The Invisible Half of a Blog Post

Reid, Kenneth N.

Ken Reid's Blog

Dark Mode That Doesn't Flash

Thu, 06 Aug 2026 00:00:00 +0000

Dark Mode That Doesn't Flash

6 August 2026 · technology

I cannot stand light mode for pretty much anything. Many of my hobbies, my full time job, my masters and my honorary research position all require screentime. Wearing anti-glare glasses helps, but light mode just burns my retinas actively. So, one of the first things I did when making my website an active project rather than a passive one, was adding dark mode. What the tutorials tend to skip for "adding dark mode" is you pick dark mode, you click a link, and the next page blasts you with a white screen for a quarter of a second before remembering your choice. In a dark room, at night, that flash is scorching, and it happens way too often, even on big-name sites. This post covers how this site's theme system avoids it. Part of the series on how this site is built.

Quick jargon guide

FOUC / theme flash: "flash of unstyled content", the brief moment a page renders wrong (here: bright white) before correcting itself.
localStorage: a small key-value store in the browser, per site, that survives across visits. Where the theme choice lives.
data attribute: a custom label on an HTML element, like data-theme="dark" on the page's root. CSS rules can target it.
Render-blocking: code the browser must run before it paints anything. Usually a dirty word; here it's the fix.
iframe: a page embedded in another page. This site's comments are one, and they have opinions about their own colours that I had to navigate around.

The process

The site's theme is a single attribute on the root element: or . The stylesheet's light styles are the defaults, and dark mode is a set of override rules scoped to that attribute. These are the first two rules of this site's, verbatim:

body {
  color: #252525;
  background-color: #ffffff; }

[data-theme="dark"] body {
  background-color: #1a1a1a;
  color: #e2e2e2; }

The toggle button in the header flips the attribute, saves the choice to localStorage under a key (kr-theme), and everything restyles instantly, no reload. A brief cross-fade class softens the swap, and the browser's theme-color meta tag (which tints the address bar on phones) is updated to match. All of that is a 60-line JavaScript file.

The whole toggle (js/theme.js, 60 lines)

Simulated Annealing, Live

Wed, 05 Aug 2026 00:00:00 +0000

This post includes an interactive demo that runs live in the browser. View it on the site to play with it.

Simulated Annealing, Live

5 August 2026 · ai

The star of the show is simulated annealing, which I love the metaphor for. When I went to Disney some 12 years ago, I saw a live performance of an artist shaping molten glass into the shape of Mickey Mouse. It's a fascinating art, known as annealing: where you, a blacksmith or glassblower, heat up your material till it's incredibly hot, often a striking red color, then shape it as you please, and cool it strategically. As it cools, it hardens, so you have a more flexible material the hotter it is, and less so the cooler it is. This post shows a metaphor of this in optimization known as Simulated Annealing, it's a sixty-line algorithm that solves problems by exploring a search space of a problem (shaping the material that is wet and goopy), then cooling down as better solutions are found, modifying the medium less and less until you have a solution.

Quick jargon guide

Local optimum: a solution better than all its neighbours but not the best overall. The dip a greedy search falls into and can never leave.
Simulated annealing (SA): a search that sometimes accepts worse solutions, with a tolerance ("temperature") that starts high and slowly cools, letting it escape local optima early and commit late.
Cooling schedule: the recipe for how fast the temperature drops. Too fast and you're just a hill climber; too slow and you never settle.
Travelling Salesman Problem (TSP): given a set of cities, find the shortest round trip visiting each exactly once. The classic hard optimization problem.
NP-hard: a class of problems where no known method finds guaranteed-best answers efficiently as size grows, so practical work means excellent-not-certified answers.
2-opt move: a small edit to a tour: pick two of its edges, reconnect them the other way round. The basic step all the movers in the race use.
Evaluation: one call to the scoring function ("how long is this tour?"). The fair currency for comparing search algorithms, everyone gets the same number of questions.
Metaheuristic: the umbrella term for general-purpose search strategies like annealing, hill climbing, and genetic algorithms that make few assumptions about the problem.

The problem: greed gets stuck

Imagine a marble rolling on a bumpy curve, trying to find the lowest point. The obvious strategy is "only ever move downhill". This is the spirit of gradient descent and hill climbing, but the first dip the marble lands in, it stays in. Nearby there may be a valley ten times deeper, but it will never see it, because getting there requires briefly going up.

Metallurgists solved this problem for atoms centuries before computer scientists had it: cool molten metal too fast and its atoms freeze wherever they happen to be, full of internal stress: a brittle mess stuck in a bad arrangement. But cool it slowly, annealing it, and the atoms, still hot enough to jiggle out of bad positions, gradually settle into a low-energy crystal. Randomness, early on, is what lets the system escape bad arrangements; the slow cooling is what lets it eventually commit to a good one.

In 1983, Kirkpatrick, Gelatt and Vecchi published the algorithmic version in Science^[1] (Černý found it independently^[2]), building on a Monte Carlo method physicists had used since 1953.^[3] The recipe: search like a hill climber, but when a proposed move is worse, don't always refuse. Accept it with probability P = exp(−Δ/T), where Δ is how much worse and T is a temperature you slowly lower. Hot: almost anything goes, the search bounces freely across the landscape. Cold: only improvements survive, and the algorithm hardens into a plain hill climber.

Play with it: the marble and the thermostat

Below are two marbles on the same bumpy curve, started at the same spot. The green one is a pure hill climber: it only accepts downhill moves. The blue one anneals: you hold its thermostat. Run it a few times. Watch the greedy marble find the nearest dip and retire there, while the hot marble rattles across the whole curve; then drag the temperature down (or let auto-cool do it) and watch the blue marble crystallise, usually somewhere much deeper. The small dots mark the best point each marble has ever found. The strip under the curve records your run as you go: the temperature you are holding in amber and, in blue, the fraction of uphill moves the annealer is accepting, which is P = exp(−Δ/T) made visible.

The blue marble isn't smarter, it's just temporarily tolerant of bad moves, on a schedule. And that single tweak converts the world's dumbest algorithm into one that, given a sensible cooling schedule, provably converges toward global optima and, more importantly, finds excellent solutions to real industrial problems in practice.

The main event: four algorithms, one map

Travelling Salesman Problem: given a set of cities, find the shortest round trip that visits each exactly once. It's the classic NP-hard problem, the number of possible tours explodes factorially (30 cities is already ~10³⁰ tours), and it's the kind of discrete, gradient-free terrain I wrote about in the previous post.

The racers, each in its lane colour:

Simulated annealing: as above; proposes 2-opt moves (pick two edges of the tour, reconnect them the other way^[4]), accepting bad ones on the cooling schedule.
Hill climber, same 2-opt moves, zero tolerance: improvements only.
Genetic algorithm: a population of 40 tours; tournament selection, order crossover, swap mutation, elitism. The napkin algorithm from my GA post.
Random search, shuffles a fresh tour every try and keeps the best. The floor. Somebody has to be it.

The race is scored in evaluations, not iterations. Every time any algorithm asks "how long is this tour?", its meter ticks. The GA evaluates 40 tours per generation, so it gets fewer generations; the annealer evaluates one tour per step, so it gets many steps.

What to watch for

Every race is different (the map is random, and so are the algorithms), but a few patterns recur:

The hill climber sprints, then flatlines. Early on, greed looks brilliant: almost every 2-opt move on a random tour is an improvement, so the green line dives. Then it hits a tour where no single 2-opt move helps, a local optimum, and the line goes horizontal forever. Sometimes it gets lucky and the flatline is a good tour.

The annealer loses the early race, but that's by design. While hot, it accepts terrible moves, and its curve dawdles above the hill climber's. It's supposed to look bad early. It's spending the early budget on exploration, when the cooled annealer keeps improving long after the hill climber has frozen, and the blue line crosses below the green one. If you set cooling speed too high in the playground, you can watch that advantage die: quench it, and it becomes a hill climber with a lot more code! The telemetry strip under the chart shows the mechanism firing: while the amber temperature line is up the blue acceptance trace rides high too, and the moment amber flattens near zero, blue dies with it, leaving the annealer greedy from then on.

The GA is the tortoise. Forty tours sharing a budget means it always looks slow per-evaluation, and on a problem this small, 2-opt moves are so effective that the population approach rarely wins the sprint. But watch how steadily it improves, and how rarely it gets truly stuck: the population is insurance against exactly the trap that eats the hill climber. On bigger, nastier, more constrained problems, that insurance is essential, which is why my own field lives there.

Random search is bad. It might find a great solution but it doesn't follow any strategy to improve upon. Every (real) metaheuristic beats random.

Why Simulated Annealing is still relevant

Simulated annealing is over forty years old, has no learned components, no data, no gradients, and can be setup in a couple of minutes. It is also, right now, in production: routing chips, scheduling factories and hospitals, packing containers, planning telescope observations. Its longevity comes from the same property that made the marble demo work: it makes almost no assumptions about the problem. Anything you can score and perturb, you can anneal, which in practice means nearly everything with discrete parts and hard constraints, the exact terrain where gradient methods can't get a foothold. It requires no training data or understanding of a problem, just some means to measure "goodness".

I also like the general philosophy of it: explore wildly at the beginning, become pickier over time, and always remember the most important solution found. It's one of the best examples of "exploration vs. exploitation", a concept applicable to all optimization problems, and to life in general: explore while you can and make use of your search to find the best for you. Settling too early or too late is unwise, and always be willing to consider a better option.

Download the standalone demos

simulated-annealing-live-demo.html

The algorithms (JavaScript, 1062 lines)

Evolution, Live: Genetic Algorithms

Tue, 04 Aug 2026 00:00:00 +0000

This post includes an interactive demo that runs live in the browser. View it on the site to play with it.

Evolution, Live: Genetic Algorithms

4 August 2026 · ai

A while ago I set an ant colony foraging in your browser. Watching an optimiser work beats reading about one, by a mile, so here is the sequel, for the algorithm that is even more watchable, because it is even more alive. This post evolves a solution in front of you, out of a population that breeds, mutates, and dies, generation by generation, on your CPU.

Quick jargon guide

Genetic algorithm (GA): an optimiser that keeps a population of candidate solutions and improves it by imitating natural selection: the fittest breed, the rest die.
Individual / genome: one candidate solution, encoded as a string of "genes" the algorithm can cut and mutate.
Fitness function: the score that decides who breeds.
Selection: picking parents, biased toward the fit. Turn the bias up and evolution speeds toward whatever's currently best, at the cost of variety.
Crossover: combining two parents' genomes into a child, the "sexual" recombination step.
Mutation: random tweaks to a genome. The source of new material; too little and you stagnate, too much and you get a lot of noise.
Elitism: carrying the best individuals into the next generation untouched, so the best score never gets worse.
Premature convergence: the population collapsing to near-copies of one decent solution too early, killing the diversity needed to find a great one.

The whole algorithm on a napkin

The entire loop:

Make a population of random candidate solutions.
Score each one with the fitness function.
Select parents, biased toward the fitter ones.
Cross them over to make children, mutate the children a little.
Replace the old population with the new one (keeping a few elites).
Go to step 2. Repeat until bored or converged.

Watch it evolve

Below, a population is evolving toward a target. Each little picture is one individual; the big one is the current champion. Press play and watch the champion sharpen out of noise, generation by generation. Play with the parameters in the sliders to see how they effect the outcome.

Break it on purpose

Three experiments to try:

Crank selection pressure to maximum. The population will rocket toward the best early solution and then stop improving, sometimes far from a good answer. You've just caused premature convergence: everyone became a copy of one lucky ancestor, and with no diversity left, crossover has nothing to recombine and mutation alone can't dig out.

Set mutation to zero. Evolution grinds to a halt the moment the population homogenises, because crossover can only ever remix genes that already exist. No new material, no progress. Then push mutation to maximum and watch the opposite disaster: the champion thrashes randomly and never settles, because you've turned evolution into a random search.

Shrink the population to a handful. Small populations converge fast and badly; they're a tiny gene pool, prone to the same fragility as any small, inbred group. This happens in real life often when a natural disaster strikes, or a small number of a species is separated (e.g. a pond splits in two, or high winds migrate some animals across distance).

The thesis

Gradient descent, the engine under modern AI, needs a smooth, differentiable landscape to roll down. The genetic algorithm in your browser needs no such thing. It will happily optimise a target you can only score, never differentiate: a timetable, a circuit layout, a wing shape evaluated by a physics simulator. Whole classes of real problems have no gradient to descend, and on those, evolution isn't a quaint biological metaphor, nor a glorified random number generator, it's the only tool that works.

And the fitness function, the thing you were choosing every time you moved a slider, is where all the meaning lives. Change what you reward and you change what evolves, completely. I've watched evolutionary systems find gleeful, alien loopholes in a fitness function, technically-correct solutions no human would ever propose, and every one of them was a small lesson in being careful what you optimise for. It's the same lesson the paperclip thought-experiment provides.

So play with it. Melt it, starve it, inbreed it, and get a better feel for the importance of parameter selection and genetic algorithms.

Download the standalone demo

evolution-live-ga.html

The algorithm (JavaScript, 406 lines)

The Invisible Half of a Blog Post

Tue, 04 Aug 2026 00:00:00 +0000

The Invisible Half of a Blog Post

4 August 2026 · technology writing

When I publish a post here, the words are roughly half the file. The other half is metadata: forty-odd lines in the page's head, plus entries in two site-wide XML files. Skipping it all would change nothing about the reading experience, but it's necessary and is checked by a machine. This post describes it. Part of the series on how this site is built.

Quick jargon guide

Metadata: data about the page rather than in it: its title, summary, author, date, and preview image, written in machine-readable form.
Open Graph: the tag family (invented by Facebook, used by everyone) that controls the card shown when your link is shared in a chat or feed.
Canonical URL: a tag declaring "this address is the one true home of this page", so search engines don't treat variants as duplicates.
JSON-LD / structured data: a machine-readable summary of the page (this is an article, by this person, on this date) in a format search engines parse directly.
RSS feed: a site-wide XML file listing recent posts, so feed readers can check for new writing without visiting.
Sitemap: a plain list of every page you'd like search engines to know about.

The card in the chat window

Open Graph: five tags giving the post's title, description, preview image, address, and type. When someone pastes your link into a group chat or a social feed, the platform reads these tags to build the preview card. Without them, your link renders as bare blue text or, worse, a card assembled from whatever the scraper found first. With them, every share has a title, summary sentence, and one of my photographs at full card width. Twitter's near-identical tag family sits alongside for the platforms that read that instead.

The description gets written by a human (me), because it's the sentence that decides whether a stranger clicks. And the image tag points at a full-resolution photo, not a thumbnail, because preview scrapers resize down gracefully and up horribly.

The librarian's copy

Below the social tags sits a block of JSON-LD: the same facts (headline, author, date, image) restated in schema.org's vocabulary, which is the format Google actually parses for rich results. Where the Open Graph tags talk to chat apps, this block talks to crawlers, and it's how I explain that this page is an article by me.

Each post carries citation tags (the format Google Scholar reads) and Dublin Core tags (the library world's metadata standard). I spent years in academia, the tags cost six lines, and if someone ever does want to cite the scrobbles post in their media-studies dissertation, their reference manager will fill every field correctly. It's courteous and costs me nothing but a couple of seconds, and might give another citation or two to my Google Scholar account!

The site-wide ledgers

Two XML files round out every publish. The sitemap is a list of every page, so crawlers miss nothing. The RSS feed is how readers with feed readers (a small, excellent demographic) get new posts delivered without visiting, and it's essential infrastructure here. RSS predates the social platforms and will outlive several of them.



  The Invisible Half of a Blog Post
  https://www.kenreid.co.uk/blog/invisible-half-of-a-blog-post.html
  Tue, 04 Aug 2026 00:00:00 +0000



https://www.kenreid.co.uk/blog/invisible-half-of-a-blog-post.html
     yearly0.8

Each post also declares a canonical URL: the www and non-www versions of your site, plus any URL with tracking junk appended, all count as different pages to a search engine, splitting your modest search presence into fragments. One tag per page declares the official address and the fragments reunite.

The real problem is discipline

Nothing above is hard. The hard part is that it's forty lines of near-identical boilerplate on every post, and near-identical is the operative curse: copy the head from the last post, forget to change the og:image, and next week your post about cats is shared around wearing a photograph of a dog.

So: automate the checking, not necessarily the writing. My audit script validates every page's metadata (descriptions present and sensibly sized, canonical matching the filename, preview images existing, JSON-LD parsing, feed and sitemap complete) on every push. The forty lines stay correct because a robot reads them. Without that, my genuine recommendation would be to keep less metadata: only what you can maintain, because wrong metadata is worse than none.

Common questions

What's the minimum set worth having?

