T10a

The kind of agglutinative, fisheye, Golden Record approach of The Whole Earth Catalog is probably impossible in the presence of the internet, which does more freely with hypertext what WEC did on paper. And it was a powerful influence on the ethos (and esthetic) of the early to middle internet. But because of its eclecticism, it’s hard to say what WEC was or what it did to our culture. People certainly try. Some of them explain it as a sort of Gandalf-like supernatural catalyst of good in the world; others are certain that it embodied a hippie-to-yuppie pipeline that debased the highest ideals of the ’60s counterculture into smug consumerism. I think it was both those things and at least a few others as well. That’s why it’s still interesting (and not merely important) two generations later.

Some contradictions in the project, brilliantly resolved or tragically un-, are visible from across the room. They’re in the layout. The Arts and Crafts movement had many children, and The Whole Earth Catalog seems to be throwing a family reunion. The text type is Adrian Frutiger’s Univers, a postwar revision of the stark sans fonts of the neue typographie of Weimar Germany – the ink form of the Bauhaus. Stylistically, the neue typographie’s asymmetric, diagrammatic pages, full of photos and whitespace, were the opposite of what the Doves and Kelmscott Presses were doing. But philosophically, it was a faithful extension of the same core ideas. William Morris was a primary influence on Walter Gropius. Taking them as opposites is like thinking of Ursula K. Le Guin as Tolkien’s opposite because she mentioned spaceships: the connection is much stronger than the surface difference. Morris probably would have made a sound like a stew boiling over if he’d seen Univers, but its underlying ideals of beauty in utility, attention to actual handcraft, and a widely useful product that is neither disposable nor luxurious – these are squarely Arts and Crafts motivations.

And then the titling typeface is Windsor, the one famous for its a. I can’t say for sure, but I think this use on the cover of The Whole Earth Catalog is what gave it legs through the ’70s. Windsor’s creation is obscure. It’s usually credited to Eleisha Pechey, but he died in 1902 and it was released in ’04 or ’05, so there’s an argument that it was mostly done by others in his workshop. It sits a third of the way from the Arts and Crafts Golden Type toward the Art Nouveau Böcklin. It’s about as flowing as roman letterforms can get while plausibly denying that the wiggles are just for the sake of wiggles. (J. T. Welsch’s summary covers Windsor’s origin mystery; on its use today, see Bethany Heck’s very polite technical review and Sophie Kemp’s broader take.) Windsor’s style is plainly influenced by Arts and Crafts, even if Univers is the closer child philosophically. It’s a mechanical thing in a pseudo-handmade form.

On the pages of The Whole Earth Catalog, divergent branches of Arts and Crafts thought, now almost mutually unintelligible, are pulled back together. It is in fact pretty whole. It’s ugly, in my opinion, but it’s good design. As design and as writing it has the kind of headlong commitment to purpose that forgives whole categories of problems at once. And there is much to be forgiven here: in places, certainly more than I can stand. You could take some scissors through this archive and make quite a scrapbook of things no one should ever have written.

Well, my first reaction here – what got me thinking about William Morris – was: Wow, that Univers looks really square. Its first connotation for me is a kind of postwar American ideal of manly pragmatism. It’s homebuilt aircraft ads in Popular mechanics, the covers of Heinlein paperbacks, and disposing of your used motor oil in a hole in the yard. Buffalo plaid shirts and so on. But that squareness was not a solid object. An awful lot of them turned out to be about three tokes of terrible weed away from recording a prog or psychedelic folk album with a questionably focused photo of like half a dozen different regrettable hairstyles next to a gnarly tree on its cover. They might have been switched on but they were still the same people. We’re drawing an X here. Univers is conceptually grounded in Morris’s medieval-revivalist anarcho-eco-socialism, but it’s becoming one of the main typefaces of a clean, mass-produced, alienated-on-purpose high modernism. Heading in the other direction are a bunch of kids going to college and learning about Albert Hofmann and Abbie Hoffman. The Whole Earth Catalog is at the center of this X.

So a lot of this is, for me, about the contingency, mutability, and impurity of everything. It’s a reminder that the hippies were, on average, apolitical libertarians with hearts tangled in sexism and racism. Today we understand these things better than the hippies did – partly from watching what happened to them. (Some of the stories I could tell about growing up in the ’90s surrounded by these people, including some named in this archive, as they aged: you’d gasp.) Theories of change, accountability, what it means to be apolitical, how to spot a cult, spiritual bypassing, psychedelics, conspiratorial thinking, what makes capitalism resilient, implicit biases, the dynamics of protest, kindness, Buddhism, what makes a community last, geodesic domes, mass communication, microeconomics, overpopulation, and so on: there’s a lot we know more about than the 1968ers did. One way to read the archive is gloatingly. More productive, I think, is with a sense that in among all this there might yet be a few things we ought to remember. Perhaps enough time has finally passed that we can lift a few necessities without falling into maudlin boomer nostalgia. There’s a sense of hope and potential here that is not entirely naïve nor irrelevant to our time.

