Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Why Doesn’t Healthcare Improve Health?, published by George on February 16, 2022 on LessWrong.

It’s almost a non-sequitur to ask whether or not a giant industry, with a well-defined goal, churning double-digits of GDP, actually works.

If pressured into an experiment we could easily design one that shows airplanes deliver people to destinations over 3000km away faster than trains (boarding time and travel to airport included). We could similarly prove they help people travel more, and more money spent on airplanes means more travel. Take a random group of 1000 people, give half 5000$ airline credits for two years, and we’ll certainly find that, on average they end up visiting more far-away places than the control.

On the other hand, this is not entirely unheard of.

A big and controversial example is the church. Where we’ve pretty conclusively proven that building huge cathedrals, converting “savage natives”, stomping out heresy and prayer have little results upon any form of well-being. But even here we must admit that, if the apparatus of science is unleashed upon religion, we will find interesting things such as people in religious communities drinking less, having longer-lasting marriages, and living longer with higher self-reported life satisfaction. We can argue about confounders and about the causal mechanism being replicable with things other than religion. But I think that, to most scientifically minded people, the lack of evidence is sufficient to not motivate them into becoming religious.

However, I keep being surprised by intelligent and scientifically minded people outright refusing to hear the evidence against medicine being useful. That’s not to say evidence against any particular therapy. There are many things that obviously work, such as emergency intervention for losing a limb in a car crash, where “almost certain death” is the clear outcome. There are many things that obviously work from a statistical perspective, such as vaccination against diseases such as covid, hep A & B, polio, HPV, yellow fever, rabies, and the like. There are also many things that seem to work, with good mechanistic evidence and promising studies, such as PRP for various forms of soft-tissue injury, or prophylactic therapy for HIV.

But this therapy-specific evidence is only more damming when viewed in the context of broader findings against medicine. Robin Hanson has written on this ad-nauseam, so I will let him do most of the talking.

1974 to 1982 the US government spent $50 million to randomly assign 7700 people in six US cities to three to five years each of either free or not free medicine, provided by the same set of doctors. . people randomly given free medicine in the late 1970s consumed 30-40% more medical services, paid one more “restricted activity day” per year to deal with the medical system, but were not noticeably healthier! (More, see also)

and

Oregon assigned a limited number of available Medicaid slots by lottery. . 8,704 (~30%) [very sick and poor US adults] were enrolled in Medicaid medical insurance. . at most see two years worth of data. . had substantially and significantly better self-reported health. . over two thirds of the health gains . appeared on the very first survey, done before lottery winners got additional medical treatment. (More)

No statistically significant effect on measures of blood pressure, cholesterol, or blood sugar. . did not reduce the predicted risk of a cardiovascular event within ten years and did not significantly change the probability that a person was a smoker or obese. . it reduced observed rates of depression by 30 percent. (More)

and

This study . is amongst the largest health insurance experiments ever conducted . in Karnataka, which spans south to central India. The sample included 10,879 households (comprising 52,292 mem...

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: REPL's and ELK, published by scottviteri on February 17, 2022 on The AI Alignment Forum.

In my previous post I talked about read-eval-print loops as providing a type signature for agents. I will now explain how you can quickly transition from this framework to an ELK solution. Notation is imported from that post.

Imagine we have two agents, a human and a strong AI, denoted H and M respectively. They both interact with the environment in lockstep, according to the following diagram.

We have the human's utility function UH:SH→Q, which is defined on the human's model of reality. We would like to lift UH to a version UM:SM→Q that the machine can use to influence the world in way that is agreeable to the human, which we can do by learning a mapping F : SM→SH and deriving UM=F∘UH.

But we haven't yet said what properties we want the ontology map F to have. I want to call two concepts sh and sm equal if they act the same with respect to transformation: ∀f. f(sh) = f(sm) → sh = sm. The issue is that since the concepts have different types we cannot feed them as arguments to the same function. So instead let's say that ∀s:S, EvalH(sh, ReadH(s)) = EvalM(sm, ReadM(s)) → sh = sm. But now we are back to the same problem where we are trying to compare concepts in two different ontologies. But this does give us a kind of inductive step where we can transfer evidence of equality between concept pairs (sh, sm) and (sh', sm'). I also believe that this kind of a coherence argument is the best we can do, since we are not allowed to peer into the semantic content of particular machine or human states when constructing the ontology map.

Consider the following two graphs.