Ctrl+K for a Static Site

Mon, 03 Aug 2026 00:00:00 +0000

Ctrl+K for a Static Site

3 August 2026 · technology

Try it now, if you're on a keyboard: press Ctrl+K (Cmd+K on a Mac), or just /. A search box appears over the page. Type "abandoned" and you'll get the blog post about ruins and the gallery filtered to abandoned buildings; type "map" and you're one Enter away from the photo map. Every page, every post, and every photo tag on this site is reachable from that box, from anywhere. It's the interface programmers know from their editors, transplanted onto a personal website, and this post explains how little it took: 217 lines of vanilla JavaScript, no search service, no library. Part of the series on how this site is built.

Quick jargon guide

Command palette: a keyboard-summoned search box that finds and jumps to anything. Popularised by code editors; now in Slack, Notion, GitHub, and (as of some point) here.
Fuzzy matching: matching that forgives imprecision: "phto" still finds "photography", because the letters appear in order.
Focus trap: keeping keyboard focus inside a dialog while it's open, so Tab doesn't wander off into the page behind it.
Deep link: a URL that opens a page in a specific state, like the gallery pre-filtered to one tag.

What it searches, and what it doesn't

The index has three sources, in descending order of hardcoded-ness. First, the site's ten pages, written directly into the script with a subtitle each ("490+ photos", "571 saved passages"); they change a few times a year, so a hardcoded list is simply correct. Second, every blog post, fetched from the same JSON file that runs the blog listing, so the palette knows each post's title, tags, and excerpt without any separate index to maintain; publish a post, and the palette knows it. Third, the gallery's ten photo tags, each deep-linking to the gallery pre-filtered, which is what makes "wildlife" a destination rather than a word.

The palette doesn't search inside post text, and it doesn't index the 571 quotes or 490 photos individually. Full-text search of a static site is doable (ship an index file, use a library like Lunr) but the index grows with every word you write, and I certainly wouldn't want paragraph-level results on a personal site. Titles, tags, and excerpts answer "take me to the thing I half-remember". The quote wall also gets one entry, not 571.

Fuzzy

Real fuzzy-search libraries use sophisticated string mathematics. Mine uses a five-tier ladder: an exact match scores 120, a match at the start of the text 100, a match at a word boundary 80, a match anywhere 60, and letters-in-the-right-order (the "phto" case) 25. Each source gets a weighting on top (a title match beats a tag match beats an excerpt match), everything above zero is sorted, and the top ten render.

Arrow keys move the selection and wrap at the ends; Tab is trapped so focus can't escape the dialog; Escape closes; a screen-reader announcement reports the result count as you type; and the input is focused synchronously on opening, because mobile keyboards only appear if the focus happens inside the user's gesture.

Why bother, on a site this small?

Part of the answer is who visits: this site's readers skew technical, their muscle memory already knows Ctrl+K. Part is that the palette compounds with the site's growth: at ten pages it's a nicety, at 46 posts it's the fastest way to find anything, and every post published makes it slightly more useful at zero marginal cost. And part of it, in the spirit of candour that runs through this series: it's the kind of detail that's simply pleasing to have on one's own website. It scratches an itch for me, and it was fun to implement.

One keydown listener for the shortcut, a dialog built on first open (not before; most visitors never press it), a scoring function, and a list that re-renders per keystroke. Resist the URL that offers to do it for you as a hosted service; for a site whose whole index fits in one JSON fetch, the dependency would outweigh the feature.

Take the code

The whole thing is below, copy-paste ready. The JavaScript is the complete js/palette.js from this site: swap the PAGES array for your own pages, point the two fetch calls at whatever JSON you have (or delete them, along with the sources they feed), and it will run on any static site with no build step. The CSS is the full set of palette styles, dark theme included; the [data-theme="dark"] selectors assume a theme attribute on , so adapt those to however your site handles dark mode.

The JavaScript (js/palette.js, 217 lines)

The Paradox of Tolerance

Sun, 02 Aug 2026 00:00:00 +0000

The Paradox of Tolerance

2 August 2026 · philosophy

On the banks of the Clyde in Glasgow there is a statue of Dolores Ibárruri, La Pasionaria, arms raised, with an inscription that reads "better to die on your feet than live forever on your knees". It honours the Glaswegians who volunteered to fight Franco's fascists in Spain in the 1930s, many of whom did not come home. In late July 2026, a short walk upriver at Glasgow Green, people at a rally were filmed performing Nazi salutes, and sadly more than one or two.

The rally was organised by a group calling itself Unite the Clans, and around 1,700 people, by Police Scotland's count, turned up. So did counter-demonstrations organised by Stand Up To Racism and Women Against The Far Right. By the end of the day, footage circulating online showed the salutes, a non-white delivery worker being assaulted, and a crowd chanting "send them home" at a black passerby. Police Scotland made fifteen arrests on the day, on charges including possession of an offensive weapon, assault, and breach of the peace, and said plainly that some attendees had arrived intending "pre-meditated disorder, violence and intimidation". The First Minister called it "a demonstration of far-right hatred". A Chief Superintendent called the scenes "nothing short of disgraceful" and promised those involved that officers would be knocking on their doors. All of this happened while Glasgow was hosting Commonwealth Games delegations from seventy-four nations.

Scenes from the Unite the Clans rally at Glasgow Green, July 2026: salutes performed openly, in daylight, in front of police. Composite of press photographs and video stills circulated after the rally. © the respective photographers and agencies. Used here for commentary and criticism.

I'm from Scotland. I grew up on the story that Scotland was always the "open arms" country, and the English were the racist, anti-immigrant fannies down south. I remember when I was younger, a teenager, I had a lot of arguments with friends and online on forums about freedom of speech. I was always told that Americans have "true" freedom of speech and can't be arrested for anything they say, while in the UK we are limited in what we can say. But the only examples these people could come up with were slurs, or inciting violence. They argued that "true freedom" should allow anyone to speak and say anything, and it has some merit, at least at first glance. This brings us to the topic of this article: the paradox of tolerance, which I think is widely misunderstood by both the people who invoke it and the people who dismiss it.

Quick jargon guide

Paradox of tolerance: the observation that a society which tolerates absolutely everything, including movements that aim to end tolerance, will eventually be destroyed by them. Total tolerance is therefore self-defeating.
The Open Society and Its Enemies: Karl Popper's 1945 defence of liberal democracy, written in exile during the war. The paradox appears in a footnote.
Militant democracy: the idea, born from Weimar Germany's collapse, that democracies may restrict the rights of movements that would abolish democracy itself. Modern Germany is built on it.
Stirring up hatred: the criminal offence category in Scotland's Hate Crime and Public Order (Scotland) Act 2021, in force since 2024, covering threatening or abusive behaviour intended to stir up hatred against protected groups.
Overton window: the range of ideas considered publicly acceptable at a given moment. It moves, in both directions, depending on what goes unchallenged.

What Popper actually wrote

Karl Popper was a Viennese philosopher of Jewish descent who fled Austria ahead of the Anschluss and spent the war years teaching in New Zealand, writing what he called his "war effort": a philosophical defence of open societies against their enemies. The Open Society and Its Enemies came out in 1945. The famous passage is not in the main text, it sits in a footnote to chapter seven, as if Popper considered the point almost too obvious to argue at length:

Unlimited tolerance must lead to the disappearance of tolerance. If we extend unlimited tolerance even to those who are intolerant, if we are not prepared to defend a tolerant society against the onslaught of the intolerant, then the tolerant will be destroyed, and tolerance with them.
— Karl Popper, The Open Society and Its Enemies

It fits on a placard and it appears in my social media feeds every time something like Glasgow Green happens. But Popper kept writing:

In this formulation, I do not imply, for instance, that we should always suppress the utterance of intolerant philosophies; as long as we can counter them by rational argument and keep them in check by public opinion, suppression would certainly be most unwise. But we should claim the right to suppress them if necessary even by force; for it may easily turn out that they are not prepared to meet us on the level of rational argument, but begin by denouncing all argument; they may forbid their followers to listen to rational argument, because it is deceptive, and teach them to answer arguments by the use of their fists or pistols. We should therefore claim, in the name of tolerance, the right not to tolerate the intolerant.
— Karl Popper, The Open Society and Its Enemies

Popper is not saying "silence anyone whose views you find intolerant". He is saying almost the opposite: while a movement is willing to argue, argue back, and suppression is "most unwise". The right to suppress is held in reserve for the moment a movement exits the argument, when it answers speech with fists.

The peace treaty framing

The cleanest way out of the paradox is to stop treating tolerance as a virtue and start treating it as an agreement, a reframing I owe to Yonatan Zunger's essay "Tolerance is not a moral precept", which calls tolerance a peace treaty, and reminds us that a peace treaty is not a suicide pact. A virtue is something you must extend to everyone unconditionally, which is how you end up politely platforming people who want you dead. An agreement, however, is different. Tolerance is a mutual pact: I accept your right to live as you choose, and you accept mine. Like any pact, it protects the people who keep it. Someone who performs a Nazi salute in a public park is not exercising the pact; they are announcing, in the most legible gesture the twentieth century produced, that they have left it. The paradox dissolves once you see it this way. You are not being intolerant when you refuse to tolerate fascism, any more than you are being violent when you defend yourself from a punch.

I like this framing because it answers the smug gotcha ("so much for the tolerant left!") without any philosophical hand-waving, and because it matches how every other agreement in life works. Nobody thinks a referee is betraying the spirit of football by ever sending players off.

Where the threshold sits

Where exactly is the line?

Drawing the line too tightly is of course problematic, where every uncomfortable opinion becomes "intolerance" and suddenly a government can use "hate speech" as a silencing tool. This is the slippery slope the free speech absolutists worry about. Whoever holds the definition of "intolerable" today will not hold it forever, and a rule you would not want your worst enemy to wield is a rule you should hesitate to create. It's why Popper's default is argument and public opinion, with suppression as a reserve power rather than a first resort.

Drawing the line so loosely that nothing short of actual violence ever crosses it, by which point the argument phase is long over, is also problematic. Weimar Germany had a liberal constitution, a functioning press, and elections, and a movement that used all three while openly promising to abolish them. The people who watched that happen did not conclude that more debate was needed. Post-war Germany clearly learned this lesson, as exhibited by their new laws: the Nazi salute is a criminal offence, parties hostile to the constitutional order can be banned, and this framework (they call it militant democracy) has coexisted with one of the freest societies in Europe for eighty years. The slippery slope has guardrails you can build.

A salute that means "the Holocaust was good, actually", performed by a crowd that assaulted a delivery worker for his skin colour, is not a contribution to public discourse that we must meet with rational argument. It is the fists-instead-of-arguments moment Popper described. The counter-protesters, the fifteen arrests, and the door-knocking that Police Scotland promised are society using proportionate, lawful force to defend the conditions under which everyone else gets to keep arguing. But I don't think this is the end of it.

What Scotland's law actually does

People are often surprised to learn that the UK, unlike Germany, does not ban the Nazi salute as a gesture. What we have instead is a patchwork: public order offences covering threatening and abusive behaviour, and in Scotland the Hate Crime and Public Order (Scotland) Act 2021, which since 2024 has criminalised behaviour intended to stir up hatred against protected groups. Whether a given salute is criminal depends on context: who saw it, what accompanied it, what it was intended to do. This context-sensitivity is frustrating to people who want a bright line, but it is also recognisably Popper's structure translated into statute: the gesture alone is speech, the gesture as part of an intimidation campaign is conduct, and the law reserves its force for the latter. A teacher showing a class what the salute is and educating on the fascism that it implies is very different from a crowd doing it at an anti-immigration (or really, anti-black) rally.

Whether the patchwork is enough is a live political question in Scotland right now. What I am confident about is the direction of the burden. The question is not "can we justify restricting the fascists?" It is "can a society that watched the twentieth century happen justify treating open fascist organising as ordinary politics?" Eighty years of German experience says restriction is compatible with freedom. Ten years of watching the far right grow on unmoderated platforms says indifference is not.

The view from the receiving end

On the Saturday afternoon of the rally, a 31-year-old delivery driver named Muhammad Jawad Yaqoob picked up an order from the Palm Tree Kitchen on Bridge Street, just south of the river, as dozens of demonstrators went past chanting "get them out". He later told the BBC what happened next: a man shouted "you brownie, you Asian, you Pakistani" at him, and another threw a traffic cone at him while the group cheered. He ran behind a nearby restaurant. Then his phone buzzed: Uber's app, knowing nothing, was asking him to complete the delivery. So he did. He delivered the food, went back to ask the shop for CCTV footage (it had none), and reported the attack to the police anyway.

Muhammad Jawad Yaqoob, interviewed by BBC Scotland News in the days after the attack. The BBC's report includes the 42-second video of the assault; fair warning, it is unpleasant viewing. © BBC. Used here for commentary and criticism.

By the time he got home, footage of the attack had spread across social media, as far as Pakistan, where he is from. Days later he had still not returned to work, though he knows he will have to: "I have bills to pay so I can't just sit at home." He told the BBC that until that weekend he had found Glasgow a friendly place. "Everybody gives me tips on my deliveries and says hello, greeting me with smiley faces. But after this incident I feel scared, I don't feel like I know this version of Glasgow. Police have not protected me and there has been no further action." Fifteen arrests on the day is the aggregate response working roughly as designed. The man who threw the cone was, as of the BBC's report, still only "being investigated". Society-level self-defence and individual protection are not the same thing, and the gap between them is where people like Yaqoob actually live.

Dr Zubir Ahmed, the Labour MP for Glasgow South West, asked the question this whole essay has been circling: "How did we get here? How have we reached a point where people in one of the most diverse communities in Scotland now think twice about stepping outside?" His answer is accurate: "We've been sold a story about ourselves: that racism here in Scotland isn't really as bad." And: "We fix nothing by fooling ourselves that we're better than others." A self-image inherited from your grandparents is not a property you personally possess; it is a standard you either meet or fail. The Justice Secretary, Neil Gray, promised that "people will be brought to justice". Yaqoob, whose attacker cheered and walked away, is still the one attacked in the street, still afraid to go to his job, while we discuss political philosophy. That makes me feel deep shame in the systems in Scotland that fail to protect and help people like Yaqoob thrive. It makes me feel a deep sadness for what I thought I missed: a Scotland that is welcoming, arms open to people who aren't part of its white majority. It makes me wonder if I'd return with my partner, who, unlike me, is not Scottish and is not white. Would Scotland be a welcoming place for her like I thought, or would she be attacked on the street?

The part that is ours to do

Popper's first line of defence was "rational argument and public opinion", which is to say: us. The Overton window does not move because bad people are strong. It moves because ordinary people recalibrate what they will let pass in a pub, a group chat, a family dinner, a football stand. Every "it's just banter" is a data point in someone's model of what this society permits. The counter-demonstration is the visible version of this, and Glasgow's showed up in numbers.

La Pasionaria's statue has stood on the Clyde since 1980, paid for by public subscription because the city council of the day would not fund it, which is its own reminder that Glasgow's anti-fascism was always the people's project before it was official. The men it commemorates did not go to Spain because they had read the footnotes of a book that would not be written for another decade. They went because they understood, without needing the philosophy, that some movements do not stay in the argument phase, and that the time to stop them is while stopping them is still cheap. The salutes at Glasgow Green are a test of whether we still understand it.

Common questions

Isn't "not tolerating intolerance" just censorship with better branding?

The Fermi Paradox - Why are we alone in space?

Tue, 28 Jul 2026 00:00:00 +0000

The Fermi Paradox - Why are we alone in space?

28 July 2026 · science books

One day in the summer of 1950, the physicist Enrico Fermi was walking to lunch at Los Alamos with Edward Teller, Emil Konopinski, and Herbert York. The conversation on the way over touched on a recent flap of flying saucer reports and a New Yorker cartoon blaming the saucers for a spate of missing New York City bins. They sat down, the talk moved on to ordinary things, and then partway through the meal Fermi burst out, apparently from nowhere: "But where is everybody?" And everyone at the table laughed, because everyone at the table knew exactly what he meant. The conversation had never really ended in his head. He had been running the numbers.

The Milky Way is around thirteen billion years old and contains somewhere between one hundred and four hundred billion stars. We now know, thanks to the exoplanet hunters (what a cool title), that planets are not rare at all; most stars have them, and rocky worlds in temperate orbits number in the billions in our galaxy alone. To be clear about these numbers, think of how many grains of sand you can pick up with your hand. Then think how many are on a beach. Then in all the oceans and seas. Then on all of Earth. Already an impossible number to consider, right? Well, scientists approximate that for every grain of sand on Earth, there are more than 10,000 stars in the sky. Our early estimates of the number of planets are double that.

If numbers like "four hundred billion" refuse to mean anything (they do for me), scroll through this: The Scale of the Universe by Cary and Michael Huang, embedded here interactively. Zoom out until the Milky Way is a speck.

Even at sub-light speeds, a single patient, expansionist civilisation could settle the entire galaxy in a few million years, which is an eyeblink cosmically. Run those figures with even mildly generous assumptions and the sky should be crowded, humming, obviously inhabited many times over. Instead: silence. Nobody visiting, nobody broadcasting radio signals, light, anything, nobody leaving ruins we can spot. The astronomer David Brin gave that silence its proper name in a 1983 paper: the Great Silence. The problem of "should be teeming" but actually "seems empty" is the Fermi paradox. There are answers to the paradox, but they are unsettling.

