis.

The third video is still technically just a quant...

This essay will require you to watch three short Youtube videos, totaling less than two minutes.

Naming things is hard. Generally speaking, a thing should be named evocatively, such that people find it memorable and sticky, or precisely, such that people can reconstruct the concept just from its title.

(So, "Moloch," or "trigger-action planning.")

This essay is about "cup-stacking skills." It's a noun that I use in phrases like "I think you're exhibiting a cup-stacking skill right now" or "I'm slowly trying to unravel this cup-stacking skill" or "I think we should consult Dave; he has the relevant cup-stacking skill."

Unfortunately, that's not a great name. Most people, encountering the name, will have to memorize both the concept and the label, rather than having to just memorize the concept and have the label stick, or just memorize the label (and being able to rederive the concept from it).

Sorry. I've made a genuine effort for the past couple of years to find a better name, and failed. Since I've failed, I need you to watch three Youtube videos.

Here's the first video.

This is me, in my kitchen, cup-stacking. It's a fun little game-slash-sport in which you stack and unstack cups in a specific pattern, to see how fast you can go. It's extremely rewarding once you get even a tiny bit good at it; you can feel things going almost-right and the pattern loops onto itself and it's very easy to just chase that feeling of smoothness for hours at a time. I've probably put between 50 and 100 hours into cup stacking over the past ten years, though at the time of filming I hadn't pulled them out much at all in the past two.

Here's the second video.

This is my partner Logan, cup-stacking for the very first time, after having watched me demonstrate the pattern exactly twice. They've got a little card for reference on the table with them, so they know each of the three end-states they're shooting for, but otherwise I told them to not worry about process or technique and just generally do their best to imitate what they'd seen in a low-stress sort of way.

The thing about the literal skill of cup-stacking is that (approximately) "anyone can do it." Even as total beginners, most people can follow the directions and hold the pattern in mind and get the cups to stack up in the right shapes.

There's obviously a big difference between someone who's practiced for 50 hours and someone who's practiced for zero. But it feels like a quantitative difference, not a qualitative one. There are little bits of the technique that I am doing that Logan doesn't yet know about, and some basic misunderstandings (they're using their hands symmetrically rather than complementarily), but in-the-absolutely-literal-sense-of-the-word essentially, Logan and I are both attempting, and succeeding at, the same task.

The cups go up, and the cups come down.

This is, in my mind, a pretty solid metaphor for most of what would count as "rationalist skill." Things like checking for truth, recognizing cognitive biases, zeroing in on cruxes, doing intelligent emotional regulation, and employing formalized techniques like TAPs or goal factoring or Gendlin's Focusing.

All of that stuff is wildly popular with a certain class of nerd (in part) because it's accessible. You can pick up the core of the concept in the course of a five-minute lecture, and test it out in the course of a five-minute timer. You can start doing it right away, as a total novice, and see it working, in the way you hoped it would work. And with 50 hours of practice, it goes much more fluidly and reliably and is integrated much better (though still not perfectly).

This is very different from, say, gymnastics, or learning programming from scratch, where many of the learning paths involve spending a lot of time establishing a foundation of background skills and concepts before ever getting to "the good stuff."

Here is the third video.

I have paused for emphasis.

The third video is still technically just a quant...

Some early career EAs are able to find good direct work, but there aren’t enough opportunities for all the promising early career people. Another option that I don’t think gets considered enough is to work at a fast-growing startup in an emerging technology (e.g. robotics or NLP), where one can often have much faster skill development within a couple of years. By fast-growing startup, I mean a company that seems decently likely to be one of the top ~20 highest valued startups founded in a given 5 year period.

Claim: it would often be more valuable to work at a fast-growing startup than direct work that doesn’t seem likely to have high positive counterfactual impact or high potential for individual growth.

The three main benefits from fast-growing startups are that the company’s growth leads to greater responsibility earlier, you can work with very high-caliber people and develop a better bar for excellence, and that fast-growing startups often work on the edges of technological development, giving insights into how the world will change over the next decade.