My intuition is that even if I don't know the labels of the above graphs, I can still infer that the bottom nodes correspond to each other. And the arrows that I get in the context of ELK are the agents' Eval transitions, leading to the following commutative diagram specification for F.

We can learn an ontology map F : SM→SH by minimizing the difference between two paths from a state sm, one in which the machine's prediction function is used and one in which the human's prediction function is used. Concretely, I propose minimizing Dist(sh1,sh2) + λ |U(sh1)-U(sh2)| where sh1 = F(EvalM(sm,om)) and sh2 = EvalH(F(sm),oh), Dist is a distance metric in SH, and observations om and oh are generated by the same underlying state S.

If you are interested in getting more detail and why I believe this circumvents existing counterexamples, please check out the full proposal.

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: Announcing Alvea—An EA COVID Vaccine Project, published by kyle fish on February 16, 2022 on The Effective Altruism Forum.

We’ve had effective COVID vaccines for more than a year, but there are still countries where less than 10% of people have received a dose. Omicron has been spreading for almost three months, but pharma companies have only just started testing variant-specific shots. mRNA vaccines are highly effective, but they’re hard to manufacture and nearly impossible to distribute in parts of the developing world.

We won’t be ready for the next variant, or the next pandemic, until these problems are solved.

An ideal vaccine platform for pandemic preparedness would be fast, effective, cheap, robust, and scalable enough to reach a critical portion of the global population as soon as any new pathogen is identified. Ideas have been floating around the EA biosecurity community for years about what such a platform would look like, and what it would take to build it, but there hasn’t been much direct work in the space.

The most recent COVID wave convinced a few of us that it was time for that to change, so we launched Alvea to rapidly develop and deploy an Omicron-specific vaccine. In doing so, we’re aiming to build a platform for responding to future COVID variants or other novel pathogens, and to massively level up the technical, operational, clinical, organizational, and logistical capabilities of the longtermist biosecurity community.

We’ve initially structured Alvea as a three-month sprint to test the hypothesis that an exceptionally bright, dedicated group of people can quickly accomplish remarkable things in this space. In the past eight weeks, we’ve built a team of 35 drug developers, logistics experts, physicians, operators, and scientists to bring the project to fruition. We’re supported by a network of consultants and partners with deep experience in every aspect of vaccine development and infectious disease response.

Many of us are longtermist EAs who believe this project has a shot at transformative impact on biosecurity. All of us are committed to radically improving vaccine development, and are working around the clock to make this happen.

Our strategy at Alvea is simple: We’re building a streamlined platform for developing and deploying DNA vaccines using safe, well-validated components. DNA vaccines function similarly to mRNA vaccines, but are easier to produce and can be shipped around the world with no special storage requirements. Ultimately, we expect people to be able to easily self-administer our vaccines anywhere on the planet.

Getting our vaccines into the clinic—and out to the places where they’re needed—requires solving a staggering number of problems in preclinical development, manufacturing, international regulation, clinical trial design, logistics, and other areas, all at the same time. In many cases, the standard approaches to these problems are riddled with crippling inefficiencies. We’re eliminating those when we can, and building from scratch when we can’t.

In the 60 days since our inception, we’ve designed twelve versions of our Omicron vaccine, responded to the emergence of the BA.2 subvariant, produced hundreds of doses of our lead candidate, run preclinical experiments in mice and sheep, kicked off scalable manufacturing processes, planned Phase I and II clinical studies, and identified potential partner countries for accelerated trials. There’s an enormous amount of work still to be done, but we are well on our way.

Alvea is led by Ethan Alley and Grigory Khimulya (Co-CEOs), Cate Hall, and Kyle Fish. Our team is growing rapidly, and we’re particularly keen to expand in the following areas:

Wet laboratory (molecular biology, in vitro and in vivo development)

Clinical trial operations and logistics

General company operations

cGM...

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: Compute Trends Across Three eras of Machine Learning, published by Jaime Sevilla on February 16, 2022 on The AI Alignment Forum.

What do you need to develop advanced Machine Learning systems? Leading companies don’t know. But they are very interested in figuring it out. They dream of replacing all these pesky workers with reliable machines who take no leave and have no morale issues.

So when they heard that throwing processing power at the problem might get you far along the way, they did not sit idly on their GPUs. But, how fast is their demand for compute growing? And is the progress regular?

Enter us. We have obsessively analyzed trends in the amount of compute spent training milestone Machine Learning models.