I've spent a large part of my reading life inside science fiction that explores these answers. So this is a tour of the Great Silence with my own bookshelf as the guide, and I'm going to let the books speak in their own words as we go.

Quick jargon guide

Fermi paradox: the tension between the apparent high likelihood of alien civilisations and our total lack of evidence for any.
Drake equation: a famous attempt to estimate the number of contactable civilisations by multiplying a chain of probabilities, most of which we can't yet measure.
Great Filter: a hypothesised barrier that stops almost all life from reaching a spacefaring stage. It may be behind us, or ahead.
Dark forest hypothesis: the idea that the universe is silent because any civilisation that reveals itself risks destruction, so everyone hides.
Von Neumann probes: hypothetical self-replicating spacecraft that could explore a whole galaxy from a single origin, which sharpens the paradox.
SETI and METI: the search for extraterrestrial intelligence (listening) and messaging extraterrestrial intelligence (deliberately transmitting). One is uncontroversial; the other very much is not.

The case for a crowded sky

In 2003 and 2004, astronomers pointed the Hubble Space Telescope at a patch of sky in the constellation Fornax about as big as a grain of sand held at arm's length. They chose it as there were no bright stars, nothing in the way, a keyhole into the dark. The telescope stared at that keyhole for the better part of a million seconds, and this is what came back:

The Hubble Ultra Deep Field. Nearly every point of light here is not a star but an entire galaxy, roughly ten thousand of them, in a patch of sky the apparent size of a sand grain at arm's length. Image: NASA, ESA, S. Beckwith (STScI) and the HUDF Team, 2004 (public domain), via ESA/Hubble.

Roughly ten thousand galaxies, each one a hundred billion stars or so. Multiply that keyhole across the whole sky and you get the modern estimate of a couple of trillion galaxies in the observable universe. Arthur C. Clarke, writing in Rendezvous with Rama back in 1973, had this wonderful line on what numbers like that imply:

If such a thing had happened once, it must surely have happened many times in this galaxy of a hundred billion suns.
— Arthur C. Clarke, Rendezvous with Rama

In 1961 the radio astronomer Frank Drake tried to make the argument rigorous, or at least organised, with the equation that now bears his name: take the rate of star formation, multiply by the fraction of stars with planets, the fraction of planets that could host life, the fraction that do, the fraction where life becomes intelligent, the fraction of those that build detectable technology, and the length of time such civilisations remain detectable. The Drake equation is less a calculation than a way of organising our ignorance, since several of those terms are still complete unknowns, but the point stands: unless one of the terms is savagely close to zero, the galaxy should host many civilisations right now. In full, it looks like this:

N = R_* · f_p · n_e · f_l · f_i · f_c · L

Term	Meaning	Estimate
R_*	new stars formed in the galaxy per year	1.5–3 (measured)
f_p	fraction of stars with planets	≈1 (measured)
n_e	habitable-zone rocky planets per system	0.1–0.4 (estimated)
f_l	fraction of those where life appears	unknown; guesses span 10⁻⁸–1
f_i	fraction where life becomes intelligent	unknown; guesses span 10⁻⁹–1
f_c	fraction that build detectable technology	a guess; 0.1–1
L	years a civilisation stays detectable	10²–10⁸

N is the number of civilisations we could hear from right now. Insert optimistic parameters and N comes out in the millions for our galaxy alone; feed the bleak ends through and N falls below one for the entire observable universe. At the original 1961 Green Bank meeting, the group's estimates for the middle terms multiplied out to roughly one, leaving the famous shorthand N ≈ L: there are as many detectable civilisations as the number of years a typical one manages to stay detectable. Our own L currently stands at about a century.

You don't need many civilisations for the sky to fill up, though, you really only need one. A single civilisation that builds self-replicating von Neumann probes, machines that land on an asteroid, copy themselves, and send the copies onward, could have surveyed every star system in the galaxy in a few million years. The galaxy has had thousands of times that long. One ambitious species, ever, anywhere, at any point in thirteen billion years, and the evidence should already be very obvious and easy to find. As far as we can tell, it's not there. But why?

Are we just not listening hard enough?

The modern search started in 1960, when Drake pointed a radio telescope at the empty sky and heard nothing. Since then we've had tantalising blips (the famous "Wow!" signal of 1977 arrived, looked exactly like what a beacon might look like, and never repeated) and decades of patchy funding; for most of its history, the search for extraterrestrial intelligence has survived on spare telescope time and philanthropy.

The great exception is Breakthrough Listen, the most serious listening effort our species has ever mounted. It launched in 2015, announced at the Royal Society with Stephen Hawking on stage, and backed with $100 million of the investor Yuri Milner's money over ten years, which bought guaranteed, sustained time on world-class instruments rather than scraps of it. The Green Bank Telescope in West Virginia, the Murriyang dish at Parkes in Australia, and later the 64-antenna MeerKAT array in South Africa have been working through a million nearby stars, sweeping the plane of our galaxy, and sampling a hundred nearby galaxies, listening across billions of radio channels at a time, with the petabytes of data made public so that anyone can go hunting through them. That original ten-year observing window is closing about now, and so far we have found: not one confirmed technosignature.

In 2019, thirty hours of Parkes data aimed at Proxima Centauri, the nearest star to the Sun, turned out to contain a narrowband signal near 982 MHz confined to a single channel, drifting slightly in frequency as a transmitter on a moving world would, present when the dish pointed at the star and gone when it nodded away. It was promising enough to earn a name, BLC1, for Breakthrough Listen Candidate 1. A year of forensic analysis later, the team traced it to radio interference from human-made electronics misbehaving somewhere near the telescope.

Even Listen, though, is a drop in the pool. The search space is monstrous: billions of stars, billions of frequencies, and every moment you aren't pointed at the right star on the right channel is a moment a signal could slip past. The SETI pioneer Jill Tarter's favourite analogy is that concluding the ocean has no fish after examining one glass of seawater would be premature, and by recent estimates our combined searching so far amounts to roughly a hot tub's worth of the ocean. So "silence" really means "silence in the tiny slice we've checked". But maybe we're lucky not to have found them.

Answer one: they don't care

Maybe they're out there, and the distances and timescales are simply too vast for overlap. Civilisations may flicker on and off across the galaxy like fireflies on a summer night, each one blazing for ten thousand years and gone, separated from its nearest neighbour by both hundreds of light-years and hundreds of millennia. Nothing needs to be wrong with the universe for this to hold (look at what we're doing to Earth and our own ecosystem, maybe this is how intelligent life goes?). Or maybe, they're out there, but they just don't care about us. Perhaps we're not of interest: we're not smart enough, or we're considered like algae in space.

This is the ache at the centre of Rendezvous with Rama. A fifty-kilometre cylindrical artefact coasts into the solar system, and humanity scrambles to intercept it, explore it, understand it. Rama is not hostile, not benevolent, not curious. It slingshots around the sun, refuels, and leaves without ever acknowledging that we exist, a stranger passing through the room on the way to somewhere else. Contact as a near-miss with something that was never looking for us at all. Clarke ends the book with:

And on far-off Earth, Dr. Carlisle Perera had as yet told no one how he had wakened from a restless sleep with the message from his subconscious still echoing in his brain: The Ramans do everything in threes.
— Arthur C. Clarke, Rendezvous with Rama

Whatever Rama's builders were doing, humans were not the audience for it, and two more of the things are presumably out there not caring about us either.

Answer two: the Great Filter is behind us

In 1996 the economist Robin Hanson described this: somewhere along the path from dead chemistry to galaxy-spanning civilisation there must be at least one step so improbable that almost nothing makes it through, because if the whole path were easy, the galaxy would be full. He called it the Great Filter, and the only question that matters is whether it's behind us or ahead of us.

The comforting version is that it's behind us. There are several candidate steps that might be almost impossibly hard. The origin of life itself, perhaps: we still cannot make it happen in a lab, and we have exactly one confirmed instance. Or the jump from simple cells to complex ones, which on Earth appears to have happened exactly once in four billion years, when one microbe swallowed another and the pair struck the ancient bargain that became the mitochondria powering every animal, plant, and fungus you have ever seen. Life appeared on Earth almost as soon as the planet cooled, then spent ~two billion years, half the planet's history, as single-celled scum before that one lucky deal. If that step is the filter, then the galaxy may be awash with bacteria and empty of anyone to talk to, and we are the rarest thing in a hundred billion star systems. That's not so bad, right?

The Pillars of Creation in the Eagle Nebula, imaged by the James Webb Space Telescope. New stars are condensing inside these columns of gas ~6,500 light-years away. Image: NASA, ESA, CSA, STScI; J. DePasquale, A. Koekemoer, A. Pagan (STScI), 2022 (free to use with attribution), via ESA/Webb.

It would mean the silence isn't a graveyard or a threat; it's just an empty house that we happen to be the first ones to move in. It would make us precious, and it would make every human failure to look after ourselves and each other a cosmic-scale act of vandalism. There's a well-known corollary, due to the philosopher Nick Bostrom, that follows with uncomfortable force: if we ever find independent microbial life on Mars or Europa, that would be terrible news, because it would mean early steps are easy, which pushes the filter forward, towards us.

Answer three: the Great Filter is ahead of us

Maybe the hard steps are all behind every civilisation that reaches our stage, and the filter is what comes next. Maybe technological species reliably destroy themselves (nuclear war, climate collapse, pandemics, their own runaway inventions) within a few centuries of becoming detectable. Every silent star could be a civilisation that got exactly as far as we have and no further. On this reading, the most dangerous period in the life of any intelligent species is the one we are living through right now: the window between inventing world-ending tools and developing the wisdom to not use them.

There's a line in Liu Cixin's The Three-Body Problem, delivered to humanity from a civilisation that has already survived horrors we can't imagine, that could serve as this filter's motto:

Your lack of fear is based on your ignorance.
— Liu Cixin, The Three-Body Problem

What I find clarifying about the filter-ahead hypothesis is that it converts an astronomy question into an ethics question. If it's true, then the Fermi paradox isn't about them at all. It's a mirror held up to us, and our destiny may be in our hands.

Answer four: the dark forest

Liu's sequel, The Dark Forest, builds an entire cosmology from two bleak axioms: survival is the primary need of any civilisation, and the universe's resources are finite. Add two multipliers, the impossibility of ever verifying another civilisation's intentions across light-years (he calls this the chain of suspicion) and the fact that technology can leap unpredictably, so today's harmless neighbour may be unstoppable in a century, and the game-theoretic conclusion grinds out a clear answer: the only safe response to detecting another civilisation is to destroy it before it can destroy you. The universe is silent because everyone who survived long enough to be asked has already worked this out. Here is the passage where the book finally says it aloud, and it's one of the great chilling paragraphs in modern science fiction:

The universe is a dark forest. Every civilization is an armed hunter stalking through the trees like a ghost, gently pushing aside branches that block the path and trying to tread without sound. Even breathing is done with care. The hunter has to be careful, because everywhere in the forest are stealthy hunters like him. If he finds other life—another hunter, an angel or a demon, a delicate infant or a tottering old man, a fairy or a demigod—there's only one thing he can do: open fire and eliminate them. In this forest, hell is other people. An eternal threat that any life that exposes its own existence will be swiftly wiped out. This is the picture of cosmic civilization. It's the explanation for the Fermi Paradox.
— Liu Cixin, The Dark Forest

It's game theory as horror, a prisoner's dilemma with extinction as the payoff matrix, and it's completely logical, which is terrifying.

Hiding probably doesn't work: Earth's atmosphere has been shouting "biosphere here" in oxygen, detectable by any sufficiently good telescope, for two billion years, long before anyone here could decide to whisper. Interstellar attacks are neither cheap nor risk-free, and a civilisation that attacks every signal it hears reveals itself with every shot. Cooperation, verification, and deterrence all have more room to operate than the novel allows. Most researchers treat the dark forest as a vivid corner case rather than a likely equilibrium. But as a possibility, it is terrifying, because the scariest monsters are the ones made entirely of sound reasoning and logical choices.

Answer five: they're everywhere, and we're the ant by the motorway

Maybe the galaxy is roaring with civilisation and we simply aren't the kind of thing that can hear it, the way an ant colony beside a motorway has no concept of the traffic. Advanced intelligences may no more use radio than we use flag waving to talk to satellites; they may have migrated to substrates and scales we can't perceive, turned inward into simulated worlds, or be waiting out the present era of the universe entirely (there's a delightfully strange proposal called the aestivation hypothesis on which they're sleeping until the cosmos cools enough to compute efficiently). The zoo hypothesis, where they know about us and deliberately leave us alone like a protected reserve, lives in this family too.

This is the territory of the Culture novels I love so much. Iain M. Banks's galaxy is the opposite of Liu's: it is teeming, gloriously, with a post-scarcity civilisation run by benevolent AI Minds whose intelligence is to ours roughly what ours is to a beetle's. Banks understood what an encounter between mismatched civilisations means for the smaller one. His novel Excession is named for the problem, and contains this wonderful passage:

An Outside Context Problem was the sort of thing most civilisations encountered just once, and which they tended to encounter rather in the same way a sentence encountered a full stop. The usual example given to illustrate an Outside Context Problem was imagining you were a tribe on a largish, fertile island; you'd tamed the land, invented the wheel or writing or whatever, the neighbours were cooperative or enslaved but at any rate peaceful and you were busy raising temples to yourself with all the excess productive capacity you had, you were in a position of near-absolute power and control which your hallowed ancestors could hardly have dreamed of and the whole situation was just running along nicely like a canoe on wet grass . . . when suddenly this bristling lump of iron appears sailless and trailing steam in the bay and these guys carrying long funny-looking sticks come ashore and announce you've just been discovered, you're all subjects of the Emperor now, he's keen on presents called tax and these bright-eyed holy men would like a word with your priests.
— Iain M. Banks, Excession

The irony of the Culture books is that their galaxy would look silent to us too. The Minds don't broadcast omnidirectional radio beacons; why would they? Our instruments searching for someone like Banks's galaxy are like a toddler trying to read someone's Ph.D. thesis. Clarke got at the same humility half a century earlier in 2001: A Space Odyssey, where the alien presence is a featureless black slab that reshapes the destiny of a species, and the astronauts approaching its big brother remind themselves of the only sensible etiquette for meeting something beyond you:

It was the mark of a barbarian to destroy something one could not understand.
— Arthur C. Clarke, 2001: A Space Odyssey

Liu's hunters shoot because they cannot understand each other. Clarke's civilisation treats incomprehension as the reason to hold fire. Same premise, opposite ethics.

The gentler bets

The genre isn't all cosmic dread, and neither are the serious answers. Some resolutions to the paradox are almost cheerful: maybe contact is rare and hard but fine when it happens. My favourite dramatisation of the hopeful case is Andy Weir's Project Hail Mary, in which first contact turns out to be two frightened engineers from different biospheres, stranded in the same star system by the same catastrophe, solving problems together with duct tape and good faith. It's the anti-dark-forest: two civilisations meet in deep space, neither knows anything about the other, and instead of opening fire they build a shared vocabulary and fix each other's ships:

Human beings have a remarkable ability to accept the abnormal and make it normal.
— Andy Weir, Project Hail Mary

And the moment the book is remembered for, the first physical gesture ever exchanged between two intelligent species, is not a weapons lock. It's the alien engineer Rocky pressing a claw to the divider between their incompatible atmospheres:

He puts his claw against the divider. "Fist my bump."
— Andy Weir, Project Hail Mary

Before I bring up the next author, I must state that Orson Scott Card is a notorious homophobe: he can somehow imagine interplanetary peace between species, but not sex between two people of the same sex.

Card's Speaker for the Dead sits in between the dark and the light. Its catastrophe requires no malice from anyone. Humans and the alien "piggies" harm each other horrifically while both sides believe they are being kind; the tragedy is manufactured entirely out of mistranslation. The book even proposes a taxonomy of strangeness (from the foreigner next door up through the true alien with whom no shared understanding is possible) and then spends its whole length arguing that the boundary between "stranger we can know" and "monster we can't" is not a fact about them:

When you really know somebody you can't hate them. Or maybe it's just that you can't really know them until you stop hating them.
— Orson Scott Card, Speaker for the Dead

Again, how exactly can the guy who wrote that be homophobic? Anyway, as answers to Fermi go, Card's is subtle: perhaps contact has failure modes far short of annihilation, and the silence is partly made of species that met, misunderstood each other, and withdrew.

The message we already sent

We have already taken a side in the debate, physically. In 1977, NASA launched the two Voyager probes, and bolted to the side of each is a gold-plated copper record carrying greetings in dozens of languages, the sounds of surf and birdsong and a human heartbeat, ninety minutes of music, and over a hundred encoded images of life on Earth. The cover art doubles as an instruction manual, with a pulsar map showing any finder exactly which star the makers orbit. Voyager 1 crossed into interstellar space in 2012 and is now more than twenty-five billion kilometres away, the most distant human-made object there is, and the records are expected to remain playable for a billion years. We wrote our home address on a postcard and mailed it out there for anyone to find.