Aurora, a >$10B self-driving car company, had ~30 employees when I joined, and ~300 when I left two years later. About 6 months after joining, I started leading a team of ~5 engineers on a high priority engineering project. That was mostly due to the company needing leaders to keep up with our growth, and my hustle and generalist skills making me well-suited for the role. That experience taught me a lot about leadership, management, and long-term engineering projects, and it seems like this type of experience is much more common in fast-growing startups. In contract, nonprofits often grow slowly or not at all.

An additional benefit is working with very high caliber people, and getting a sense of what a highly successful company looks like. It’s useful to have a well-calibrated bar for who you should work with in the future and who to hire -- I think it would be pretty valuable if more people in the community had well-defined standards of excellence. To quantify this, I think I probably worked with at least 5 of the best 100 people who have worked on self-driving cars in the past decade, and at least 15 people that could get hired to lead a team at basically any self-driving or robotics company.

It’s also useful to see trends in fast-growing startups to understand how the world is changing. At Aurora, I learned about how people are thinking about ML engineering and deploying ML products, and which parts of the ML industry were real or all-hype. Learnings on the front of developing technologies seem a lot more useful for doing impactful work later than learning about random web apps, because the learnings are more applicable for direct work (e.g. AI research).

Some early career EAs are able to find good direct work, but there aren’t enough opportunities for all the promising early career people. Another option that I don’t think gets considered enough is to work at a fast-growing startup in an emerging technology (e.g. robotics or NLP), where one can often have much faster skill development within a couple of years. By fast-growing startup, I mean a company that seems decently likely to be one of the top ~20 highest valued startups founded in a given 5 year period.

Claim: it would often be more valuable to work at a fast-growing startup than direct work that doesn’t seem likely to have high positive counterfactual impact or high potential for individual growth.

The three main benefits from fast-growing startups are that the company’s growth leads to greater responsibility earlier, you can work with very high-caliber people and develop a better bar for excellence, and that fast-growing startups often work on the edges of technological development, giving insights into how the world will change over the next decade.

Aurora, a >$10B self-driving car company, had ~30 employees when I joined, and ~300 when I left two years later. About 6 months after joining, I started leading a team of ~5 engineers on a high priority engineering project. That was mostly due to the company needing leaders to keep up with our growth, and my hustle and generalist skills making me well-suited for the role. That experience taught me a lot about leadership, management, and long-term engineering projects, and it seems like this type of experience is much more common in fast-growing startups. In contract, nonprofits often grow slowly or not at all.

An additional benefit is working with very high caliber people, and getting a sense of what a highly successful company looks like. It’s useful to have a well-calibrated bar for who you should work with in the future and who to hire -- I think it would be pretty valuable if more people in the community had well-defined standards of excellence. To quantify this, I think I probably worked with at least 5 of the best 100 people who have worked on self-driving cars in the past decade, and at least 15 people that could get hired to lead a team at basically any self-driving or robotics company.

It’s also useful to see trends in fast-growing startups to understand how the world is changing. At Aurora, I learned about how people are thinking about ML engineering and deploying ML products, and which parts of the ML industry were real or all-hype. Learnings on the front of developing technologies seem a lot more useful for doing impactful work later than learning about random web apps, because the learnings are more applicable for direct work (e.g. AI research).

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. Published on October 11, 2021.

Crossposted to the EA Forum

Introduction

Imagine you are tasked with curing a disease which hasn’t appeared yet. Setting aside why you would know about such a disease’s emergence in the future, how would you go about curing it? You can’t directly or indirectly gather data about the disease itself, as it doesn’t exist yet; you can’t try new drugs to see if they work, as the disease doesn’t exist yet; you can’t do anything that is experimental in any way on the disease itself, as. you get the gist. You would be forgiven for thinking that there is nothing you can do about it right now.

AI Alignment looks even more hopeless: it’s about solving a never-before seen problem on a technology which doesn’t exist yet.

Yet researchers are actually working on it! There are papers, books, unconferences, whole online forums full of alignment research, both conceptual and applied. Some of these work in a purely abstract and theoretical realm, others study our best current analogous of Human Level AI/Transformative AI/AGI, still others iterate on current technologies that seem precursors to these kinds of AI, typically large language models. With so many different approaches, how can we know if we’re making progress or just grasping at straws?