Our analysis shows that:

Before the Deep Learning era, training compute approximately followed Moore’s law, doubling every ≈20 months.

The Deep Learning era starts somewhere between 2010 and 2012. After that, doubling time speeds up to ≈5-6 months.

Arguably, between 2015 and 2016 a separate trend of large-scale models emerged, with massive training runs sponsored by large corporations. During this trend, the amount of training compute is 2 to 3 orders of magnitude (OOMs) bigger than systems following the Deep Learning era trend. However, the growth of compute in large-scale models seems slower, with a doubling time of ≈10 months.

Not enough for you? Here are some fresh takeaways:

Trends in compute are slower than previously reported! But they are still ongoing. I’d say slow and steady, but the rate of growth is blazingly fast, still doubling every 6 months. This probably means that you should double the timelines for all previous analyses that relied on AI and Compute’s previous result.

We think the framing of the three eras of ML is very helpful! Remember, we are suggesting to split the history of ML into the Pre-Deep Learning Era, the Deep Learning Era and the Large-Scale Era. And we think this framing can help you make sense of what has happened in the last two decades of ML research.

We have curated an awesome public database of milestone ML models! Please use it for your own analyses (don’t forget to cite us!). If you want to play around with the data, we are maintaining an interactive visualization of it here.

Compute is a strategic resource for developing advanced ML models. Better understanding the progress of our compute capabilities will help us better navigate the advent of transformative AI.

In the future, we will also be looking at the other key resource for training machine learning models: data. Stay tuned for more!

Read the full paper now on the arXiv

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: Elon Musk donates $5.7 Billion, will work with Igor Kurganov [Crosspost], published by Nathan Young on February 16, 2022 on The Effective Altruism Forum.

Bloomberg reports the following - without permission"Elon Musk, it seems, is finally becoming a philanthropist on the scale of his billionaire peers.

The electric car and space mogul gifted $5.7 billion worth of Tesla Inc. stock to charity in the span of 10 days in November -- many times more than he’s given away through his eponymous foundation in the two decades since it was founded. Where that donation is going is a mystery, but it’s just one more signal that the world’s richest person is taking philanthropy more seriously.

The decision by Musk, 50, to donate more than 5 million shares in the electric-car maker was disclosed in a regulatory filing Monday night, and comes on the heels of some of his biggest-ever philanthropic commitments -- though nothing has come close to the scale of billions of dollars. It would also help reduce what he called the biggest tax bill in U.S. history.

The Musk Foundation in the past couple of years has made eight-figure grants to the city and local school system near his South Texas spaceport, a $100 million ready-made competition to fight climate change and millions of dollars to a pair of Covid-19 researchers.

Musk at the construction site for a new Tesla Gigafactory near Gruenheide, Germany in 2020.

Almost all of those recipients have been primarily working with Igor Kurganov, a professional poker player-turned-philanthropist, who Musk has recently enlisted to keep in contact with grantees and consider their proposals.

Effective Altruism

Kurganov, who has won more than $18 million during his poker career, is active in the world of so-called effective altruism, a philanthropic and philosophical movement that tries to have the greatest impact by carefully spending money to solve problems. Kurganov is the co-founder of Raising for Effective Giving, an organization created by a group of poker players that recommends “highly cost-effective charities.” He’s also an adviser to the Forethought Foundation, a project by the Centre for Effective Altruism.

Calls and emails to Kurganov haven’t been returned. Jared Birchall, the managing director of Musk’s family office, and Aaron Beckman, Tesla’s senior manager of stock administration, didn’t respond to requests for comment.

If Musk’s foundation is turning to effective altruism, it hasn’t showed yet, said Alixandra Barasch, an associate professor of marketing at New York University who has done research on the movement.

Musk doesn’t regularly publicize his gifts and tax documents that provide such disclosures have a years-long delay. What Musk has shared recently includes $10 million to Brownsville, Texas, near where SpaceX is located, to revitalize its downtown. So far, that money has gone to a grants program for local property owners, lighting projects and murals.

Scale Up

Saving lives is the primary focus of effective altruism, and “murals don’t do that much good in terms of the way it’s measured and defined in the movement,” Barasch said.

Other gifts from Musk, like the $20 million for Cameron County schools, the $100 million commitment for the XPrize Carbon Removal competition and $5 million for Khan Academy, similarly don’t have the focus that effective altruism typically does, Barasch said. But scale is also a factor, she said, and if Musk’s foundation is influenced by the philosophy, his giving could increase significantly.