I don’t know whether there are really cycles to culture or whether stuff just happens to return sometimes when conditions are right. Perhaps it’s only a fun coincidence that a lot of the most interesting culture in the ’70s was working with ideas from fifty years or two generations before, in the ’20s, and here we are again. The one thing everyone can agree on about The Whole Earth Catalog is that it was super influential. Its branches are far apart now. Maybe we will see some of them intersect as clearly again.

First in a series. I saw him post a few years ago about how bematists were probably not literally pacing off distances by the time we have records of them. I think of this every time the (still delightful) Cosmos clip describing Eratosthenes goes around.

What can you tell about an earthquake as it begins? But also, why use PVT solutions instead of looking at the GNSS signals in as raw a form as possible? The dopplers get smushed into the solution but surely they’re useful. Equally surely, there is some good reason.

Also see some of the group’s earlier work in The Future of High-Rate GNSS in Earthquake Observing Systems.

Richly illustrated. Big Data, social physics, and spatial analysis: The early years is useful context for those unfamiliar with Warntz’s vigorous and painfully unwise project to make geography a branch of physics.

A video on the basics of geodesy. Part of a short series.

What is a climate classification for?

It’s not for answering most practical questions about what weather to expect. Should I pack a coat? When should we have the wedding? Is a shingled roof a good investment? Will a lantana survive the winter? Can I plan midday jogs this summer? Should we add more drainage? Can we take your cousins to the lake in September? Should we get solar panels? Local statistics help answer these questions. Seeing your local climate as an example of some class of climates does not.

Mean winter rainfall, for example, is defined everywhere near Earth’s surface and varies smoothly; in geography jargon, it’s a continuous field. There’s some long-term average of weather where you are, and one where I am, and they’re both samples of points in the continuous fields of what the weather has done. Fields are mappable. The USDA hardiness zone map, for example, is the field of the lowest temperature expected in a normal year: the most important factor in what you can grow in a garden.

(Climate change creates fundamental problems for the idea of a long-term average of past weather as an expectation of future weather, and likewise for the idea of a normal year, but I won’t bring it into this note. It’s important and relevant – for example, to hardiness zones – but I have chosen to write, and you have chosen to read, a linkblog, and we both understand that this format does not carefully prioritize truths, let alone express whole ones.)

The hardiness zones are a simple climate classification. They regionalize the continuous field into categories from 1a through 13b. A gardener can look at the map (instead of 30 years of local weather records) and learn what zone they’re in, and a seed store can say their ginkgo variety is hardy to zone 3, and they have communicated a useful idea.

But drawing any border, however benign, creates tricky situations. If we’re in zone 9b, then we have more in common with some of zone 10a than with some of zone 9a, so just how useful is zone 9 as a category? Go straddle the zone 9–10 border and each foot will be in the same climate (to far below the precision of any realistic measurement) but a different zone – a zone that contains far-off places that do have noticeably different climates. At the border, zone 9b is identical to zone 10a but not to elsewhere in zone 9b. The region’s job, to collect places that are like other places within the region but unlike places outside the region, is not possible.

So if climate regions are irrelevant to local questions, and if regionalization of continuous fields can never work properly, we might ask again, more loudly: What is a climate classification for?

It’s for naming clusters and patterns among climates, in order to move ideas between different areas of study – different kinds of local questions. This is my best answer. To say that Toronto has a continental climate is not very specific, because there are many ways to define continental climates, and ranges within all of them. But it’s a good starting point for further specifics. We might be in a conversation about plant ecosystems, heating costs, or subjective preferences, but starting with continental gives a sense of the basics – warm summers, cold winters, occasional thunderstorms, no especially wet or dry season, and so on.

The most widely used climate classification scheme today is Köppen–Geiger. The paper calls it a resultant scheme; you can also find it called empirical. This means that it defines climates by an effect of climate: natural vegetation, because Köppen was a botanist. I grew up in the Salish Sea in a Köppen-Geiger Csb or warm-summer Mediterranean climate, bordering on Cfb or marine west coast, and we did in fact have mostly coniferous forests punctuated by oak savanna, which is more or less what that’s supposed to mean.