Intuitively, the latter two approaches sound more like how we should produce knowledge: they take their epistemic strategies (ways of producing knowledge) out of Science and Engineering, the two cornerstones of knowledge and technology in the modern world. Yet recall that in alignment, models of the actual problem and/or technology can’t be evaluated experimentally, and one cannot try and iterate on proposed solutions directly. So when we take inspiration from Science and Engineering (and I want people to do that), we must be careful and realize that most of the guarantees and checks we associate with both are simply not present in alignment, for the decidedly pedestrian reason that Human Level AI/Transformative AI/AGI doesn’t exist yet.

I thus claim that:

Epistemic strategies from Science and Engineering don’t dominate other strategies in alignment research.

Given the hardness of grounding knowledge in alignment, we should leverage every epistemic strategy we can find and get away with.

These epistemic strategies should be made explicit and examined. Both the ones taken or adapted from Science and Engineering, and every other one (for example the more theoretical and philosophical strategies favored in conceptual alignment research).

This matters a lot, as it underlies many issues and confusions in how alignment is discussed, taught, created and criticized. Having such a diverse array of epistemic strategies is fascinating, but their implicit nature makes it challenging to communicate with newcomers, outsiders, and even fellow researchers leveraging different strategies. Here is a non-exhaustive list of issues that boil down to epistemic strategy confusion:

There is a strong natural bias towards believing that taking epistemic strategies from Science and Engineering automatically leads to work that is valuable for alignment.

This leads to a flurry of work (often by well-meaning ML researchers/engineers) that doesn’t tackle or help with alignment, and might push capabilities (in an imagined tradeoff with the assumed alignment benefits of the work)

Very important: that doesn’t mean I consider work based on ML as useless for alignment. Just that the sort of ML-based work that actually tries to solve the problem tends to be by people who understand that one must be careful when transfering epistemic strategies to alignment.

There is a strong natural bias towards disparaging any epistemic strategy that doesn’t align neatly with the main strategies of Science and Engineering (even when the alternative epistemic strategies are 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. Published on October 11, 2021.

Crossposted to the EA Forum

Introduction

Imagine you are tasked with curing a disease which hasn’t appeared yet. Setting aside why you would know about such a disease’s emergence in the future, how would you go about curing it? You can’t directly or indirectly gather data about the disease itself, as it doesn’t exist yet; you can’t try new drugs to see if they work, as the disease doesn’t exist yet; you can’t do anything that is experimental in any way on the disease itself, as. you get the gist. You would be forgiven for thinking that there is nothing you can do about it right now.

AI Alignment looks even more hopeless: it’s about solving a never-before seen problem on a technology which doesn’t exist yet.

Yet researchers are actually working on it! There are papers, books, unconferences, whole online forums full of alignment research, both conceptual and applied. Some of these work in a purely abstract and theoretical realm, others study our best current analogous of Human Level AI/Transformative AI/AGI, still others iterate on current technologies that seem precursors to these kinds of AI, typically large language models. With so many different approaches, how can we know if we’re making progress or just grasping at straws?

Intuitively, the latter two approaches sound more like how we should produce knowledge: they take their epistemic strategies (ways of producing knowledge) out of Science and Engineering, the two cornerstones of knowledge and technology in the modern world. Yet recall that in alignment, models of the actual problem and/or technology can’t be evaluated experimentally, and one cannot try and iterate on proposed solutions directly. So when we take inspiration from Science and Engineering (and I want people to do that), we must be careful and realize that most of the guarantees and checks we associate with both are simply not present in alignment, for the decidedly pedestrian reason that Human Level AI/Transformative AI/AGI doesn’t exist yet.

I thus claim that:

Epistemic strategies from Science and Engineering don’t dominate other strategies in alignment research.

Given the hardness of grounding knowledge in alignment, we should leverage every epistemic strategy we can find and get away with.

These epistemic strategies should be made explicit and examined. Both the ones taken or adapted from Science and Engineering, and every other one (for example the more theoretical and philosophical strategies favored in conceptual alignment research).