“Looking at those amounts I’m like, ‘Holy moly, there’s millions of dollars, it’s a lot of money,’” Barasch said. But “it’s nothing compared to what he’s worth.”

Musk has a $227 billion personal fortune, according to the Bloomberg Billionaires Index. The donation of $5.7 billion of Tesla stock could be a sign that...

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] 4. The “short-term predictor”, published by Steve Byrnes on February 16, 2022 on The AI Alignment Forum.

4.1 Post summary / Table of contents

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

The previous two posts (#2, #3) presented a big picture of the brain, consisting of a Steering Subsystem (brainstem and hypothalamus) and Learning Subsystem (everything else), with the latter “learning from scratch” in a particular sense defined in Post #2.

I suggested that our explicit goals (e.g. “I want to be an astronaut!”) emerge from an interaction between these two subsystems, and that understanding that interaction is critical if we want to assess how to sculpt the motivations of a brain-like AGI, so that it winds up trying to do things that we want it to be trying to do, and thus avoid the kinds of catastrophic accidents I discussed in Post #1.

These next three posts (#4–#6) are working our way up to that story. This post provides an ingredient that we’ll need: “the short-term predictor”.

Short-term prediction is one of the things the Learning Subsystem does—I’ll talk about others in future posts. A short-term predictor has a supervisory signal (a.k.a. “ground truth”) from somewhere, and then uses a learning algorithm to build a predictive model that anticipates that signal a short time (e.g. a fraction of a second) in the future.

This post will be a general discussion of how short-term predictors work and why they’re important. They will turn out to be a key building block of motivation and reinforcement learning, as we’ll see in the subsequent two posts.

Teaser for the next couple posts: The next post (#5) will discuss how a certain kind of closed-loop circuit wrapped around a short-term predictor turns it into a “long-term predictor”, which has connections to the temporal difference (TD) learning algorithm. I will argue that the brain has a large number of these long-term predictors, built out of telencephalon-brainstem loops, one of which is akin to the “critic” part of actor-critic reinforcement learning. The “actor” part is the subject of Post #6.

Table of contents:

Section 4.2 gives a motivating example, of flinching just before getting hit in the face. This can be formulated as a supervised learning problem, in the sense that there is a ground-truth signal to learn from. (If you just got hit in the face, then you should have flinched!) The resulting circuit is what I call a “short-term predictor”.

Section 4.3 defines terminology: “context signals”, “output signals”, and “supervisory signals”. (In ML terminology, these correspond respectively to “trained model inputs”, “trained model outputs”, and “labels”.)

Section 4.4 offers a sketch of an extremely simple short-term predictor that could be built out of biological neurons, just so you can have something concrete in mind.

Section 4.5 discusses the benefits of short-term predictors compared to alternative approaches including (in the flinching example) a hardwired circuit for deciding when to flinch, or a reinforcement learning (RL) agent that is rewarded for appropriate flinching. For the latter, a short-term predictor can learn faster than an RL agent because it gets an error gradient “for free” each query—or in simpler terms, when it screws up, it gets some indication of what it did wrong, e.g. whether the error is an overshoot vs. undershoot.

Sections 4.6-4.8 cover various examples of short-term predictors in the human brain. None of these are especially important for AGI safety—the really important one is the topic of the next post—but they come up sufficiently often that they warrant a brief discussion:

Section 4.6 covers the cerebellum, with my theory that it's a collection of ≈300,000 short-term predictors, used to (in effect) reduce the latency on ≈300,000 signals...

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: Should there be a BERI for animal welfare or global poverty?, published by sawyer on February 14, 2022 on The Effective Altruism Forum.

The Berkeley Existential Risk Initiative provides operational and administrative support to researchers working on existential risks in a university setting. The idea is to make operations faster and more flexible for these groups—not only to make them more directly effective, but also to improve morale by unblocking tasks and projects they care about that are hard to do efficiently through other means (e.g. existing university administration channels).

Some specific examples: Hiring a software engineer, graphic designer, web developer, or copy editor; hiring a research assistant for month on 2 days notice; paying monthly software subscriptions; paying international contractors; buying stuff from a supplier that your university doesn't support.

How beneficial would this same service be to researchers working on global poverty or animal welfare (or other cause areas for that matter) in a university setting?

I'll add some of my preliminary arguments for and against 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: Paper is True, published by Zvi on February 14, 2022 on LessWrong.