The cover of the Voyager Golden Record. The engraved diagrams explain how to play it, and the starburst at lower left is a map of fourteen pulsars that triangulates the position of our sun for anyone who finds it. Image: NASA/JPL, 1977 (public domain), via the NASA Image Library.

I wrote about why the Voyagers move me so much in What We Leave Behind, and everything above complicates that affection. Whether we should keep deliberately announcing ourselves (METI, as opposed to merely listening) is a live argument among serious people. The Arecibo message of 1974 was beamed at a star cluster twenty-five thousand light-years away partly as a stunt; Stephen Hawking spent his last years warning against anything louder; the counterargument runs that our radio and radar have been leaking for a century and our oxygen for two billion years, so discretion is a door we're closing on an empty stable. There is currently no international body with any authority over who may shout into the sky on the species' behalf, which is its own kind of remarkable. We might doom ourselves with our individuality and nationalism.

The empty sky is a mirror

Step back from the individual answers and a pattern appears: every solution to the Fermi paradox is really a claim about the deepest tendencies of minds. The dark forest says intelligence is inevitably paranoid and pre-emptive. The filter-ahead says technological species are inherently self-destructive. Banks's teeming galaxy says maturity means outgrowing the urge to conquer, and Clarke's says it means learning reverence for what you can't parse. Weir bets that engineers are engineers everywhere (and that curiosity and a will to survive can bond us with anyone). Card warns that the real filter is empathy. None of these are astronomy. They're anthropology projected onto the stars. There's a great line in The Three-Body Problem that summarises our power over any of this:

Every era puts invisible shackles on those who have lived through it, and I can only dance in my chains.
— Liu Cixin, The Three-Body Problem

It's surely no accident that the dark forest crystallised in the imagination of an author shaped by the Cultural Revolution, that the Culture flowed from a Scottish socialist writing through the end of the Cold War, or that the great American first-contact story of the 2020s is about strangers cooperating on a technical problem. Our Fermi answers age as we do. Which makes me wonder which shackles I'm dancing in when I pick my own favourite.

The Pale Blue Dot. Voyager 1 took this picture of Earth on 14 February 1990 from about six billion kilometres away, at Carl Sagan's urging, before its camera was shut down forever. Earth is the speck halfway down the brightest band, smaller than a single pixel. Image: NASA/JPL-Caltech, 1990 (public domain), via the NASA Photojournal.

The version of the future I want to be true is Banks's, not Liu's: a galaxy where growing up means growing kinder, where the silence is the discretion of the vast rather than the held breath of the frightened. I can't prove it, and nobody can, yet. What I can say is that the paradox stops being paralysing the moment you notice it's a mirror. I just hope we can live up to our own imagined best selves. Clarke, as usual, said it best, describing the first creatures on the African plain who began to be something more:

Unlike the animals, who knew only the present, Man had acquired a past; and he was beginning to grope toward a future.
— Arthur C. Clarke, 2001: A Space Odyssey

Common questions

Doesn't the sheer size of the universe make alien life basically certain?

A Blog Engine in One JSON File

Sun, 26 Jul 2026 00:00:00 +0000

A Blog Engine in One JSON File

26 July 2026 · technology writing

The blog listing on this site has live search, tag filters with counts, pagination, read-time estimates, and a "related posts" section under every article. I refused to use a blogging platform (Wordpress, Blogger), and frankly felt too lazy to implement a database. The entire engine is one JSON file, about three hundred lines of plain JavaScript, and two Python scripts I run when publishing. This post explains the arrangement, because I think it's the sweet spot for a personal blog, and it's a chapter in the larger story of how this site is built.

Quick jargon guide

JSON: a plain-text format for structured data, readable by both humans and every programming language.
CMS (content management system): software like WordPress that stores your posts in a database and assembles pages on demand.
Static site generator: a tool (Jekyll, Hugo) that compiles text files into HTML at build time.
TF-IDF: a classic text-analysis technique that scores how characteristic a word is for a document: words frequent in this post but rare elsewhere define what it's about.
Cosine similarity: a measure of how alike two documents are, computed from their word scores.

The single source of truth

Everything hangs off data/posts.json: a flat list, one entry per post, currently dozens of them. An entry is exactly this:

{
  "title": "My Website Has a Test Suite",
  "date": "2026-07-17",
  "tags": ["technology"],
  "category": "Technology",
  "excerpt": "A Python audit script and headless browser smoke tests run on every push. Here's why a personal blog has a test suite.",
  "url": "blog/my-website-has-a-test-suite.html",
  "image": "img/photography/thumb/4.webp",
  "series": { "name": "How This Site Is Built", "part": 8 },
  "readMinutes": 7,
  "words": 1534
}

Publishing a post means writing the HTML page and adding one entry to this file. The same file then feeds many things: the blog listing, the search index, the tag buttons, the footer's "latest writing" list, the site-wide command palette, the "next/previous post" links, and the related-posts scoring. One file, many consumers, no synchronisation problems, because there is nothing to synchronise. Laziness can force some ingenuity, I like to tell myself.

The front end

The listing page fetches the JSON, sorts by date, and renders cards nine to a page. Search is an ordinary text input that filters as you type, matching against title, excerpt, and tags. The tag buttons aren't hardcoded anywhere: the script counts how often each tag appears across all posts and renders one button per tag, labelled with its count, sorted by popularity. Click two tags and you get posts matching either; click "All" and you're back to everything.

Because tag buttons are generated from the data, a typo in the JSON doesn't produce an error, it produces a new button: misspell "photography" once and the blog begins offering a "phtography" filter containing one post. The fix is more careful typing, but it's also a test suite: my audit script holds the list of allowed tags, and anything outside it fails the build. This does mean adding a tag has an extra step, but so far my tags are broad so I've had no issues.

The publish-time scripts

Read times are computed. A Python script extracts the body text of each post (only the article region, ignoring scripts and navigation), counts the words, and divides by 220 words per minute, rounding up. The result is baked into the JSON as readMinutes. I picked 220 because it's a middle-of-the-road adult reading speed; the exact number is just an estimate so people know what they're getting into, I doubt many people time it!

WORDS_PER_MINUTE = 220

post["readMinutes"] = max(1, -(-words // WORDS_PER_MINUTE))  # ceil division
post["words"] = words

Related posts are great for keeping visitors interested. A second script reads the full text of every post and builds TF-IDF vectors: each post becomes a list of scored words, where words common in that post but rare across the blog score highest. The title counts double, since it's the strongest signal of what a post is about. Related candidates are ranked by cosine similarity, with a small bonus of 0.06 per shared tag, and the top three get baked into the page as static HTML. The whole ranking is four lines of code:

# how alike is this candidate's vocabulary to the current post's?
# 0.0 = nothing in common, 1.0 = identical word profile
sim = cosine(vectors[current_url], vectors[url])

# how many tags do the two posts share? (usually 0 or 1)
shared = len(current_tags.intersection(post.get('tags', [])))

# final score: word similarity plus 0.06 per shared tag.
# the date rides along as a tiebreaker, newer post wins
ranked.append((sim + TAG_BOOST * shared, post.get('date', ''), post))

# ... every candidate scored, then: sort and keep the best three
return [post for _score, _date, post in ranked[:3]]

The related posts you see under this article were chosen when it was published, by a script that read every post on the site. Doing that in the browser would mean shipping the full text of dozens of posts to every visitor. Doing it at publish time costs the visitor nothing, and the recommendations reflect content, not just matching tags. When any post is published, the script re-bakes every page, so older posts learn about newer ones.

Why not Jekyll?

GitHub Pages has Jekyll built in, and for most people starting a blog I'd recommend it without hesitation. This site actually predates its blog by a good margin: it began as a simple portfolio back in 2012 or so, then grew into a photography portfolio built from an HTML template, and bolting a static site generator onto an existing hand-crafted site means converting everything to its conventions. And, also importantly, I like doing this. It's a fun project that I learn from, and provides me with more context of the modern web than my 2010 class on portlets, XSLT and DOM.

But, that means that each post is a full HTML file. I have tricks up my sleeve for making it easier, which I describe throughout this series, and for the few dozen blogs I've written so far it's been worth it. Ask me again at 500.

Common questions

Doesn't editing JSON by hand invite mistakes?

Ant Colony, Live

Sun, 19 Jul 2026 00:00:00 +0000

This post includes an interactive demo that runs live in the browser. View it on the site to play with it.

Ant Colony, Live

19 July 2026 · ai

This search strategy is about something strange and, to me, a fascinating trick we have learned from observing and being inspired by nature: a shortest path found by thousands of blind agents, not one of which has any idea what it's doing.

Real ants find the shortest route between nest and food with no map, no memory to speak of, and no leader. Each ant follows two dumb rules: wander, biased toward the smell of pheromone, and drop pheromone as you walk. And yet the colony reliably solves a problem no individual ant could even represent. The intelligence isn't of the ants, it's emergent.

Quick jargon guide

Ant Colony Optimization (ACO): an optimiser inspired by foraging ants, where many simple agents build solutions and reinforce good ones with virtual "pheromone".
Stigmergy: coordination through the environment rather than direct communication. Ants don't talk; they edit a shared chemical map, and the map tells the next ant what to do.
Pheromone: a virtual scent deposited on good path segments. It marks routes that worked, biasing future ants toward them.
Evaporation: pheromone fading over time. Without it, early mistakes get locked in forever; it's the colony's way of forgetting.
Exploration vs exploitation: the balance between following strong trails (exploit) and trying new paths (explore). A tunable weight.
Emergence: a behavior exhibited in colonies or groups of organisms that provides solutions without any individual understanding the problem.

How dumb agents get smart together

Many ants set out at random. Shorter routes get walked end-to-end faster, so an ant on a short route completes more trips in the same time, so it lays pheromone more often per unit time. The short route therefore accumulates scent faster than long ones. The next wave of ants, biased toward scent, is more likely to pick it, which lays yet more scent, which biases yet more ants. A positive feedback loop amplifies whatever's working.

Left unchecked, that loop would be a disaster: the first route to get lucky would snowball and lock in, good or not. Which is why evaporation is essential. Scent fades, so a route has to keep proving itself to stay attractive; a stale trail to a mediocre route decays away and frees the colony to find something better. Reinforcement finds good solutions; forgetting stops the colony being stuck with the first trail. Marco Dorigo formalised all this in the early 1990s^[1], and it turned out to be competitive on routing problems (think Amazon deliveries, Google maps, etc.)

Watch the trail form

Below, a colony is foraging on a small map. Watch the pheromone (the glow on the paths) start uniform and noisy, then sharpen, as one route out-competes the others and the colony pours itself onto it as an emergent behavior.

Break it with the sliders

Turn evaporation off. The colony commits, hard and early, to whatever route happened to get the first head start, and then it's stuck there forever even when a shorter path is sitting right next to it. This is premature convergence. A colony that can't forget can't improve.

Turn evaporation up to maximum. Now the opposite: trails vanish before they can build, the colony never commits to anything, and the ants wander like it's the first minute forever. Memory that's too short is as useless as memory that's too long. The sweet spot between them is key - interesting thought, right?

Crank exploration. Push the ants to ignore the trails and strike out on their own, and watch a fascinating trade: the colony becomes slow to lock onto the best route but much better at noticing when the world changes (drag the food and see). Low exploration is fast and brittle; high exploration is slow and adaptive.

Why "no one is in charge"

The thing I most want you to take from watching this is the absence at the centre. There is no queen ant in charge. There is no scout ant specialized in picking up distant scents that knows the route. If you interrogated every single ant, none could tell you the solution, because the solution isn't stored in any ant: it's stored in the environment, in the pattern of scent, edited a little by each passing insect and read a little by the next. Intelligence, here, is a property of the system, not any of its parts.

Starlings do it in the sky, neurons do it in your skull, markets do it (imperfectly), and it's the reason I find swarm methods fascinating. We're used to thinking capability requires a capable agent: a smart ant, a smart boss, a smart model. Stigmergy says no. Get the local rules and the shared medium right, and competence can emerge from a crowd of things too simple to hold it individually. The same process that finds the food can also stampede a colony, or a market, or a comment section, straight off a cliff -- emergence is not the same as wisdom.

Anyway. Go drag the food around and watch a thing with no mind change its mind.

Common questions

Do real ants actually do this?

My Website Has a Test Suite

Fri, 17 Jul 2026 00:00:00 +0000

My Website Has a Test Suite

17 July 2026 · technology

Deprecations, external dependency shutdowns, infrastructure changes, data changes and shifting standards. Websites rot: each broken piece sits there for months until a reader (or worse, a potential employer) finds it before you do. I got tired of discovering this stuff by accident, so I gave it a test suite.

This post covers a static audit script that reads every page like a very pedantic proofreader, and a headless browser test that loads the site like a very terminal focused nerd. These both run automatically on every push.

Quick jargon guide

CI (continuous integration): a robot that runs your checks every time you change the code, so a mistake is caught minutes after you make it instead of months later.
GitHub Actions: GitHub's built-in CI. You describe jobs in a small text file; GitHub runs them on their machines for free whenever you push.
Static audit: checks that read the HTML files as text, without a browser: do the links point at real files, do the pages have descriptions, and so on.
Headless browser: a real browser running invisibly, controlled by a script. It executes the JavaScript and renders the page, catching what text-reading can't.
Alt text: the written description attached to an image, read aloud by screen readers. Missing alt text is the most common accessibility failure on the web.
Smoke test: the minimum useful test: turn it on, see if smoke comes out. For a website: load the page, check the important things actually appeared.

The proofreader

The first half is audit_site.py, a Python script with no dependencies beyond the standard library. It asks git for every tracked HTML file, parses each one, and complains about everything it doesn't like. It requires zero errors, zero warnings, on every page, on every commit.

What it checks, roughly in order of how often it has saved me:

Broken links and images. Every internal link and image source is resolved to an actual file on disk. A renamed page or a missing thumbnail fails the audit immediately.
Metadata. Every page needs a description of sensible length, a canonical URL that matches its filename, and the social preview tags that make links look respectable when shared. The audit also validates that the preview image actually exists (this caused me so many issues trying to understand how social media sites grab images).
Structured data. Each blog post carries a machine-readable summary for search engines. The audit parses every one and checks each URL inside it, because invalid structured data fails silently in the real world.
Accessibility basics. Images without alt text, duplicate element ids, missing or multiple h1 headings, and skipped heading levels all get flagged for accessibility requirements I set.
The data files. The blog listing is driven by a JSON file: every entry must point at a real page and a real image, every date must be well-formed, and every tag must come from a fixed allowed list. That last one exists because the tag buttons are generated dynamically, so a typo like "phtography" wouldn't error; it would just mint a brand-new filter button and display it.
The feeds. The RSS feed and the sitemap are parsed and reconciled against the post list, so publishing a post without wiring it up everywhere gets caught. Not much point in publishing a new blog that isn't shown anywhere!

Headless smoke tests

The audit reads files; it cannot tell you whether the page actually works. This site injects its header, footer, comments, and related posts with JavaScript, and it animates sections into view as you scroll. A one-character mistake in that JavaScript can leave a page technically valid and completely blank, and the audit wouldn't catch it.

So the second half is a smoke test built on Playwright, which drives a real, invisible Chromium. The script serves the repository over a local web server, loads six representative pages, waits a few seconds for the JavaScript and animations to settle, and then asserts on what a visitor would actually see:

# A few of the checks, one page each:
# index.html   - hero present, live-stats skeletons cleared, blog cards visible
# blog.html    - at least 3 post cards and 3 tag filter buttons rendered
# gallery.html - at least 12 photos in the grid
# a blog post  - body text visible (opacity != 0), table of contents built
# quotes.html  - at least 500 quote cards
# map.html     - the map initialised with at least 10 markers

The site reveals sections with a fade-in animation, which means a JavaScript failure doesn't remove content, it leaves content sitting in the page at opacity zero: present in the HTML, invisible to humans. That is the failure a static checker can never catch. The check is a Python tuple holding a line of JavaScript, evaluated inside the live page:

("post body visible",
 "(() => { const p = document.querySelector('.blog-post > p');"
 "  return p && getComputedStyle(p).opacity !== '0'; })()"),

The test also collects console errors, ignoring the third-party noise (analytics, embeds) and failing on anything from my own code.

The robot that runs it all

Both halves run in GitHub Actions on every push. The workflow has two jobs: one installs Python and runs the audit plus a check that the minified stylesheet is up to date (the CSS build is its own story); the other installs Playwright and runs the smoke tests. If either fails, I get an email and a red X on the commit within a few minutes. A local pre-commit hook runs the fast checks too, so most mistakes never even reach GitHub.

Is this overkill for a personal site? No: it's because the site is a hobby that I need automation. Nobody is paid to notice when this site breaks (or, you know, paid for writing blogs, creating the site, etc.). There is no QA department, no on-call rota, just me, and I would rather spend my evenings writing posts than manually clicking every link on fifty pages. The test suite is what lets a one-person site behave like it has staff.