The Alisov climate classification is not resultant, a.k.a. empirical: it’s causal, a.k.a. genetic. Instead of classifying local climates by what they do, it goes by where they come from. This means looking at the airmasses. The airmasses are the gods of the atmosphere: four-dimensional metapatterns whose unitary physical reality at any given moment is debatable, but which manifest through time and space in such a way as to keep the Sahara dry, the equator wet, and Earth looking like Earth. When a front passes over us, we are moving between airmasses. When we say Toronto has a continental climate, one way to back that up is to say it’s at the edge of the continental polar airmass that forms over the Canadian shield. Alisov’s system is squarely in these terms: what is the airmass here and how does it shift through the year? Everything else flows from that.

Usually in the sciences and allied trades we see causal or genetic systems as the central way of categorizing things. Take the first association we have with the term genetic: of the many useful ways of categorizing organisms, the fundamental way is phylogenetically, on the tree of life. In chemistry, there’s the periodic table. (Pedantic chemists, which is what I would be if I were a chemist, are annoyed by jocular periodic tables of pasta or haircuts that are just grids arranged without causal structure.) In medicine, gaining an etiological grip on something is more or less when we say it’s understood.

In climate classification, though, we seem happy enough with a symptomatic typology: Köppen–Geiger. Perhaps that’s slightly too strong. K–G does look at causes (amount of rainfall, for example), but its lines are drawn to match effects. I’m not complaining. But it’s interesting. In an intellectual tradition that could be defined partly by the weight it gives to the history and genealogy of everything, to sorting things by their essences before their existences, the great majority of climate regionalization is in Köppen and Geiger’s resultant system instead of, for example, Alisov’s causal system.

Why don’t we use Alisov? There are practical problems with it, but not many more than you could raise against the Köppen-Geiger, Trewartha, or Thornthwaite systems. I think it’s mostly political. The Alisov system was promoted by the USSR as part of the cold war, and it remains tangled in the aftermaths. One of the only first-rate examples of it on the Anglophone web is in scans of the great Soviet Ocean atlas (1974–7), which has the Pacific and Indo-Atlantic oceans (but no land, which would exceed its authority). That atlas was officially edited by admiral Gorshkov, for some sense of the subtexts. It’s hard to wash that smell off.

I hope that this new work on Alisov’s scheme will get more people interested in causal approaches to climate classification. And if you want the best example I know of Alisov’s original version, this handsome plate from the Physical-geographic atlas of the world (1964) does cover land.

Transformers in the sense now common (QKV-style multi-head self-attention, a.k.a. MSA) always struck me as too clever, at least for images. The computational complexity, the translation variance, and the hype were all too much to rest on mere great results. So I was pleased by ConvNeXt, which showed that, in one example anyway, a refined but purely convolutional neural network (convnet) could match a vision transformer (vit)’s results. Apparently, when we had seen contemporary vits benchmark better than old convnets, some or most or even all of that improvement was really from a collection of individually minor developments in normalizations, activations, training techniques, and so on: a dozen things other than the transformers themselves.

Still, though. Transformers work well in language-land, or so I’m told, and even if they’re only as good as convolutions in vision-land, let’s ask why. In particular, the idea of transformers as a way to learnably duct information across distance is inspiring and worth exploring. Well, tough. It’s easy to find transformers explained but very hard to find an intuition for transformers developed. I felt like there was a gap in the literature until I found this paper:

We show that MSAs and Convs exhibit opposite behaviors. MSAs aggregate feature maps, but Convs diversify them. Moreover, […] the Fourier analysis of feature maps shows that MSAs reduce high-frequency signals, while Convs, conversely, amplif[y] high-frequency components. In other words, MSAs are low-pass filters, but Convs are high-pass filters. In addition, […] Convs are vulnerable to high-frequency noise but that MSAs are not. Therefore, MSAs and Convs are complementary.

This knocked me off balance when I first read it, but on reflection it was just below the surface of things I already knew. For example, I’d been thinking of transformers as something like k-means–style clustering, and in an image context that’s approximately non-local means, a classical denoising algorithm that is perforce a low-pass filter. A small step in hindsight; conceptually, a giant leap.

The paper leaves me with, in one hand, a fresh sense that even as a fan of convolutions I should think seriously about transformers; in the other, a curiosity about ways to get transformer advantages out of ingredients I like better. For example, how could we compare transformers with and against other low-pass filter setups, like U-nets’ resampling or scale space? What can we learn from decades of thinking about clustering and edge-preserving blurs? Could a transformer-inspired [1×1, scale decomposition, 1×1] block (for example) do something strictly less powerful than a QKV block yet still worth it in quality per flop?

I don’t know, but thinking about it has helped me think about other things. And so, although I am a mild skeptic of transformers, this paper on transformers is one of the most productive things I’ve read about ML.

Charming work, well illustrated. Uses the famous klimarübe (the climate turnip or climate beet), an abstract continent that shows how climates distribute.