Postscript to: Rock is Strong

Always choosing Rock, broadly construed, is freerolling on the efforts of others and/or you are lying. You are in at some important sense defecting. The problem of an information cascade is real. You are not contributing new true information, instead reinforcing the information and heuristics already out there.

Often this is the most accurate option available to you, or at least the most efficient. Doing the work yourself is hard. Many can’t do better that way and even for those who can it is a lot of work. On average those who disagree are more wrong and end up doing worse. Other times, accuracy is not what is desired. None of that makes it easy to pick the correct rock, but often it is straightforward enough.

The defection that comes from not doing your part for the long term epistemic commons often is made up for not only by your lack of a practical option to do otherwise, but also in some cases by the gains from accuracy.

Rewarding those who worship the rock, like rewarding any form of defection, is also defecting. The Queen causes everyone to die in a volcanic eruption because she set up bad incentives and is encouraging everyone to spend down the commons.

This can also be thought of partially as explore versus exploit. The majority of actions should be exploit, and most people should mostly be exploiting. There is more variance in ability to explore than exploit, so most people should explore even less than average and let others explore and report back. Exploration is a public good, so it needs to be rewarded beyond its private payoff or there won’t be enough of it. When competition is sufficiently intense or other factors reward exploitation too much, exploration can die out entirely.

This is the context for a Marginal Revolution post that nagged at me enough to give me enough initial motivation to write this whole thing. Quoted in full, and this is not about the examples it is about the method:

No, you don’t always have to agree with the majority of the educated people, but I would say this. For whatever set of views you think is justified, try to stick to the versions of those views held by well-educated, reasonable, analytically-inclined people. You will end up smarter over time, and in better places. Peer effects are strong, including across your ideological partners.

When I hear that a particular group defends liberty, such as the Ottawa truckers’ convoy, while this is partially true it makes me nervous. As a whole, they also seem to believe a lot of nonsense and to be, in procedural terms, not exactly where I would want them on scientific method and the like. Fair numbers of them seem to hold offensive beliefs as well. Whine about The Guardian if you like, but I haven’t seen any rebuttal of this portrait of the views of their leaders. Ugh.

I recall taking a lot of heat for my 2007 critique of Ron Paul and his movement, but that example illustrates my points perfectly. Those people did defend liberty in a variety of relevant ways, but so many of them have ended up in worse spaces. And that is exactly what I predicted way back when.

Look for strong analytical abilities, and if you don’t see it, run the other way.

Here is a defense of the Freedom Convoy. You can read it for yourself, but it doesn’t change my mind. Here is I think a wiser account. I’ll say it again: “Look for strong analytical abilities, and if you don’t see it, run the other way.” I’m running.

When I saw this, it set off loud alarm bells in my head. Something felt deeply wrong.

Essentially it was saying that if the Wrong Kind of People, who expressed Wrong Views that oppose Narrative, were advocating or backing something, you needed to run away even if they were right this time. Either the Covi...

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: Observations about writing and commenting on the internet, published by dynomight on February 15, 2022 on LessWrong.

I’m not famous or successful, so why should you care what I think? Well, I have some observations about the dynamics of writing on the internet that I think my (even more non-famous and non-successful) self would have benefited from when I started.

Human experience is vast.

The whole idea of writing is crazy: You have a pattern in your brain-meat, which you try to encode it into a linear series of words. Then someone else reads those words and tries to reconstruct the pattern in their brain-meat. But in this dance, how much of the work is being done by the words versus the lifetime of associations each person has built up around them?

Rather than a full blueprint for an idea, writing is often more like saying “Hey, look at concept #23827! Now look at concept #821! Now look at concept #112234! Are your neurons tingling in the way mine are? I hope so because there’s no way to check, bye!”

We have different personalities and spend our lives getting exposed to different information and thinking about different things. What concept #821 triggers for you may be vastly different than what it triggers for me.

I suspect that even when writing works, readers are often taking quite a “different trip” than the writer intended. Personally, I figure that’s fine and it’s better to just let people take their own trip, rather than going to insane lengths in a hopeless quest to make everything precise.

So: No matter what you do, sometimes your writing will fail. It’s impossible to predict all the ways it will fail. Really, it’s amazing that it works at all. But still, it’s possible to reduce the frequency of failure.

Failures can seem baffling.