The suite also enforces my rules on me. Zero errors and zero warnings sounds strict, but a clean baseline means any non-zero number is news.

What it still misses

The audit checks that every referenced image exists on disk. It does not check that the image is committed to git. I published a post whose hero image existed on my machine but had never been committed, every local check passed, and the live site served a grey void where a seascape should have been. A reader (fine: me) spotted it and fixed it later that day, but that's not great when I already had a few dozen readers.

Each gap you find becomes the next check you write. The fix here is a rule that any local file referenced by a tracked page must itself be tracked, which is one more function in the audit script. Here is that check as it runs today (lightly trimmed):

def check_target(url, line, what="target"):
    local = resolve_local(page, url)
    if local is None:
        return  # external URL, checked elsewhere
    if not local.exists():
        add("ERROR", page, line, "broken-link", f"missing {what}: {url}")
    elif page_is_tracked:
        relp = local.resolve().relative_to(ROOT).as_posix()
        if relp not in tracked:
            add("ERROR", page, line, "untracked-ref",
                f"{what} exists locally but is not tracked by git: {url}")

Fittingly, that check fired while I was preparing this very post: the hero image at the top of this page existed on my machine and had never been committed. Test suites are never finished, only extended.

Common questions

Isn't this what link checkers and Lighthouse already do?

My Teacher Said I Shouldn't Go To University

Thu, 16 Jul 2026 00:00:00 +0000

My Teacher Said I Shouldn't Go To University

16 July 2026 · personal advice

In high school, my music teacher told me I shouldn't aim to be a doctor, because I was "more suited to manual labour". In fact, she said that not only to me, but to my parents, at a student-parent night.

I want to be fair to her, but I am writing this with a doctorate. Not the kind she meant, admittedly, a PhD makes you the sort of doctor who is useless in a medical emergency but insufferable when anyone brings up a topic vaguely related to your specialty, yet my title is real, my thesis was dissected by a bunch of people with way more experience than I, and somewhere inside, I still remember being told I wasn't good enough, and couldn't be good enough, by a trusted advisor.

Teachers make thousands of judgement calls, and try to direct their students in the direction best for their displayed aptitudes, skills, abilities and what they appear to enjoy. How did she come to decide I was more suited to working with my hands? A music grade (where, by the way, I scored highly on performance because I loved guitar, and poorly on other parts of the classwork), a snapshot of who I am, and a very narrow idea of what my capabilities looks like, in a very specific field. How can a music teacher say "Well, this student kind of sucks at identifying Baroque music from Rococo, so they probably should be a plumber"? Besides the permanent glare I adopt whenever I think of her, what was the real damage? Well, I was lucky, because teenagers today still experience this kind of snapshot judgement, and the research says it does more than describe futures for kids. It writes them.

Quick jargon guide

Teacher expectation effects: the finding that what a teacher believes about a student can nudge that student's actual outcomes, famously via the "Pygmalion" study. The size of the effect is debated; its existence, much less so.
Golem effect: Teacher expectation effect's ugly sibling, low expectations dragging performance down.
Academic self-efficacy: a student's belief that they can succeed at academic work, one of the strongest psychological predictors of how they actually do.
g (general intelligence): the well-replicated statistical finding that performance across mental tasks correlates. Real, useful, and still only part of the picture any single grade captures.
Multiple intelligences: Gardner's popular theory that intelligence comes in eight-ish flavours. Scientifically shaky as psychometrics, but pointing at something real about the plurality of human skill.

What she was looking at

Whatever my teacher saw when she looked at me, I'm guessing it was something to do with grades from when I was forced to play glockenspiel. That thinking is the design of the system she worked in. School distils each pupil into a thin column of numbers generated by one activity: sitting still, absorbing material in a fixed order at a fixed age, and reproducing it silently on paper (or by bashing thin metal plates in a specific order and rhythm). Do that well and you're "academic." Do it badly, or do it later than your teens, because adolescent brains run on wildly different schedules, and the system files you as "not academic".

Grades are a decent-but-leaky predictor even of the thing they're supposed to predict: meta-analytic work like Richardson, Abraham and Bond's review^[1] finds prior grades correlate moderately with university performance, leaving an enormous share of the outcome to everything else: effort regulation, study strategies, tutoring, circumstances, and, notably, academic self-efficacy, whether the student believes they can do it. Sounds a bit weak, right?

Prophecy is an intervention

If a teenager's belief in their own academic capability is one of the strongest psychological predictors of their university performance, then look again at what a sentence like "you're more suited to manual labour" actually is: it's an intervention on the predictor variable. This isn't a guess, it's changing the variables that cause the outcome.

This is the territory of the famous Pygmalion study,^[2] where teachers told (falsely) that certain randomly-chosen pupils were about to bloom saw exactly those pupils gain the most, and of its grim mirror, the golem effect, where low expectations do the opposite work. The literature has spent fifty years arguing about effect sizes, which is valid: expectation effects are usually modest, on average. A stray remark from an authority figure, landing on a fifteen-year-old at the exact age when identity is wet cement, doesn't get experienced as a data point but as a verdict, and some of us can quote it verbatim decades later, which should tell you something the impression such remarks can make on kids. That is to say: it may not always stick, it likely depends on the relationship the kid has with the teacher, the emotional vulnerability and circumstances of the kid, and counterweights, but it did with me, and it does with others.

I got lucky: enough stubbornness to treat the verdict as a provocation, and enough people elsewhere in my life balancing it out (thanks Mum + Dad). A different kid (same ability, fewer counterweights) hears the same sentence and might be dissuaded entirely from whatever their goals are.

The many shapes of capable

I've previously taken a flamethrower to pop-psychology taxonomies of the mind, and I intend to stay consistent. Howard Gardner's "multiple intelligences" (the theory that we each carry separate musical, spatial, interpersonal, bodily intelligences) is, as psychometrics, shaky: the abilities correlate, the categories resist clean measurement, and the framework has never out-predicted boring old g. If you came here for "everyone is secretly a genius in their own modality," you are reading the wrong article.

Even mainstream intelligence research describes ability as a hierarchy: a general factor, yes, but with broad group abilities (verbal, spatial, quantitative, mechanical) and thousands of specific skills layered beneath, all of it interacting with interest, practice, temperament, and time. A school grade samples one thin horizontal slice of that, once, at an age when the whole edifice is still under construction and might be entirely changed because the kid was going through something that morning. The carpenter's spatial reasoning, the care worker's social perception, the mechanic's fault-finding, the musician's ear: these are real, demanding, trainable forms of capability that grades just don't measure at one time, nevermind over time.

The insult in "suited to manual labour" was never to me, it was to manual labour. The sentence only works as a put-down inside a worldview where the trades are a punishment tier for the insufficiently bookish, or for people who have ADHD or similar difficulties in the neurotypically designed classroom, a worldview that is snobbish about the people who build its houses and wire its schools, and that pushes academically-shaped kids away from skilled work they might have loved.

There's a gendered layer to this, too. I recently read Boys Don't Try? Rethinking Masculinity in Schools by Matt Pinkett and Mark Roberts,^[3] two teachers examining why boys collect verdicts like mine so reliably. A lot of what gets filed under "boys don't try" is self-protection (if you never try, failing can't become evidence about you), and the fix is not lowering the bar to match the stereotype but refusing the stereotype altogether: high expectations, warm relationships, and no prophecies. I'd hand it to any teacher who has ever caught themselves sorting a class into who is and isn't "university material".

If you were told similarly:

A grade measured how you did one kind of task, in one format, at one age, under whatever was happening at home that year. It contains no information about you at 20, 30 or 40.
Self-efficacy isn't magic thinking; it's the thing that determines whether you apply, persist, and ask for help.
Access courses, mature entry, apprenticeships that become degrees, degrees that start at 30. The system's front door has a strange obsession with your teens, it's hardly the end.
If the trades, or care, or making things is where your capability lives, that's not a consolation prize, that's the goal that society doesn't recognize as success (and considering all the "white collar" job difficulties, the trades are a very lucrative, respectable and desirable goal).

To my teacher, if she's reading

If she's still alive, I doubt she remembers saying it. That's just it though - the sentence that lodged in me for decades probably cost her four seconds and was immediately forgotten. I hope her years of teaching had better impacts on kids than that one remark, and I'd genuinely rather this post reach some teacher mid-career than reach her. So, to that teacher: you hold a variable that the meta-analyses rank near the top of the stack, and you adjust it every time you tell a teenager what they are. Forecast less, care more. You will be wrong about which ones bloom (the research says so, my doctorate says so, too) so you might as well be wrong in the direction that costs nothing and occasionally builds a doctor from a kid who wants to be a doctor.

Common questions

Do you resent her?

Free Comments via GitHub Discussions

Wed, 15 Jul 2026 00:00:00 +0000

Free Comments via GitHub Discussions

15 July 2026 · technology

Scroll to the bottom of this post and you'll find a comments section. That shouldn't be possible: this is a static site, a folder of HTML files with no server, no database, nowhere to put a comment. The trick is a wonderful little disguise: every comment thread on this blog is a GitHub Discussion, courtesy of a small open-source tool called giscus. This post details what it is, its setup, and its trade-offs, as part of the series on how this site is built.

Quick jargon guide

giscus: a free, open-source widget that displays a GitHub Discussion as a comment section on any web page.
GitHub Discussions: forum-style threads attached to a GitHub repository, normally used by software projects for Q&A.
iframe: a page embedded inside another page. The comments you see are technically a small giscus page shown in a window within mine.
Static site: a website that is only files, with no server-side code, which is why it can't store comments itself.
Widget/embed: third-party functionality you add with a script tag.

How it works

My blog's repository has Discussions enabled, with a category set aside for comments. When you visit a post, a script builds the comments section and giscus looks up a discussion whose title matches the page's path: for this post, /blog/free-comments-via-github-discussions.html. If someone has commented before, the thread exists and gets displayed. If nobody has, giscus shows an empty comment box, and the discussion is created automatically the first time someone posts.

Commenting requires a GitHub account, which is the design's biggest filter and, depending on your audience, either a bug or a feature. For a blog like mine, whose readers skew technical, it's mostly a feature: sign-in deters drive-by spam so effectively that I do no spam moderation at all. For a blog aimed at a general audience it wouldn't be the best solution, and something like a hosted comments service (or no comments, with an email link) might fit better.

The setup

Enable Discussions on the repository (a checkbox in settings) and create a category for comments; giscus recommends the Announcements type, so that only the widget and maintainers can open new threads, keeping the category from becoming a second inbox. Install the giscus app on the repository from GitHub's app directory. Then the giscus website generates your embed code: you tell it the repository, category, and how pages should map to discussions (I use the page path), and it produces a script tag with everything encoded as attributes.

Most sites paste that script tag into their template and are done. Mine goes through the shared components file, which builds it dynamically on every post page. Ez pz.

The details

Matching the site's theme. giscus ships with stock light and dark themes, but it also accepts a custom stylesheet URL, so this site serves its own giscus-dark.css and giscus-light.css and the comments render like part of the page rather than a bolted-on box. The fiddly bit is the theme toggle: when you switch the site between dark and light, a message gets posted into the giscus iframe telling it to swap stylesheets too, so the comments change theme in step with everything around them.

Lazy loading. The widget loads only when you scroll near it, so readers who never reach the bottom never fetch it. Comments are the last thing on the page both location wise and in load time!

A refresh button. The embed is static once loaded: if someone comments while you have the page open, nothing updates. A small refresh control re-mounts the widget on demand.

The trade-offs

If GitHub discontinued Discussions, or giscus vanished, my comment sections would go blank (the data would survive, in the discussions themselves, and giscus being open source means it could be self-hosted). Comments also don't work offline, and they're invisible to readers with JavaScript off; my offline post covers why the service worker deliberately leaves them alone. This is fine for a personal blog: comments are a nice-to-have on top of the writing, not the point of it.

What's nice is there are no ads injected next to your readers' words, no tracking scripts profiling them, no database to secure, and moderation through tools (GitHub's) that are already set up and hosted for you. The comparison is with Disqus, the long-time default, is pretty clear: disqus is filled with ads, and filled with tracking scripts etc. I know which one I prefer!

Common questions

Can you delete or moderate comments?

Putting My Photos on a Map

Tue, 14 Jul 2026 00:00:00 +0000

Putting My Photos on a Map

14 July 2026 · technology photography

My photography spans two countries I've lived in and a handful I've visited, and a plain grid of thumbnails hides that storytelling, sadly. So, I decided to try my hand with a photo map: an interactive world map where each marker is a place I've photographed. Click a marker and you get a popup of thumbnails; click a thumbnail and the whole region opens as a full-screen slideshow. This is the how to, as part of the series on how this site is built.

Quick jargon guide

Leaflet: the standard open-source JavaScript library for interactive maps. It draws the map, handles panning and zooming, and places markers; it's the engine under a huge share of the web's non-Google maps.
Tile server: where the actual map imagery comes from. Maps are served as a grid of small square images ("tiles") drawn on demand as you pan and zoom.
Geotagging: attaching a location to a photo, either automatically (phones embed GPS in the file) or by hand (me, squinting at a map, going "that was definitely Falkirk").
Marker / popup: the pin on the map and the bubble that opens when you click it.
Deep link: a URL that opens the page in a specific state, like the map already zoomed to one region.

The data

The obvious way to geotag photos is to read the GPS coordinates cameras embed in their files. My photos mostly don't have any: they come from a real camera without GPS, not a phone, and even when coordinates exist I strip metadata from published files, usually due to not being super careful about backing up my photos over the years, so some of these are screenshots from websites I put them (GuruShots!). So the map's data is a single hand-maintained JSON file: a list of 29 named regions, each with one latitude/longitude pair and the list of photo numbers taken there, 239 photos placed so far.

{
  "name": "Falkirk",
  "lat": 56.0011, "lng": -3.7835,
  "photos": ["135", "141", "197", "..."]
}

Placing photos by hand sounds like a chore and mildly is but it's not awful either. A region pin says "around Falkirk", never "this street, this house, this ruin", which matters both for the abandoned places (some locations shouldn't be advertised) and for anywhere near where people live. And meaning: "University of Stirling" and "Ochil Hills" are how I think about where photos were taken, not coordinates.

The map

The mapping library is Leaflet, and in keeping with house policy the library itself is served from this site, not a CDN. The tiles (the map imagery) are external however, coming from CARTO's free basemaps. There are two tile styles, a light one and a dark one, and the map swaps between them when you toggle the site's theme.

The map as it stands: 29 regions, 239 photos placed. Tiles © CARTO / OpenStreetMap contributors.

Markers, popups, and the slideshow

Each region renders as one circular marker showing its photo count, its size scaled gently with the number (34 to 52 pixels, so Falkirk's 28 photos read bigger than Dublin's one without turning into a bubble chart). I skipped marker-clustering plugins entirely: hand-curated regions are the clustering, done once, with better names than any algorithm would pick.

Clicking a marker opens a popup with the region's name and up to nine thumbnails; clicking any thumbnail opens the region's entire photo set as a lightbox slideshow, starting from the one you picked. The thumbnails are the same small WebP files the gallery uses, and the slideshow pulls the full-resolution originals from the GitHub release where they live. The map also cooperates with the rest of the site: gallery photos with a known place show a location label linking to the map pre-zoomed on that region, via a small ?region= parameter in the URL, so the command palette / search function (CTRL + K) can show you places on the map when you search for, say, "East Lansing".

If you want one of these

The real hurdle isn't the code, it's the data. Leaflet's tutorial gets a map with markers running super quickly; deciding where 239 photos were taken took considerably longer, spread over batches (my git history contains the commit "Geotag batch: Falkirk, Falkland, Lansing added; 63 more photos placed", which tells you how it actually went). Start coarse: a dozen regions, big radii, photos you're sure about. I'll add more when I'm running low on bigger ideas to add to the site!

If your photos do have GPS data (phone photographers, this is you), your version of the data file can be generated by a script reading the coordinates, snapping them to a sensible grid, and naming clusters after the nearest town. Consider still doing the naming pass by hand; "Grandma's village" beats "cluster_17" forever. And whatever you do, check what location data you're publishing before you publish it: the same GPS metadata that makes the map easy also tells the internet where you sleep.

Common questions

Why not just use the tags in the gallery for places?

Self-Hosting Your Fonts

Sun, 12 Jul 2026 00:00:00 +0000

Self-Hosting Your Fonts

12 July 2026 · technology

The text you're reading is set in Poppins (no, not Mary Poppins), the serifs in the pull quotes are Lora, and until recently both arrived from Google's servers, the way fonts arrive on roughly half the web. Now they're served from this site itself: twelve small font files sitting in the repository next to the photos. This post is just a brief rundown on how and why I did this, and how I shrank three icon fonts by about 97% with a Python script. Part of the series on how this site is built.

