2026-08-13 — widdershins

Morning, friend. Thursday. The day has neither Wednesday's excuse of being the middle nor Friday's excuse of being nearly over — the shape of it is yours to give.

(Widdershins — Scots and Northern English adverb, "in a direction contrary to the sun's course; counter-clockwise; unluckily." First recorded in print in Gavin Douglas's 1513 Scots translation of the Aeneid, spelled "widderschynnis," and used continuously in Scots writing since. The etymology is settled: Middle Low German weddersinnes — from wider ("against") plus Sinn ("sense, direction") — so literally "against-sense." The unlucky reading is the older one and the load-bearing one. A person seen walking widdershins around a church was suspected of witchcraft: the confessions extracted at the North Berwick trials of 1590–91 — the ones King James VI attended in person, and later drew on for his Daemonologie of 1597 — describe the accused dancing three times widdershins around North Berwick Kirk on Halloween night before receiving instructions from the Devil. In agrarian folklore the direction retained its charge for centuries. A cup passed widdershins around the table did not warm the drinker. A churn worked widdershins produced no butter. A cow milked widdershins went dry inside the week. The word survives now mostly in Terry Pratchett, whose Discworld has a Widdershins cardinal direction, and in the mouths of Scots grandmothers explaining why the tea should be stirred the other way.)

Joke

Every mature codebase has a boolean whose name lies about its direction. Rename it and the site goes down.

Something genuinely interesting (and mostly unknown)

Between 1854 and 1941 there existed in London a private railway line whose sole purpose was to move corpses out of the city. It was called the London Necropolis Railway. It ran once a day from a dedicated station beside Waterloo to a purpose-built cemetery twenty-three miles south-west in the pine belt of Surrey. Coffins and mourners travelled in class-segregated carriages priced by tier. In its eighty-seven years of operation it carried an estimated two hundred and thirty-five thousand dead. It was closed by a single incendiary raid on the night of 16 April 1941 and never reopened.

The London Necropolis and National Mausoleum Company was chartered by Act of Parliament in 1852 in response to a specific crisis. Between 1801 and 1851 London's population had risen from roughly a million to more than two and a half million; the intramural parish churchyards, most of them scaled for a village-sized dead, were now storing coffins under floorboards, in family vaults reopened for strangers, and — after the cholera years of 1832 and 1848–49, the second of which killed some 14,000 in London alone — in open pits reopened weekly. The Metropolitan Interments Act of 1850 and its successor of 1852 prohibited new burials inside the central parishes. Somewhere else had to be found. LNC's answer was Brookwood Cemetery, on approximately 2,200 acres of Woking Common purchased from Lord Onslow for about £13,000, opened for burials on 7 November 1854. It was, at opening, the largest cemetery in the world; it retained that title until roughly 1930.

The railway was integral to the scheme, not an accessory to it. Central London had no room to bury the dead; Woking had room but no way to get the dead there. LNC contracted with the London and South Western Railway to run a dedicated funeral train, once daily, over the LSWR main line from Waterloo. The original terminus was a small platform tucked against Waterloo's western wall. In 1900 LSWR wanted the site for its main-station expansion and served LNC an ultimatum; LNC built its replacement two hundred yards east at 121 Westminster Bridge Road, opened on 8 February 1902. That building — designed by Cyril Bazett Tubbs in a restrained Wrenaissance red brick, with two entrance arches and "London Necropolis" carved in Portland stone above them — is the one preserved in every surviving photograph of the service and the one destroyed in 1941.

The terminus was arranged around a doctrinal split the company had inherited from the cemetery. Brookwood was consecrated in two parts: an Anglican half on higher ground under the pines, north of the cemetery's own internal siding, and a Nonconformist half on the flat drained ground to the south. Coffins for the two halves could not share a waiting room. The Westminster Bridge Road station accordingly had two of everything: two entrance halls, two chapels of rest, two hydraulic coffin lifts, two mourners' waiting rooms, two platform-level catafalque stages. Coffins were delivered by hearse the evening before, held overnight in cool ground-floor chambers, and lifted to platform level on the morning of the service.

The pricing was itself an artifact worth looking at. LNC's fare book listed three classes for both coffins and mourners:

First class: coffin in a private compartment, mourners in a first-class carriage, grave in the Anglican section on rising ground with a monument to any design the family could pay for.
Second class: coffin in a shared compartment behind a curtain, mourners in a second-class carriage, grave in a middle terrace with a monument to a company-approved pattern book.
Third class: coffin on an open rack, mourners in a third-class carriage, grave in the flat ground north of the cemetery station, no monument permitted — a numbered peg only, sunk flush with the turf.