This matters a lot, as it underlies many issues and confusions in how alignment is discussed, taught, created and criticized. Having such a diverse array of epistemic strategies is fascinating, but their implicit nature makes it challenging to communicate with newcomers, outsiders, and even fellow researchers leveraging different strategies. Here is a non-exhaustive list of issues that boil down to epistemic strategy confusion:

There is a strong natural bias towards believing that taking epistemic strategies from Science and Engineering automatically leads to work that is valuable for alignment.

This leads to a flurry of work (often by well-meaning ML researchers/engineers) that doesn’t tackle or help with alignment, and might push capabilities (in an imagined tradeoff with the assumed alignment benefits of the work)

Very important: that doesn’t mean I consider work based on ML as useless for alignment. Just that the sort of ML-based work that actually tries to solve the problem tends to be by people who understand that one must be careful when transfering epistemic strategies to alignment.

There is a strong natural bias towards disparaging any epistemic strategy that doesn’t align neatly with the main strategies of Science and Engineering (even when the alternative epistemic strategies are 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. Published on October 10, 2021.

Over the last two years I have been researching and advising numerous government officials on how to do long-term thinking well. Now you are probably not going to be shocked to hear me say it but: making long-term decisions is difficult. Most institutions don’t do it very well and the feedback loops to tell us what works are, as you would expect, long. Yet, that said, it is neither a new challenge nor an uncommon challenge. Many groups of people have faced this problem before and developed tools, strategies and approaches that seem to be working. I have been trying to pull all of this together to paint a rough picture of what best practice in long-term thinking looks like, and advise governments accordingly.

To do that work well I did of course engage in some depth with relevant academic work, including the research on longtermism both from within academia and on this forum. And lo and behold it seemed to me like the space was divided into two distinct camps: longtermist theorists and long-term practitioners. The theorists wonder why policy makers do not listen to them and the policy practitioners wonder why academics are not producing work relevant to them. As a practitioner, it seemed that on some days I would say something that was obvious to me and a researcher would be excited by how novel and useful it is, yet on other days I just could not understand the things longtermist researchers were doing and why it mattered. This post is an attempt to bridge this divide.

The post is in three parts:

Section A is descriptive. I invite you to look around my world, at the politicians, policy makers and risk planners who think long-term. I draw examples from fields as diverse as defence, forestry, tech policy and global development looking for common threads and patterns that give us some idea of how we should be making our long-term plans and decisions. My hope is to both be informative about current best practice in long-term planning but also to give a sense of where I am coming from as a practitioner thinking about the long-term.

Section B is applied. There are of course differences between how a UK government policy official will think about the long-term, and how longtermists might think about the long-term. I take some of the ideas described in Section A and try applying them to some longtermist questions. I don’t have all the answers but I hope to suggest areas for future research and exploration.

Section C is constructive. I reflect on how my experience as a practitioner of longtermism shapes my view of the academic longtermist community. I then make some recommendations about how longtermists can better produce useful practical research.

Section A: Welcome to my world, let me show you around

Imagine that you are a politician or policy maker. You believe that the future matters a lot and that preventing existential risk is important, but you are uncertain about how best to achieve long-term goals. So for a starting point you look for existing examples of good long-term planning and long-term decision making.

At first good examples of long-term policy thinking can be hard to spot. Political incentives that push policy-makers towards short-term plans [1] or towards making long-term decisions primarily based on ideology [2]. There are however places where there is seemingly good long-term policy making to learn from, especially a step away from the most politicised topics. And if we look across enough institutions we start to get a picture of a best practice approach to long-term thinking.

Now I don’t want to claim that current best practice represents the only way to do long term thinking. But I do think it makes sense to set common sense (or best practice) as a prior and not to differ too much from it without a good reason. (I discuss this more...

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. Published on October 10, 2021.

Over the last two years I have been researching and advising numerous government officials on how to do long-term thinking well. Now you are probably not going to be shocked to hear me say it but: making long-term decisions is difficult. Most institutions don’t do it very well and the feedback loops to tell us what works are, as you would expect, long. Yet, that said, it is neither a new challenge nor an uncommon challenge. Many groups of people have faced this problem before and developed tools, strategies and approaches that seem to be working. I have been trying to pull all of this together to paint a rough picture of what best practice in long-term thinking looks like, and advise governments accordingly.