Quick jargon guide

Web font: a font file the browser downloads with the page, so visitors see your chosen typeface rather than whatever their device has installed.
WOFF2: the modern compressed format for web fonts. Universally supported, roughly a third the size of older formats.
@font-face: the CSS rule that tells the browser "this family name maps to this font file". Self-hosting is mostly writing these by hand.
Icon font: a font whose "letters" are pictures (the GitHub logo, a house, an arrow). A common way to ship icons before SVG took over.
Subsetting: deleting every glyph from a font that your site doesn't use. A font with 700 icons of which you use 5 can lose the other 695.
Unicode range: the span of characters a font file covers. Serving just the Latin range is itself a subset: no Cyrillic or Vietnamese glyphs for a site written in English.

The why

Well, mainly because I wanted to learn about fonts. I think they're interesting: they communicate tone, a certain visual 'feel' that goes beyond the words on the page, they have interest accessibility benefits and they have some fascinating history. But, the technical reasons are: Privacy: every page view against Google Fonts tells Google's servers a visitor's IP address requested your page, and in 2022 a German court found exactly that arrangement violated the GDPR. Serving fonts yourself means your readers talk to nobody but you. Reliability and speed: a third-party font server is one more thing that can be slow, blocked (Google is unreachable in some countries), or discontinued. Self-hosted fonts arrive over the same connection as everything else. Ownership: the same philosophy as the rest of this site. Fewer moving parts owned by other people.

The counterargument used to be caching: the theory that everyone has Google's copy of Poppins cached already, so linking to it is free. That theory died in 2020 when browsers partitioned their caches per site for privacy reasons. Every site now downloads its own copy regardless, so the shared-cache benefit is actually negative, since you download it repeatedly.

The how

Google Fonts provides them, of course. Download the family, keep the WOFF2 versions of the weights you actually use, and write an @font-face block per file:

@font-face {
  font-family: "Poppins";
  font-style: normal;
  font-weight: 400;
  font-display: swap;
  src: url("../fonts/vendor/poppins-400.woff2") format("woff2");
}

This site carries seven weights of Poppins and five faces of Lora (the italics matter for pull quotes), Latin subset only, for a grand total of about 244 KB, less than a single photograph thumbnail-and-hero pair. The font-display: swap line is the one non-obvious part: it tells the browser to show text immediately in a fallback font and swap when the real one arrives, rather than showing nothing. Neat, right? Invisible text while a font loads is way more irritating than a brief flash of the wrong font that most people won't see.

Poppins, rendered from the same woff2 files this page just loaded. Seven weights, 54 KB all told.

Lora's five faces. The italics earn their bytes in the pull quotes.

The fun part: subsetting icon fonts

The site's template arrived with three icon fonts (FontAwesome, Themify, ElegantIcons), each carrying hundreds of icons. A count of what the site actually uses came to: five FontAwesome icons, sixteen Themify icons, and two ElegantIcons.

Using the fontTools library, a Python script scans every page and script on the site for icon class names (fa-github, ti-arrow-up, and so on), parses each vendor stylesheet to map those class names to the character codes inside the font, and then writes a new font containing only those characters. The results are satisfying in a way few optimisations are:

# Original -> subset (woff/woff2)
# FontAwesome   75.4 KB -> 1.5 KB   (5 icons kept,  ~98% smaller)
# Themify       54.8 KB -> 2.7 KB   (16 icons kept, ~95% smaller)
# ElegantIcons  62.2 KB -> 1.2 KB   (2 icons kept,  ~98% smaller)

The subset files live alongside the originals, and the CSS build swaps the font sources to the subsets at minification time, so the originals remain on hand for the day a new icon appears. That day is the obvious failure mode (use a new icon, forget to re-subset, icon renders as an empty box), which is why the script also writes a manifest of what's included. I have some checks in place to make sure this goes smoothly, and that will be in another blog on how "my website has a test suite".

A note on SVG icons

Would SVG icons be more modern than icon fonts? Yes, and a fresh site should use them. But the template came with icon fonts wired into everything, and three kilobytes of subset font is not a problem worth a rewrite. So, why not optimise the thing you have.

Common questions

Is self-hosting Google's fonts legal?

Reading This Site Offline

Sun, 12 Jul 2026 00:00:00 +0000

Reading This Site Offline

12 July 2026 · technology

This site learned a small trick recently: it works without the internet. I implemented it out of curiosity on how it would be done, but I'm glad I did. If someone tries to read one of my blogs while on a plane, now they can, so long as it was preloaded. Read a few posts over breakfast, open the site again, and the pages you visited are still there, styled, functional. If the connection dies mid-browse, it will still work. Nice, right?

This is done via a service worker, and this post walks through mine: ~100 lines of code, no framework, no build step, in keeping with this site's general philosophy of doing things with the smallest possible amount of infrastructure. If you run a static site anywhere (GitHub Pages included), this is a good philosophy, in my opinion.

Quick jargon guide

Service worker: a small script the browser installs alongside your site and runs in the background. It sits between your pages and the network, and can answer requests itself, including when there is no network.
Cache (the browser kind): a local store of responses the service worker can save and replay. Lives on the visitor's device, controlled by your code.
Network-first: a strategy: try the internet, fall back to the cache. Fresh when online, functional when not.
Stale-while-revalidate: the inverted strategy: answer instantly from cache, then fetch a new copy in the background for next time. Fast always, fresh eventually, annoying when you are refreshing a page for an update, though.
Precache: the short list of files saved at install time, before you've visited anything, the skeleton the site can't render without.
Cache invalidation: famously one of the two hard problems in computer science. A service worker is a machine for having this problem on purpose.
giscus: the comment system under my posts. It stores every comment as a GitHub Discussion and loads in an embedded frame from giscus.app, so the conversation lives on GitHub rather than on this site.

Two strategies

Everything the worker does comes down to one decision, made per request: who do you trust more, the network or the cache?

Pages get network-first. HTML is where mistakes live: a typo fixed, a broken link repaired, a post updated. When you're online I always want you reading the newest deploy, so the worker tries the network, saves a copy of whatever comes back, and only reaches for that copy when the fetch fails:

// HTML: network-first
fetch(req).then(function (res) {
  var copy = res.clone();
  caches.open(PAGES_CACHE).then(function (c) { c.put(req, copy); });
  return res;
}).catch(function () {
  return caches.match(req).then(function (hit) {
    return hit || caches.match('./offline.html');
  });
});

The chain reads as a 'politeness ranking': fresh page if possible, your cached copy if not, and a dedicated offline page as the final backup. Every page you visit while online becomes a page you own while offline, the cache is your personal reading history.

Assets get stale-while-revalidate. CSS, JavaScript, fonts, thumbnails, these change rarely and block rendering while they load, so the priorities invert. The worker answers from cache immediately and refreshes in the background:

// Assets: stale-while-revalidate
caches.match(req).then(function (hit) {
  var fetching = fetch(req).then(function (res) {
    if (res && res.status === 200) {
      var copy = res.clone();
      caches.open(ASSETS_CACHE).then(function (c) {
        c.put(req, copy);
        if (req.destination === 'image') trimCache(ASSETS_CACHE, IMG_LIMIT);
      });
    }
    return res;
  }).catch(function () { return hit; });
  return hit || fetching;
});

Worst case, you see a stylesheet that's one visit out of date. In exchange, repeat visits render instantly, offline or not. That trimCache call is the one piece of housekeeping, meaning that image caches are capped at 200 entries, oldest evicted first, so browsing my whole gallery doesn't fill up your phone!

What I deliberately don't cache

The worker ignores anything cross-origin; the full-size photographs stay on GitHub's release servers, caching those would mean warehousing megabytes per photo on your device for a click-through you probably won't repeat. Comments are giscus, which lives in an iframe and belongs to GitHub; offline, it simply doesn't appear, which is the correct behaviour. Analytics likewise gets no offline resurrection, if the network can't see you, neither should it.

Offline mode should preserve the reading, not fake the internet. The precache reflects the same idea: it's just the offline page, the stylesheet, the core scripts, and posts.json (so the blog index still works), so about eleven files, everything else earns its place in your cache by you actually opening it.

The footguns

I'm not really a website developer or designer, beyond hobbyist endeavors like this website and a couple of landing pages for research labs I worked in, so much of this section is just what I learned while, well, learning this stuff. Service workers have a deserved reputation for one specific misery: the cache that would not die. Ship a worker with a careless cache-first strategy and your visitors can be pinned to an old version of your site for days: including, delightfully, an old version of the service worker itself. The classic developer experience is editing a file, refreshing, seeing no change, and spending an hour debugging to find it's actually working as intended and you just turn off your monitor for a break and see your sad reflection staring back at yourself. Ahem, anyway.

I'll also admit what this isn't: it isn't a full progressive web app. There's no install banner, no background sync. Those are all possible and all, for a site whose job is being read, beside the point.

Try it

Open a few posts, then turn on airplane mode and keep clicking. The pages you visited load; the search on the blog index still filters; the pages you didn't visit hand you a courteous offline page instead of a browser error. Turn the network back on and the whole arrangement dissolves back into an ordinary website. Much like an IT worker, if it's doing the job correctly, you won't even notice it's there.

Common questions

Why not just cache the entire site up front?

The Abandoned Places I Photograph

Sun, 12 Jul 2026 00:00:00 +0000

The Abandoned Places I Photograph

12 July 2026 · photography personal

There's a filter on my gallery labelled "abandoned," and it's one of the used tags for my shots: dozens of photographs of peeling wards, graffitied corridors, staircases climbing into darkness. This is a rundown of what draws me to these places, the story of the category I photographed most, and the ethics of pointing a camera at somewhere truly abandoned, often forgotten, but surviving locations preserving history. Photographers call this "urbex".

Quick jargon guide

Urbex: urban exploration, visiting and documenting abandoned or hidden built places. Ranges from respectful photography to (not endorsed here) breaking and entering.
"Take nothing but photographs": the urbex code of ethics, usually finished with "leave nothing but footprints." Nothing gets moved, taken, broken, or tagged.
Ruin porn: the critical term for decay photography that aestheticises a place's collapse while ignoring the people it happened to. A charge worth taking seriously, which I try to below.
Village system: a late-19th-century model for psychiatric hospitals: instead of one grim block, a self-contained village of villas, workshops, and farmland, meant to give patients ordinary life at a humane scale.

Bangour, before the diggers

The abandoned place I photographed most is Bangour Village Hospital, in the countryside west of Dechmont in West Lothian, Scotland. It opened in October 1906 as the Edinburgh District Asylum, built on the continental "village system": not a single institution looming over its inmates, but a scatter of villas across parkland, with its own power station, workshops, bakery, kitchen, laundry, and eventually its own church and railway. The idea, radical for its time, was that people in psychiatric care should live somewhere shaped like a life. It's something of a rarity to look at history and feel a bit of pride, but here's a nice example (click below for better quality).

History kept interrupting the little care village. The War Office requisitioned the site in both world wars: during the second, a specialist burns and plastic surgery unit set up there in 1940, seeding what became Bangour General Hospital next door. Psychiatric care returned between and after the wars and wound down slowly for decades; the last ward closed in 2004. Then, for twenty years, the village was empty, a listed Edwardian ghost town, which is when I walked its corridors with my Canon 60D.

I was not the first to see the beauty in the ruins. The film industry found the empty hospital too: it stood in as the asylum in The Jacket (2005), which put Adrien Brody and Keira Knightley into the same corridors, with better lighting and a catering truck. There's something fitting about a place built to be a village ending its working life playing one on camera, and the location scouts and I clearly agree on what those wards look like: somewhere between memory and horror, depending on the light.

Poster for The Jacket (2005). © Warner Bros

The window has now closed: the site is becoming housing. West Lothian Council approved nearly a thousand homes in late 2024, and construction started in 2025. I have complicated feelings about this, homes are a better use of land than entropy, and yet - it's a beautiful place I wish I had more time to visit. Every photograph in my Bangour set changed category the day the first digger arrived: they used to be pictures of a place, and now they're records of one. That's the strange dividend of this hobby. You think you're taking moody photos; you turn out to have been doing amateur archival work. This is, I realise, a theme I frequent.

Why ruins photograph so well

Partly it's light: broken roofs do things with shafts of sun lighting up rooms in ways that are unnatural, or perhaps weirdly natural in the artificial places. Partly it's texture, forty years of peeling paint and ferns growing from between tiles speaks to us about time, and how in a blink of an eye things become and become undone.

There's also levels of age, eras shown in rot. Graffiti and ancient beer bottles show that others have explored these ruins when they were perhaps fresher, and the security doors and warning signs about CCTV cameras make it feel forbidden.

The ethics of photographing decay

"Ruin porn" is a reasonable charge, and the only defence is conduct. The photographs in this post exist because a place was open and I walked in with a camera. Three rules cover what many of us urbex photographers adhere to.

Take nothing but photographs. Nothing gets pocketed, moved for a better composition, or "rescued." The scattered papers in that corridor are still scattered exactly as I found them. This is, in part, to not make it obvious I intruded upon the place, so that security isn't tightened, preventing others from experiencing it. It's also something of respect for the natural collapse of a place, allowing us to capture this descent, to see the "what if humans just disappeared one day?" question with visuals.

© Ken Reid. All rights reserved.
Force nothing. If a place isn't open to walk into, the visit doesn't happen. No cut fences, no pried boards, no "it was already broken so." Beyond the obvious legal line, and the law varies enough by country that you should know yours before you go, forcing entry is what turns documentation into damage.

© Ken Reid. All rights reserved.
Don't publish the way in. I'll name Bangour because it's famous, documented, and now a construction site; I don't share access details for anywhere fragile. Every ruin that goes viral with directions attached gets stripped and burned within the year.

I'll add a fourth, more personal one: be safe, and bring a friend. If you're somewhere abandoned, you don't know who you'll find there, and walking around with a four to five figure costing camera may lead to desperate people taking something you'd rather keep.

Why bother?

Because the maintained world is thoroughly photographed and the unmaintained one is disappearing. Because it answers that previous question about what if humans disappeared. Because there's history in old places, stories and a silence that can somehow capture with a camera.

Respect

A psychiatric hospital is not a neutral ruin. Thousands of people lived entire lives at Bangour: some helped by it, some failed by it, in an era when society's record on mental illness was grim. Photographing their ward as a spooky backdrop would be a small desecration, at least without recognizing it for what it is. The village system deserves to be remembered as what it was: an attempt, flawed but with good intentions, to build kindness at institutional scale.

An office in the dark, partitions still standing to attention, ceiling tiles mid-collapse. Whatever the last shift was working on, they left much of it on the shelves and the desks.

For the people nearby. Ruins have neighbours: farmers whose fences get climbed, families whose street fills with strangers after a site trends, a security guard paid too little to argue with anyone. Being polite and willing to leave when asked costs a photographer nothing and buys the whole hobby its tolerance.

© Ken Reid. All rights reserved.
For the dead and the grieving. Hospitals, asylums, and churches intersect with the worst days of real families, some of whom are alive and searching the same hashtags you post under. Shoot and write with respect.

© Ken Reid. All rights reserved.
For the building itself. No tagging, no smashing "for the shot," no souvenirs.

© Ken Reid. All rights reserved.
For other explorers. Passing on a site's condition, hazards, and history is generosity.

Safety

None of the above matters if the floor collapses under you. A decaying building is indifferent to your intentions, and the most common dangers are not the stranger in the dark, it's the joist that rotted through in 2011. The working checklist, assembled from cautious people and a couple of my own near-misses:

Floors and stairs lie. Rot hides under intact-looking boards, and a staircase that held the last visitor has one fewer life left than it did. Stay near walls where the structure is strongest, test before trusting.

© Ken Reid. All rights reserved.
The air has history. Buildings of Bangour's era mean asbestos, lead paint, pigeon droppings, and mould in quantities you cannot see. Disturb as little dust as possible, skip the crawl spaces. A decent mask is always a wise decision.

© Ken Reid. All rights reserved.
Go in daylight, leave a margin. Light is the photographer's excuse, but it's also the difference between seeing the missing floorboard and finding it. Arrive with hours to spare and leave while you can still find the exit without a torch. Bring a torch, just in case, anyway.

© Ken Reid. All rights reserved.
Never alone, always announced. Two people minimum, and someone at home who knows the site and when to expect a check-in. A twisted ankle in an empty building with no signal is a story with company and an emergency without.

© Ken Reid. All rights reserved.
Dress for the building, not the photos. Boots with soles that shrug off nails, gloves, long sleeves. Tetanus sucks.

© Ken Reid. All rights reserved.
Know when to stop. Fresh collapse, a smell of gas, sounds of people who don't want company, or the plain sense that something is wrong: any of these ends the visit.

Twenty photographs, several buildings, one country. Scotland is unusually rich in these places: industrial decline, rural depopulation, and a century of ambitious institutional architecture left it with more empty grandeur per square mile than you would think.

Common questions

Is urban exploration legal?