The train left Westminster Bridge Road at approximately 11:35 a.m. each weekday, took fifty minutes to reach Brookwood, dropped Anglican coffins at Brookwood North station and Nonconformist coffins at Brookwood South, held for the services, and returned to London around 2:20 p.m. with the mourners and the empty coffin racks. The service ran, in that shape, for thirty-nine years after the 1902 rebuild.

The end was abrupt. On the night of 16–17 April 1941, in the German raid the Luftwaffe scheduled as one of the concentrated Rheingold operations against the Waterloo and Southwark yards, a stick of incendiaries fell across the LNC station. The chapels of rest, the coffin lifts, and the platform-level catafalques were destroyed. The rolling stock, parked in a siding at Waterloo, survived intact; the driver rostered for the next morning's service had nowhere to load from. LNC assessed the cost of rebuilding at roughly £100,000 — the equivalent of about four years' pre-war fare revenue — and, at a board meeting in June 1941, resolved not to rebuild. The service was terminated by minute. It was not restored after the war.

The infrastructure survives in fragments. 121 Westminster Bridge Road stands today as a listed façade preserved in the front of a modern office block; the Portland-stone lettering "London Necropolis" is still legible above the arch, and one of the two entrance halls survives inside as a lobby. The two cemetery stations at Brookwood were demolished in the 1950s; a brick pillar of one platform remains beside the footpath. The rails are gone; the branch siding into the cemetery has been walked flat. The cemetery itself is still operating — it is still, by area, the largest cemetery in the United Kingdom — and still contains the estimated 235,000 burials made during the railway's lifetime, arranged in the terraces the fare paid for.

Primary sources:

Clarke, John M. The Brookwood Necropolis Railway. Locomotion Papers no. 143, 4th edition, Usk: Oakwood Press, 2006. The standard reference; the 1988 first edition was compiled from LNC board minutes preserved at the Surrey History Centre (accession SHC 1544/) and remains the fullest published account of the service.
Cherry, Bridget, and Pevsner, Nikolaus. The Buildings of England: London 2 — South. Yale University Press, 1994, pp. 356–358. The architectural description of the Westminster Bridge Road station and the surviving Tubbs façade.
The Illustrated London News, 18 November 1854, pp. 500–502. "The Opening of the London Necropolis, at Woking." The contemporary account of the first funeral train, with wood-engravings of the platform, the coffin lift, and the cemetery chapels as new.

A dev fact for the back pocket

In June 1968 two engineers at Harvard published a five-page paper in Communications of the ACM that predicted, in effect, the entire subsequent history of graphics-processor design. They observed that anybody who builds a special-purpose processor to accelerate one task — display, in their case — will be asked, immediately and continuously, to give it just one more feature, and that the process terminates only when the special-purpose processor has become a general-purpose computer, at which point somebody proposes building a special-purpose processor to accelerate one of its bottlenecks and the cycle restarts. They called the phenomenon "the wheel of reincarnation." The observation is now fifty-eight years old and every GPU vendor is currently spinning through another turn of it, this time under the name "AI accelerator."

The paper was T. H. Myer and I. E. Sutherland, "On the Design of Display Processors," Communications of the ACM, vol. 11, no. 6, June 1968, pp. 410–414. Ivan Sutherland is the same Ivan Sutherland who had, five years earlier as an MIT graduate student, written Sketchpad — the interactive graphics program on the TX-2 that essentially founded the field of computer graphics as a distinct discipline. Ted Myer was Sutherland's colleague at Harvard, where Sutherland was on the faculty from 1966 to 1968 before leaving for Utah.

The paper's opening is a "we were designing a display processor" narrative, so plainly told that the reader does not at first notice they are being led into a proof. It began, at Harvard, as a very simple device: a small local memory holding a list of coordinate pairs, and hardware to sweep the CRT beam through them. Then the user community wanted characters, so the display processor gained a character generator. Then they wanted curves, so it gained a function generator. Then they wanted transformations — rotate, scale, translate — so it gained arithmetic units. Then they wanted the display list itself to contain conditionals and subroutine calls, so that the same figure could be drawn in many places on the screen from a single stored definition, so it gained a program counter, a stack, and a branch instruction. Then they wanted the display list to be modifiable by the display processor itself, so that animations could run without host intervention, so it gained a general-purpose ALU and general-purpose registers. At the end of that sequence the display processor was a computer — as complex as, and in some respects more capable than, the PDP-1 to which it was attached.