To do that work well I did of course engage in some depth with relevant academic work, including the research on longtermism both from within academia and on this forum. And lo and behold it seemed to me like the space was divided into two distinct camps: longtermist theorists and long-term practitioners. The theorists wonder why policy makers do not listen to them and the policy practitioners wonder why academics are not producing work relevant to them. As a practitioner, it seemed that on some days I would say something that was obvious to me and a researcher would be excited by how novel and useful it is, yet on other days I just could not understand the things longtermist researchers were doing and why it mattered. This post is an attempt to bridge this divide.

The post is in three parts:

Section A is descriptive. I invite you to look around my world, at the politicians, policy makers and risk planners who think long-term. I draw examples from fields as diverse as defence, forestry, tech policy and global development looking for common threads and patterns that give us some idea of how we should be making our long-term plans and decisions. My hope is to both be informative about current best practice in long-term planning but also to give a sense of where I am coming from as a practitioner thinking about the long-term.

Section B is applied. There are of course differences between how a UK government policy official will think about the long-term, and how longtermists might think about the long-term. I take some of the ideas described in Section A and try applying them to some longtermist questions. I don’t have all the answers but I hope to suggest areas for future research and exploration.

Section C is constructive. I reflect on how my experience as a practitioner of longtermism shapes my view of the academic longtermist community. I then make some recommendations about how longtermists can better produce useful practical research.

Section A: Welcome to my world, let me show you around

Imagine that you are a politician or policy maker. You believe that the future matters a lot and that preventing existential risk is important, but you are uncertain about how best to achieve long-term goals. So for a starting point you look for existing examples of good long-term planning and long-term decision making.

At first good examples of long-term policy thinking can be hard to spot. Political incentives that push policy-makers towards short-term plans [1] or towards making long-term decisions primarily based on ideology [2]. There are however places where there is seemingly good long-term policy making to learn from, especially a step away from the most politicised topics. And if we look across enough institutions we start to get a picture of a best practice approach to long-term thinking.

Now I don’t want to claim that current best practice represents the only way to do long term thinking. But I do think it makes sense to set common sense (or best practice) as a prior and not to differ too much from it without a good reason. (I discuss this more...

If humans stopped imitating, civilization would vanish.

It is 6:37 in the morning as I’m sitting down to write this, late in August, and I and the three-year-old are down in the old stable we use as study. As I look over my notes, she observes me. Mist is leaking through a crack in the window. Then she gets up, finds a pocket calculator, rummages around with the notes on the table a bit, and starts typing.

“The unicorn ate the human beings”, she says aloud. “Like wheat.”

Human beings cannot help picking up skills if they live immersed in communities where the skills are practiced.

If that statement is correct, or at least a good first approximation of the truth, as I believe it to be, the aim of education is to make sure as many people as possible can immerse themselves in skillful communities. We should, as far as possible, extract learners out of schools – where the incompetent live surrounded by, and therefore imitate, others at the same level of incompetence – and into productive environments.

Historically, this approach to learning has been the norm. Our grandparents lived in direct contact with the environments they would have to navigate as adults, and they could observe and imitate those who had already mastered it. But as modernity, working like a giant centrifuge, started separating everything that in traditional societies co-mingled, it became impossible for children to learn through pure imitation. They were segregated by age groups and denied access to the offices, factories, workshops, stores, whose skills they would have to master to become successful adults.

But maybe we can return to our premodern condition? Maybe we can build systems that allow novices to reintegrate themselves into communities of practice - but at scale?

It is a question of architecture, and can be phrased as such: How do we design workspaces where novices can enter at scale, and where they can easily observe the knowledge that guides experienced practitioners? And how can we do that without obstructing the work being done?

Historically, the rooms we’ve been able to build have been too limited for this. They put a low upper bound on how many novices could be absorbed by productive communities. Peter Paul Rubens, while developing his baroque masterpieces in the early 1600s, could only allow a handful of apprentices in his atelier. How would it have looked with hundreds! The demands of the incompetent would have overwhelmed Rubens, making work impossible, turning his workshop into a school!