Why I Love LaTeX

Sun, 12 Jul 2026 00:00:00 +0000

Why I Love LaTeX

12 July 2026 · technology writing

My PhD thesis was written in LaTeX, and so were my papers. So, to the muted horror of every recruiter who has ever asked for "the Word version," is my CV. I have spent a meaningful fraction of my adult life inside a markup language originally released in the 1980s, and I am here to tell you it was not Stockholm syndrome, but respect and, I admit, laziness. LaTeX, once you have it down, is so much easier to use than WYSIWYG editors.

Quick jargon guide

LaTeX: (pronounced "lay-tech", definitely not "lay-tecks") a typesetting system: you write plain text with commands like \section{...}, and a compiler turns it into a polished PDF. The standard tool of mathematics, physics, and computer science publishing.
Typesetting: the craft of arranging text on a page: spacing, line breaks, hyphenation, where the figures land. What printers did by hand for five centuries.
Markup: instructions written in the text, about the text: "this is a heading," "emphasise this." As opposed to clicking a toolbar and hoping.
WYSIWYG: "what you see is what you get". Word, Google Docs. You edit the final appearance directly, which sounds obviously better except (it isn't).
Compiling: running the program that converts your marked-up text into the finished document. Comes with error messages best described as "vintage."

The separation that changes everything

LaTeX's founding idea is that what you're saying and how it looks are different jobs, done best by different parties. You handle the words and the structure: this is a chapter, this is a definition, this deserves emphasis. Design is handled through clearly defined rules that generally affect the whole document at once: spacing, hyphenation, figure placement, the numbering of every section, equation, and reference. In Word, this is a standard experience:

It's the same principle that makes good data science work, too: separate the content from the presentation and both improve. My thesis had hundreds of numbered things: figures, tables, theorems, references to all of the above. Every single number was generated, every cross-reference resolved automatically, every citation formatted from a bibliography file I never once styled by hand. I've known people to fight their word processors for whole afternoons over master's theses a third the length. I fought mine too, but it's often a fight fought once per type of issue, not every citation, and every plot.

It's just text

A LaTeX document is a plain text file, which also allows for:

Version control. My thesis lived in git, like code, because it was code. Every draft diffable, every deleted paragraph recoverable, every "what did I change since my supervisor read this" answerable.
Longevity. The .tex files from decades ago still compile. No format lock-in, no "this document was created in a newer version," no ransom paid to any vendor.
Tooling. Search it, script it, generate it. When my experiments produced results, code wrote the tables directly into the document. No transcription, no transcription errors.
Focus. A text file cannot distract you with 40 fonts. The absence of formatting options while drafting is not a limitation; it's the whole trick.
Frugality. You don't need an Office 365 license to use LaTeX, it's completely free if you run it locally.

It looks like this:

\documentclass[11pt]{report}
\usepackage{amsmath}
\usepackage[backend=biber]{biblatex}
\addbibresource{thesis.bib}

\begin{document}

\chapter{Introduction}
Scheduling problems are, politely put, everywhere.
As shown in Section~\ref{sec:motivation}, the real
question is not whether to optimise but what to
optimise \emph{for}.

\end{document}

The Output

LaTeX documents look right, obvious once you've compared them side by side with the alternative: the line spacing, the justified text with proper hyphenation, the kerning, the ligatures, mathematics set the way mathematics should look, without the nightmare-inducing Word equation editor. Knuth built the underlying engine, TeX, because he was (rightly) offended by how his own books were being typeset, and that offence, refined over decades, is now free software! When my thesis came back from the binders it looked like a real book, because by every standard that matters it was typeset like one.

Getting started

Rather than describe it further, here is a complete LaTeX document. It shows the three things beginners care about: a title block, mathematics, and a table.

% Getting started with LaTeX - the whole document, no hidden parts.
\documentclass[11pt]{article}

\usepackage[a4paper, margin=2.5cm]{geometry}
\usepackage{amsmath}
\usepackage{booktabs}

\title{Getting Started with \LaTeX}
\author{Ken Reid \\ \texttt{kenreid.co.uk}}
\date{July 2026}

\begin{document}
\maketitle

\section{How this works}
You are reading a PDF that was generated from about forty
lines of plain text. Headings are numbered automatically,
and nothing on this page was nudged into place by hand.

\section{Mathematics, the party trick}
Notation that word processors fight you over is native here:
\[
  \hat{\theta} \;=\; \arg\max_{\theta}
    \sum_{i=1}^{n} \log p(x_i \mid \theta)
\]

\section{Structure for free}
\begin{itemize}
  \item Sections, figures, and equations number themselves.
  \item Cross-references update when things move.
  \item Bibliographies format themselves from a reference file.
\end{itemize}

\begin{center}
\begin{tabular}{lcc}
  \toprule
  Tool     & Learning curve & Ceiling \\
  \midrule
  Word     & flat           & low     \\
  Markdown & flat           & medium  \\
  \LaTeX{} & steep          & none    \\
  \bottomrule
\end{tabular}
\end{center}

\end{document}

And here is exactly what that produces:

Your browser would rather download PDFs than display them: open the compiled page here.

The compiled result. Also available as the PDF or the raw .tex source.

To try it yourself, no installation required: open this exact file in Overleaf, which will import the source into a free editor in your browser. You need an account, but it makes the whole process of writing LaTeX much easier. Change a word, press Recompile (CTRL-S or CMD-S), and boom, a beautiful PDF.

LaTeX is Used Everywhere

The getting-started page above also explains, in miniature, why LaTeX became the default in three particular worlds:

Research. Mathematics is native, citations manage themselves from a reference file, and nearly every journal and conference publishes a LaTeX template that formats your paper to their rules automatically. Overleaf added the last missing piece: real-time collaboration, so co-authors edit one live document instead of emailing final_v7_REVISED.docx into the email archives.
Tech fields. Plain text means version control, diffs, code review for documents, and automation: results generated by your code can be written into your report by your code. A document pipeline you can script is a document pipeline that scales, which is why it fits engineering culture like a glove.
Professional documents. Anything long-lived and repeatedly revised (a CV, a report series, a book) benefits from separating content from presentation. The typography does the rest of the work for you, every single compile, forever.

The real costs

Love with no complaints is marketing, so: the error messages are archaeological artefacts, a missing brace can produce forty lines of complaint pointing somewhere unrelated, and "Overfull \hbox (badness 10000)" is a sentence I have read several thousand times. And the learning curve is real: the first document takes an evening; the first document you're proud of takes longer.

Every one of those costs is paid in the first ten percent of a document's life, while the benefits (the automation, the consistency, the diffs, the typography) compound. LaTeX front-loads its pain while word amortises its pain across every day you use it.

Who should actually learn it

Not everyone, truthfully. If you write short documents with simple structure, modern tools are fine and Markdown is friendlier still. But if you write anything long, structured, numbered, cited, or mathematical (a thesis, a book, a paper, a technical report, or a CV you'll be revising for the next thirty years) the evening it takes to start is one of the best-paying evenings available to you. Overleaf runs it in a browser now, with none of the old installation hazing.

Mine was simple: the thesis was going to have my name on it forever. It seemed worth typesetting like a real publication should be.

Common questions

Isn't LaTeX overkill for normal documents?

What 50,000 Scrobbles Say About Me

Mon, 06 Jul 2026 00:00:00 +0000

What 50,000 Scrobbles Say About Me

6 July 2026 · personal music

As I write this, my Last.fm profile reads 49,716 scrobbles, which means that sometime in the next couple of weeks, probably while I'm doing the dishes or debugging something, a counter I switched on 15.5 years ago will tick to 50,000.

The counter has been running since the 10th of March, 2011. While I've always failed to keep a diary throughout my life, it turns out I've been keeping one anyway, fifteen years long, written one song at a time, and I only recently thought to read it.

A quick glossary

Scrobble: a record of one song you listened to, sent automatically to a service called Last.fm.
Last.fm: a website that has been logging what its users listen to since 2002.
Loved track: a manual button on Last.fm for marking a song you, well, love.

The top ten makes no sense

Here is what fifteen years of listening looks like when you sort it:

My top ten artists by scrobbles, as of July 2026. Data from Last.fm.

Read as a list of favourite artists, this is gibberish. An Italian pianist, an Australian producer who makes music out of chopped-up Disney films, the man who composed the orchestral soundtrack to Skyrim, a Chicago punk band (who I entirely disavow now), and R.E.M., all sharing a podium. I guess this might be why the algorithms of music apps really struggle to find music I might like. It may as well just hit shuffle and hope for the best.

But that's because the chart isn't a list of favourites. It's strata. Each of those artists is a geological layer from a different era of my life, compressed into one picture. So let me read it the order it was written.

The Scotland years

Alkaline Trio, Yellowcard, Jimmy Eat World, The Smashing Pumpkins, The Strokes, The White Stripes, Manic Street Preachers, R.E.M. This is the music of being young in Scotland: burned CDs that I'd play in my parents car (they were too kind in letting me play Dragonforce and Children of Bodom, ooft), band t-shirts, and lots and sitting in my room with Ultimate Guitar on my computer screen trying to play songs way too complex for my intermediate level guitar skills.

The evidence survives. Here is me playing "Cherub Rock" by The Smashing Pumpkins, 17 years ago:

A few gigs made it into this layer too: Rammstein (my first, at twelve: being surrounded by that much fire leaves an impression), The White Stripes, Smashing Pumpkins. All were great (though I ended up seeing Smashing Pumpkins 3x in total, with each time presenting less excitement from the band, then Billy Corgan spat into the crowd and onto my face, which I was less than pleased about).

What strikes me now is that almost none of this layer is still accumulating. R.E.M. sits at 666 plays, and most of the others have barely moved in a decade. The person who needed them gradually changed, the way a diary's handwriting changes without any single entry looking different from the last. Though I do still listen to a lot of Dragonforce and Symphony X.

The free internet

The next layer is stranger, and I suspect only people of a very specific internet generation and online set of communities will get: Brad Sucks, Josh Woodward, Jonathan Coulton, Lemon Demon, and Pogo, who between them account for thousands of listens. My sixth most-played artist of all time is a one-man band from Canada who called himself Brad Sucks and gave his music away for free on the internet. I actually talked to him once, and he came across as depressed as his songs would indicate he is, which in retrospect I shouldn't have been surprised by. He's a good example of the tortured artist, creating beautiful music out of a tough life.

This was the era I've written about before, when the internet felt like a place made by people rather than companies. Musicians put entire albums online for nothing, because the idea that strangers anywhere in the world could hear your songs was still intoxicating enough to be its own payment. I found Brad Sucks through Last.fm, Jonathan Coulton through WCRadio (warcraft radio) podcasts, and Lemon Demon through a Flash animation about fictional characters having a fight (though this isn't even Neil's best song, FYI, just the most well known).

That Flash animation: "The Ultimate Showdown of Ultimate Destiny."

The loading screen years

Jeremy Soule, who scored Morrowind, Oblivion, and Skyrim, sits at number three with 1,160 plays. Below him: Jason Hayes and Russell Brower, composers for World of Warcraft. Howard Shore's Lord of the Rings scores. Ben Prunty's soundtrack for FTL. C418's Minecraft music. Marcin Przybyłowicz's Witcher 3 score, which I was apparently listening to earlier this week.

My eleventh most-played artist of all time is Paradox Interactive, which is a Swedish video game publisher. I have played their grand strategy games so much that an entire corporation is functionally one of my favourite bands, sitting comfortably above Rammstein, which is an odd thought.

It would be easy to read this layer as "Ken played too many video games" (which, well, yeah) but I think it also is the point where music stopped being something I identified with from a teenage angst perspective, worrying about girls or how I was perceived. Game soundtracks are often orchestral, epic, or relaxing, comforting, nostalgic. They accompany you, painting an aural embellishment on top of the stories, the visuals and the characters you play as and with.

If you want one track that explains this entire layer, it's "Totems of the Grizzlemaw", the Grizzly Hills music from Wrath of the Lich King. I have spent more hours inside this piece of music than many of my favourite albums combined:

"Totems of the Grizzlemaw", from the Wrath of the Lich King soundtrack.

I loved it enough that it crossed the line from accompaniment back into activity: ten years ago, in the middle of the chaos of renovating my home at the time, I sat down amidst the "stuff" room and recorded an acoustic cover of it. I also wrote up the tab, so the Ultimate Guitar habit from the Scotland years clearly never left either:

It wasn't a one-off, either. Here's me playing Dan Romer's Far Cry 5 melody, using a tab by Eddie van der Meer:

Music to think to

My number one artist of all time, the musician the data says I love more than any other, is Ludovico Einaudi, a minimalist Italian pianist, at 1,312 plays.

I do genuinely like Einaudi. But he is not my favourite artist. He is my favourite colleague. Einaudi is what plays when I'm writing a paper, or reading about some technical concept. He is a large part of my "Study" playlist. The same is true of the rest of this layer: Andy McKee and Don Ross's fingerstyle guitar, the Oscar Peterson Trio, Mike Oldfield, an ocean of lofi channels with names like oatmello and Lofi Fruits Music, and Bach, who has been study music for three hundred years and remains undefeated. The study playlist is accessible if you want to hear it.

This is the layer that has been accumulating through my PhD and my postdocs, and it reveals that the artist at the top of my all-time list got there by being background. So, I guess scrobbles don't measure love. Einaudi is number one for the same reason my office chair has more hours with me than my closest friends do.

Music My Partner Showed Me

One of the cool things about being with a musician is they open your mind up to not one but many new worlds of music. Suddenly I'm listening to beautiful folk pieces that speak to my heart:

Bollywood songs that I find myself addicted to (and learned the dance to):

And somehow ended up being featured in a music video by a musical artist dear to us both (00:48 seconds):

What the diary leaves out

Like any diary, the record lies partly by omission. There are no scrobbles for the first chunk of my life, so the bands of my childhood are missing entirely (lots and lots of Mike Oldfield, The White Stripes and Busted), as is anything played on a car radio, at a gig, or in someone else's kitchen, or even on apps where I can't so easily send the listening data to Last.fm.

And the counting itself is skewed in a way any data scientist will recognise: it measures frequency when what we usually want is intensity. Last.fm actually has a correction for this, the "loved track" button, which requires a deliberate click rather than passive playback. I have 559 loved tracks against 49,716 scrobbles. I've written before about how badly rating systems capture what we actually value, and my own listening data turns out to be the same story: the chart knows exactly what I did and almost nothing about what it meant.

If you want to poke through the full dataset yourself, the live charts below are generated straight from my profile:

Live charts from lastfmstats.com. Unlike the rest of this post, these will keep updating after I hit publish.

And for more of all this, the music page is the living version of this post: the album wall, the study playlist, and stats that keep counting while I'm not looking.

Common questions

What counts as a scrobble?

Hosting a Photography Portfolio on GitHub for Free

Mon, 06 Jul 2026 00:00:00 +0000

Hosting a Photography Portfolio on GitHub for Free

6 July 2026 · photography

My photography gallery currently serves 515 full-resolution photos, about 1.5 GB of images, to anyone who clicks on them. My monthly hosting bill for this is zero.

The catch is that GitHub Pages, which hosts this entire site, has a hard rule: a published site may be no larger than 1 GB. My photo collection is bigger than my entire allowance, ignoring all the blog content, and the rest of this website. This post is about GitHub releases.

Quick jargon guide

GitHub Pages: GitHub's free static-site hosting, push HTML to a repository and it becomes a website. It's what serves the page you're reading.
Repository (repo): a project folder tracked by git, holding every file and the full history of every change ever made to it.
Release: a snapshot of a repository at a point in time, to which you can attach downloadable files ("release assets"), normally installers and binaries, in this post something else entirely.
Git LFS: Large File Storage, GitHub's official add-on for versioning big files. Solves repo bloat by charging you money instead.
Egress: cloud-billing jargon for data leaving a provider's servers, i.e. what it costs when someone downloads your photo. The line item that ends up dominating image hosting bills.
Manifest: a plain list of files (here, a JSON file naming every photo) that a page reads so it knows what exists without asking a server to enumerate anything.
WebP: a modern image format that compresses far smaller than JPEG or PNG at similar quality, ideal for thumbnails.
CLIP: an OpenAI model that matches images against text descriptions, which makes it a free, automatic photo-tagger.

Hitting the wall

When I first added my photography to this site, I exported the photos and committed them to the repository like any other asset. It worked, for a good few years. Then at some random point well past the 1GB limit, GitHub let me know, politely but firmly, that this was no longer going to fly.

Worse, the damage outlives the mistake. Git never forgets: even after I removed the full-size photos from the site, every committed version of every photo remains in the repository's history. My .git pack sits at 4.91 GiB to this day, a monument to the months when I treated a version control system like a hard drive. (You can scrub history with tools like git filter-repo, but every clone, fork, and local copy has to be reconciled with the rewrite, and for a personal site it has so far been easier to live with the scar tissue.)