Here are two examples of how things I’ve written have failed:

I wrote an article suggesting ultrasonic humidifiers might put particulates into the air and harm health. The median response was something like this:"Please stop polluting the internet with speculation. If you can’t support your argument with peer-reviewed research papers you shouldn’t write it at all."This was. puzzling, because I had dozens of citations starting very early in the article. But many people in different forums independently left comments like this.

I wrote an article about gas stoves and nitrogen dioxide where I stated in the overview that a normal range hood might not solve the problem. A top comment was that a high-quality hood would solve the problem and the fact that I claimed this without providing any data showed I was biased and bad (so bad).That’s fair, although after the overview finished there was a link to the “Can a range hood fix this?” section where I discuss an article that tested hoods in various houses and found that some worked well but most of them reduced nitrogen dioxide by 30% or less, and wait did I say that was fair?

Please holster your tiny violin, I know this is all normal. Many people get fed up with stuff like this and resolve to ignore all comments. That’s not my position. My position is that I failed and when I fail I’d like to know about it.

First, though, why does this happen?

What happens in forums isn’t always about you.

Why do people go to forums? Partly for links, yes, but also because they like the community. When an article on an interesting topic comes up, some people decide they’d rather see what people they trust think before investing time reading the article. And once they are there, maybe they see a comment they want to respond to.

I’d love to see statistics for this, but I’d guess that only a small fraction of people finish most articles before commenting, and many don’t even look at the article. This is not necessarily a bad thing! Sometimes the discussion detaches and goes into all sorts of interesting and...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The Territory, published by LoganStrohl on February 15, 2022 on LessWrong.

I don’t know if you’ve ever had the experience of hiking in the wilderness with a map and compass but no cell service. I recommend it, if you haven’t.

I somehow did not quite all-the-way understand what a map even is until I was lost on my own under these circumstances. I knew in a “factual knowledge” sense that maps were drawings of the land, and I’d even used them as a kid and teenager to help my family navigate on road trips. But when I was lost in a national park, trying to find my way back to my car, I confronted the incompleteness of my knowledge of maps. There was a shift.

My map had trail lines drawn on it, with labels like “Canyon Trail”. I’d pause my walking to look at the shape of “Canyon Trail”, noting that it intersected "Overlook Trail" somewhere off to the left of where I was standing. Then I would walk again—attempting, I think, to "follow Canyon Trail to Overlook Trail".

I would move back and forth between walking and map consultation, making sure I remembered which way the trails were supposed to go, constantly placing and replacing myself within the borders of the lines drawn on the paper. The more distressed I felt about being lost, the more often I turned to the map, looking for something to hold on to.

The shift happened after. (this is sort of embarrassing, it’s so simple. But it’s true.) The shift happened after, having oriented myself toward “North”, I happened to lower the map a little bit, probably out of exhaustion. I held it a bit below eye level, so that it was no longer taking up my whole field of vision.

I looked at the squiggly blue line on the map, and the close-together lines that I knew indicated steepness. And I saw to my left, because the map was not blocking my vision, a creek. Up ahead, I saw a steep hill.

I realized that the blue line was probably a drawing of that creek.

The contour lines were a drawing of that hill.

And then this wild rushing sensation began to wash over me. I was starting to get it. Slowly, I tilted the paper in my hands from a vertical position, partially blocking my view...

...to a horizontal position, parallel to the ground.

I held the map that way, looking out at the world the cartographer had tried to draw, and it was as though the territory rose up to meet the map, while the map spread itself across the surface of the territory. And I said to myself, “It’s a picture!”

For the first time, I understood in a practical way that a map is meant to be a top-down picture of the real world.

Before I had this realization, I wasn't behaving as though I knew myself to be in the territory, using the map as a tool. I was acting as if I were traversing the map, using my body as a kind of clunky video game controller. I had been treating the map as the terrain I "really" had to navigate.

But once I stopped playing that game, and started actually traversing the forest I was in, things went very differently. I spent most of my time looking at creeks and trees and hills, making sure I knew how the real world around me was shaped. And from that perspective, I looked down at the map to help me predict what I’d see next.

And I found my car shortly thereafter.

There are ways to increase some kinds of knowledge that largely involve staring at maps. Perhaps your own map is not clearly labeled in places, or it’s somehow inconsistent with itself, or it doesn’t match the map of an expert.

This is why it’s often valuable to clearly articulate your beliefs, even just to yourself. It’s valuable to ask yourself what you expect, and to notice when you feel confused about that. It’s valuable to ask other people what they think, or to read their books and blog posts, especially when you have reason to believe they know important things that you don’t.

But the ma...