In scaling access to valued knowledge, states, starting in the 18th century, opted for specialized learning environments. Fenced-in houses where novices lived segregated from the environments they were trying to master – called schools. By severing the bond between novices and masters, schools allowed us to increase the spread of select knowledge.

But schools have frustratingly slow rates of knowledge transfer and rapid learning loss. And they are marked by various psychological issues: disorders, boredom. It seems sitting in long rows of benches studying books is a bad fit for a species with learning instincts attuned to experiential learning.

We could have apprenticeships, which were effective but didn’t scale, or we could have schools, which were ineffective but scaleable. We couldn’t have both. This was the trade-off we faced. We choose school.

But the constraints that forced that trade-off on us are being overcome. We no longer need to separate novices from the environments they are to master. At this point in our technological development, it is possible to architect workspaces where novices can learn by doing, and more importantly by doing in the presence of competent practitioners. We can construct digitally mediated rooms where experts can coexist with large numbers of novices, allowing for a return to premodern forms of knowledge transfer. In this essay, I will explore how that would look.

Humans are Culture-Seeking

If you measure two-and-a-hal...

If humans stopped imitating, civilization would vanish.

It is 6:37 in the morning as I’m sitting down to write this, late in August, and I and the three-year-old are down in the old stable we use as study. As I look over my notes, she observes me. Mist is leaking through a crack in the window. Then she gets up, finds a pocket calculator, rummages around with the notes on the table a bit, and starts typing.

“The unicorn ate the human beings”, she says aloud. “Like wheat.”

Human beings cannot help picking up skills if they live immersed in communities where the skills are practiced.

If that statement is correct, or at least a good first approximation of the truth, as I believe it to be, the aim of education is to make sure as many people as possible can immerse themselves in skillful communities. We should, as far as possible, extract learners out of schools – where the incompetent live surrounded by, and therefore imitate, others at the same level of incompetence – and into productive environments.

Historically, this approach to learning has been the norm. Our grandparents lived in direct contact with the environments they would have to navigate as adults, and they could observe and imitate those who had already mastered it. But as modernity, working like a giant centrifuge, started separating everything that in traditional societies co-mingled, it became impossible for children to learn through pure imitation. They were segregated by age groups and denied access to the offices, factories, workshops, stores, whose skills they would have to master to become successful adults.

But maybe we can return to our premodern condition? Maybe we can build systems that allow novices to reintegrate themselves into communities of practice - but at scale?

It is a question of architecture, and can be phrased as such: How do we design workspaces where novices can enter at scale, and where they can easily observe the knowledge that guides experienced practitioners? And how can we do that without obstructing the work being done?

Historically, the rooms we’ve been able to build have been too limited for this. They put a low upper bound on how many novices could be absorbed by productive communities. Peter Paul Rubens, while developing his baroque masterpieces in the early 1600s, could only allow a handful of apprentices in his atelier. How would it have looked with hundreds! The demands of the incompetent would have overwhelmed Rubens, making work impossible, turning his workshop into a school!

In scaling access to valued knowledge, states, starting in the 18th century, opted for specialized learning environments. Fenced-in houses where novices lived segregated from the environments they were trying to master – called schools. By severing the bond between novices and masters, schools allowed us to increase the spread of select knowledge.

But schools have frustratingly slow rates of knowledge transfer and rapid learning loss. And they are marked by various psychological issues: disorders, boredom. It seems sitting in long rows of benches studying books is a bad fit for a species with learning instincts attuned to experiential learning.

We could have apprenticeships, which were effective but didn’t scale, or we could have schools, which were ineffective but scaleable. We couldn’t have both. This was the trade-off we faced. We choose school.

But the constraints that forced that trade-off on us are being overcome. We no longer need to separate novices from the environments they are to master. At this point in our technological development, it is possible to architect workspaces where novices can learn by doing, and more importantly by doing in the presence of competent practitioners. We can construct digitally mediated rooms where experts can coexist with large numbers of novices, allowing for a return to premodern forms of knowledge transfer. In this essay, I will explore how that would look.

Humans are Culture-Seeking

If you measure two-and-a-hal...