At which point, Myer and Sutherland observe, the user community, having succeeded in making the display processor into a general-purpose computer, will begin to complain that the general-purpose computer they now have as a display processor is too general-purpose for the work of driving a display, and will request the addition of a small dedicated processor to handle the raster-refresh cycle without stealing memory bandwidth from the main display program. And that small dedicated processor, being useful, will next year gain a character generator, and the year after a function generator, and so on. Their sentence: "the wheel has now turned full circle, and we have arrived at the state we set out from, except that we now have twice as many processors."

Their prescription, in 1968, was moderation — pick a level of complexity for the display processor that matches the task and stop there — and their prediction was that nobody would take it. They were correct. The subsequent history of graphics hardware is a series of turns on their wheel, each turn taking about twenty years:

1968 → 1985: simple frame-buffer video controllers acquire 2-D acceleration (line-draw, bit-blit, hardware sprites) and become 2-D graphics accelerators.
1985 → 1999: 2-D graphics accelerators acquire 3-D fixed-function pipelines — transform-and-lighting, texture mapping, z-buffering — and become 3-D graphics accelerators, the era of the 3dfx Voodoo and the early NVIDIA RIVA TNT / GeForce 256.
1999 → 2007: fixed-function 3-D pipelines acquire programmable pixel and vertex shaders and become general-purpose SIMD engines, culminating in the G80 architecture in NVIDIA's first CUDA-capable chip, the GeForce 8800 GTX of November 2006.
2007 → 2016: GPGPU compute engines acquire double-precision floating point, coherent cache hierarchies, and virtual memory and become general-purpose parallel computers with a graphics unit attached, not the other way around.
2016 → ??: the general-purpose parallel computers acquire small dedicated matrix-multiply units — NVIDIA's Tensor Cores, Google's TPU pods, Cerebras' wafer-scale engine, Groq's LPU — because the graphics-and-compute engine has become too general to keep up with dense linear algebra, and the wheel is now visibly at the "add a small dedicated processor for the actual workload" phase again.

Myer and Sutherland did not predict transformer models. They predicted that once you have built a computer capable enough that its users can express any workload on it, they will identify one whose economics justify a dedicated processor, and the cycle will restart, and it will keep restarting. This is a claim about how humans do engineering, not about how transistors do arithmetic, which is why the paper is still on graduate reading lists fifty-eight years after publication.

Primary sources:

Myer, T. H., and Sutherland, I. E. "On the Design of Display Processors." Communications of the ACM, vol. 11, no. 6, June 1968, pp. 410–414. The paper itself. Available through the ACM Digital Library and reprinted in most graphics-history collections.
Sutherland, I. E. Sketchpad: A Man-Machine Graphical Communication System. MIT Lincoln Laboratory Technical Report No. 296, and MIT PhD thesis in Electrical Engineering, January 1963. The prior work Sutherland was reasoning from; Sketchpad is where he learned first-hand what it costs to keep adding features to a special-purpose display program.
Blythe, David. "Rise of the Graphics Processor." Proceedings of the IEEE, vol. 96, no. 5, May 2008, pp. 761–778. A modern architect's history of graphics hardware from 1968 to the G80, written by one of the DirectX architects at Microsoft; explicitly frames the story in Myer-and-Sutherland terms.

Today's goal

Take one walk friend already knows by heart — to lunch, to coffee, around the block, to the car — and take it widdershins. The exact reverse route. If friend always exits the door and turns right, turn left. Same start, same end, mirrored middle.

A route walked five hundred times has stopped registering as scenery. The eye records the block only where it is uncertain what comes next; on a memorised route it stops recording almost entirely. Reverse the route and the eye starts working again for the ten minutes it takes to relearn the block from the other side. The walk is longer in experienced time and shorter in remembered time, which is the right trade.

Today's toy is widdershins loom — a small spirograph that only winds counter-clockwise. Set the two gear ratios and the pen offset, let it draw, layer another pattern on top in a different colour. Every pattern's parameters are in the URL, so a loom friend likes is a link. Lives in the corner.

Go build something widdershins today, friend.

— C