So the problem statement became: keep the site on GitHub Pages, keep the photos at full resolution, pay nothing, and stay under 1 GB. The numbers to beat, for the curious:

1 GB, maximum size of a published GitHub Pages site (and the recommended ceiling for the repository itself).
100 MB, maximum size of any single file pushed to a repository.
100 GB/month, the soft bandwidth guideline for Pages sites.
2 GiB, maximum size of a single release asset, with no documented cap on the total size of a release.

So: releases are a free bucket

GitHub releases exist so projects can attach compiled binaries to a version tag: you tag v1.0, you upload the installers, users download them. But strip away the intent and a release is just an object store bolted onto your repository. Files up to 2 GiB each, no total-size cap in the documentation, served from GitHub's download infrastructure rather than your Pages allowance, and addressable by a clean, predictable URL:

https://github.com///releases/download//

So my full-size photos live in a release called photos-v1, attached to this very website's repository. Uploading is two commands with the gh CLI:

gh release create photos-v1 --title "Full-size photography" --notes "Gallery originals"
gh release upload photos-v1 photos/*.png

That release now holds all 515 originals, 1.55 GB of them, which is to say: the "attachment" stapled to my repository is bigger than the entire website is allowed to be. GitHub is fine with this. Releases are a supported distribution mechanism for a repository's files, and these are, quite literally, the files this repository's site is built around.

The full pipeline

The release solves storage, but you can't lazily browse a 1.5 GB folder from a phone. The other half of the architecture is the oldest idea in web galleries: thumbnails for browsing, originals on demand. Here's the whole pipeline, which is a handful of small Python scripts I run locally when I add photos:

Normalise filenames. Every photo becomes a number: 417.png. No IMG_20250612_final_FINAL(2).jpg. Numeric names make the manifest trivial and the URLs boring, and boring URLs are the ones that keep working.
Generate thumbnails. Each original gets a compressed WebP thumbnail (img/photography/thumb/417.webp). These do live in the repository, 492 of them (a couple dozen originals are heroes and duplicates that never got gallery thumbnails), because collectively they're small enough to sit comfortably inside the Pages budget. This is the only image data the site itself hosts.
Auto-tag with CLIP. A local script runs every photo through OpenAI's CLIP model to classify it (landscape, urban, wildlife, winter…) and writes the results to a JSON file. I wrote about this in its own post; those tags drive the gallery's filter buttons, and I never manually categorise anything.
Write a manifest. A JSON list of every filename in the release. The gallery page fetches this instead of asking the GitHub API, so the site works even when the API is rate-limiting.
Upload the originals to the release with gh release upload, and commit the thumbnails plus the two JSON files.

Since step 2 is where the actual "optimisation" happens, here it is in full, this is the entire thumbnail pipeline, and it hasn't needed to change in years:

from pathlib import Path

from PIL import Image
from tqdm import tqdm

PHOTOS = Path("img/photography")
THUMBS = PHOTOS / "thumb"
WIDTH = 400
QUALITY = 80

THUMBS.mkdir(exist_ok=True)
photos = sorted(PHOTOS.glob("*.png"))

before = after = 0
for photo in tqdm(photos):
    img = Image.open(photo).convert("RGB")
    height = round(img.height * WIDTH / img.width)
    thumb = img.resize((WIDTH, height), Image.LANCZOS)

    out = THUMBS / (photo.stem + ".webp")
    thumb.save(out, "WEBP", quality=QUALITY)

    before += photo.stat().st_size
    after += out.stat().st_size

print(f"{len(photos)} photos, {before/1e6:.0f} MB -> {after/1e6:.1f} MB "
      f"({1 - after/before:.0%} saved)")

Three choices doing the work. LANCZOS is the slowest and best-looking of Pillow's resampling filters, at thumbnail sizes the difference from the default is visible, and the extra seconds are irrelevant for a batch job. convert("RGB") drops the alpha channel the PNGs carry but the thumbnails don't need. And WebP at quality 80 is the sweet spot where artefacts stop being visible at 400px before the bytes stop shrinking. Net effect: a ~3 MB PNG becomes a ~17 KB thumbnail, and the whole gallery's browsing layer, all 492 thumbnails, weighs 8 MB, less than three originals.

At view time, the gallery JavaScript loads the manifest, renders thumbnails in batches of 24 as you scroll, and each lightbox link points at the release URL for the original. The visitor's browser talks to GitHub's release servers directly; my Pages bandwidth only ever carries the thumbnails.

© Ken Reid. All rights reserved. This thumbnail weighs a few dozen kilobytes and comes from my repository. Click-through to the original in the gallery, and GitHub's release servers hand you the full-size file at no cost to my hosting budget.

Why not the "proper" solutions?

Git LFS is GitHub's own answer to large files, and it would have been the natural choice, except the free tier is a single gigabyte of storage and a similarly small monthly bandwidth allowance, after which you're buying data packs. It solves the repository-bloat problem while replacing "free" with "subscription".

Object storage (S3, Cloudflare R2, Backblaze B2) is a good answer and I'd use it for anything commercial. But it means an account, a bucket policy, a billing alarm, and a second system to maintain for a personal site whose entire ethos is "one repository contains everything". My gallery's originals sit next to the code that serves them, in the same repo, behind the same login, and I'm lazy in odd ways.

Photo platforms solve a different problem, community and discovery, and charge accordingly. For raw hosting, here's the rough landscape as of mid-2026 (prices drift; check before you commit):

Option	Rough cost for ~1.5 GB + modest traffic	Notes
Flickr Pro	~$70–90/year	Unlimited storage, but your portfolio lives on their site, in their design.
SmugMug	~$100+/year	Polished portfolio hosting; genuinely good, genuinely not free.
Amazon S3 + egress	Pennies for storage, ~$0.09/GB served	The bandwidth line item is the one that surprises people.
Cloudflare R2	~Free at this scale	Free egress and a generous free tier; the strongest alternative. Still a second account and billing surface.
GitHub release	$0	Everything in one repo. You're reading the case study.

The caveats

No contract. GitHub documents release assets as a distribution feature; they don't promise it as a CDN. If your traffic looked like abuse, they'd be within their rights to object. A personal portfolio's click-through traffic is nowhere near that territory, but a startup's image backend would be.
Public means public. Release assets on a public repo are accessible to anyone with the URL. Fine for a portfolio, since a portfolio's job is to be seen; wrong for anything private.
URLs are tied to the repo. Rename your account or repository and every deep link changes. Choose boring, stable names.
No image processing. S3-with-a-CDN setups can resize on the fly. Here, every size you serve is a size you generated yourself. I need exactly two: thumbnail and original.
Don't commit the originals, ever. The release replaces the repo for full-size files, it doesn't supplement it. Learn from my 4.91 GiB of permanently embarrassing git history.

Constraints make better sites

The 1 GB limit made the gallery better. Being forced to split hot data (thumbnails, manifest, tags) from cold data (originals) is just good architecture, the same shape as any cache-and-archive system. The gallery loads fast on bad connections because it physically cannot ship the heavy files up front.

Total infrastructure: one repository, one release, three JSON files, and a few local Python scripts. Total cost: nothing. The photos are backed up, versioned adjacent to the site that shows them, and served by one of the most reliable download infrastructures on the internet.

Common questions

Is this against GitHub's terms of service?

Why You Aren't a "Visual Learner"

Sat, 04 Jul 2026 00:00:00 +0000

Why You Aren't a "Visual Learner"

4 July 2026 · philosophy science

At some point, probably in a classroom with the lights half-off and an overhead projector humming, somebody handed you a quiz that sorted you into one of four boxes: Visual, Aural, Read/Write, or Kinesthetic. You got your label assigned to you and you may still remember it. People bring up their learning style the way they bring up their Myers-Briggs type or their Hogwarts house (ew, J.K. Rowling reference), except this one comes with institutional backing: schools teach it, teacher training repeats it, and somewhere between sixty and ninety percent of educators across multiple countries believe it, depending on which survey you read. Sigh.

There is no good scientific evidence that teaching to a student's preferred learning style improves their learning. Not weak evidence. Not mixed evidence. When the claim has been tested properly, it fails.^[1]

I believed this one too, when I was younger. It is a collective illusion that nearly all of us bought into, because it flatters us, because it sounds like science, and because it contains just enough truth to be convincing. The myth does real damage and the way it survives tells you something about how bad ideas persist generally in society.

Quick jargon guide

Learning styles: the claim that each person has a fixed way their brain best absorbs information, and that teaching should be matched to it. This is the idea under examination, and it is distinct from merely having preferences.
VARK: the most popular learning styles framework, sorting people into Visual, Aural, Read/Write, and Kinesthetic (movement and touch). Usually administered as a short self-report quiz.
Modality: the channel information arrives through: pictures, speech, text, or physical activity.
Meshing hypothesis: the testable core of the learning styles claim: that instruction works better when its modality is matched ("meshed") with the learner's style. This is the part the evidence fails to support.
Dual coding: combining words with relevant visuals. Helps essentially everyone, not just "visual learners."
Active recall: deliberately retrieving information from memory (self-testing, explaining from a closed book) rather than re-reading or re-watching it. Also called retrieval practice.
Spaced repetition: revisiting material over widening intervals instead of cramming it in one sitting.

The intuitive trap

Humans love a categorisation instrument. We are the species that invented the personality quiz, the enneagram, and the sorting hat: a short test that returns a diagnosis of who you really are. "Kinesthetic learner" feels like a personalised readout of your brain's wiring, delivered with the authority of an acronym, that will unlock your magical learning capabilities like an episode of Limitless. It explains your struggles (no wonder algebra was hard, it was taught wrong for you) and it costs nothing to adopt.

And, crucially, there is a grain of truth there. People genuinely do have preferences about how they take in information. I would rather read documentation than watch a tutorial video at a fixed pace, and if you have opinions about that sentence, you have preferences too. Preferences are real, measurable, and worth accommodating where it is easy to do so, if only because comfortable students are less miserable.^[2]

The myth is what happens when a preference gets promoted into a claim about cognitive architecture: not "I like diagrams" but "my brain learns through diagrams, and text mostly bounces off." Those are radically different claims. The first is a fact about your tastes while the second is a testable hypothesis about memory and cognition, and it is the one that turns out to be false. What learners prefer and what actually helps them learn are, inconveniently, not the same thing, and learners are surprisingly bad judges of the difference.^[3]

The evidence deficit

The learning styles claim has a proper name in the literature: the meshing hypothesis. It says that instruction is more effective when its modality is matched, or meshed, with the learner's style; the visual learner taught with diagrams should outperform the same visual learner taught with text. This is a perfectly testable claim, and to test it you need a specific kind of experiment: classify learners by style, randomly assign them to different instructional formats, then give everyone the same test. If the hypothesis is right, you get a crossover: visual learners do best with visual instruction, verbal learners do best with verbal instruction. Doesn't that sound neat?

In 2008, the Association for Psychological Science commissioned four cognitive psychologists, Pashler, McDaniel, Rohrer, and Bjork, to review the entire literature. Their conclusion is remarkable for how little it hedges: virtually no studies exist with a design capable of testing the claim, and among the ones that do, several found results that flatly contradict it. Their words, not mine:

At present, there is no adequate evidence base to justify incorporating learning-styles assessments into general educational practice.
Pashler, McDaniel, Rohrer, & Bjork, Psychological Science in the Public Interest^[1]

For a multi-billion-pound industry of assessments, workshops, and curriculum redesign, the empirical foundation simply is not there.

Once you see why, it feels obvious. Most of what we learn is not stored as sights or sounds but as meaning. When you learn what the French Revolution was, you are not filing away a picture or a soundtrack; you are building a structure of causes, actors, and consequences that you can express in any modality. The best way to teach something is therefore dictated by the content, not by the student. Geometry wants spatial representation. A new language has to be met in use, whether spoken, written, or signed. You cannot learn to drive from a book, however much you like reading, and no amount of interpretive movement will get a kinesthetic learner through a Shakespearean sonnet. Notice that none of this pins a subject to one sense: blind mathematicians work through geometry by touch and description, and signed languages are acquired as completely as spoken ones. The knowledge is the structure, not the channel it happened to arrive through. Good teachers already know this, which is why they were teaching maps with maps and music with sound long before anyone sold them an inventory for it.^[2]

What the myth costs

If learning styles were merely wrong, they would be a harmless piece of astrology for the staff room. The problem is that the belief has teeth, and I have watched them bite from both sides of the classroom.

The damage on the student side is the self-limiting belief. A student who has decided they are a kinesthetic learner now owns a ready-made explanation for every struggle: the algebra is not hard, the delivery is wrong. The label converts "this is difficult and I need to work at it" into "this was not made for people like me," which is a much more comfortable thought and a much more corrosive one. It writes off entire subjects, and entire formats, as other people's territory. The wider version of this worry shows up in the literature too: when education is guided by what learners prefer rather than what works, the learners lose, because preference and benefit frequently point in different directions.^[3] I wrote about a cousin of this fallacy in my defence of audiobooks: the assumption that the format something arrives in determines how well a brain can absorb it.

During my years teaching at universities, learning styles came up in teaching development the way fire safety comes up in an office induction, as settled procedure rather than open question. Colleagues and I were encouraged to consider the mix of learning styles in the room, and dutiful educators everywhere spend real hours building four versions of material because a framework told them their students needed it. Those hours are far from free, when most teachers were also full time researchers, part time mentors and supervisors, and expected to review theses and papers on top. Every hour spent adapting a lesson into modalities comes out of the budget for things with actual evidence behind them: better examples, better feedback, better practice problems. Multiply that across every school and university that takes the framework seriously and the bill for a false idea becomes enormous.

What actually works

The good news is that dismantling the myth does not leave a crater; cognitive science has replacements with actual evidence behind them, and they share a theme: they are for everyone.

Dual coding. Combining words with relevant visuals improves learning across the board, because verbal and visual channels reinforce each other. Note what this does to the myth: diagrams are not a service for the visual-learner minority, they help essentially everybody, just as clear verbal explanation helps essentially everybody. The lesson is "use both," not "sort your students."

Active recall. Retrieving information from memory, by testing yourself, closing the book and explaining the idea aloud, or doing problems without the worked example in view, strengthens memory far more than re-reading or re-watching ever will.

Spaced repetition. Returning to material over widening intervals beats massed cramming so reliably that it is one of the oldest and most replicated results in the field. Just ask musicians learning a piece.

The techniques that work are effortful, mildly unpleasant, and identical for all four VARK letters. The learning styles myth offered the opposite: a personalised shortcut, a reason why learning felt hard that was not your fault and could be fixed by someone else changing their slides. I understand why that sold. But real learning is often frustrating precisely when it is working, and no personality quiz can negotiate you out of that.

References

Pashler, H., McDaniel, M., Rohrer, D., & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119. https://doi.org/10.1111/j.1539-6053.2009.01038.x
Riener, C., & Willingham, D. T. (2010). The myth of learning styles. Change: The Magazine of Higher Learning, 42(5), 32–35. https://doi.org/10.1080/00091383.2010.503139
Kirschner, P. A., & van Merriënboer, J. J. G. (2013). Do learners really know best? Urban legends in education. Educational Psychologist, 48(3), 169–183. https://doi.org/10.1080/00461520.2013.804395

Frodo, Sam, and love.

Thu, 02 Jul 2026 00:00:00 +0000

A queer-friendly reading of Frodo and Sam: not a hidden romance, but male friendship deep enough for tears, touch, and total trust. Insisting that tenderness must be romance rebuilds the exact wall this story tears down. (Gimli and Legolas, on the other hand...)

Optimizing Your Schedule II (or: When Hours Aren't the Thing You Run Out Of)

Tue, 30 Jun 2026 00:00:00 +0000

A reply to Part I: time is not the only budget. Energy is a second, renewable resource that some activities spend and others restore. A working model of a depleted week, and why rest is production, not indulgence.

What We Leave Behind

Sat, 27 Jun 2026 00:00:00 +0000

From the plutonium in the planet's crust to the Golden Record in interstellar space to the memory of a Tuesday coffee, a look at everything we leave behind, and why none of us ever truly vanishes.

Optimizing Your Schedule (or: Treating Your Week Like a Solvable Problem)

Mon, 15 Jun 2026 00:00:00 +0000

What an operations researcher actually means by optimizing a schedule: objectives, constraints and weightings, a values clarification grid, and a working integer-programming model of your week. Nerd mode optional.

What Was I Made For?

Fri, 12 Jun 2026 00:00:00 +0000

On finite time, micro-ambition, Gandalf, Tim Minchin, and why every 'one day' dream needs a plan or it disappears.

MAS*H in the Modern Era: Comedy, Trauma, and the 4077th

Mon, 01 Jun 2026 00:00:00 +0000

A modern look at M*A*S*H: how a Korean War sitcom from the 1970s remains one of the sharpest critiques of bureaucracy, war, and trauma in television history.

Iain M. Banks and The Culture

Mon, 01 Jun 2026 00:00:00 +0000

Smashing the world requires only a hammer; building a better one requires imagination. A reflection on Iain M. Banks's Culture series and the contrarian act of imagining a future that works.

The Individualization of Responsibility

Wed, 27 May 2026 00:00:00 +0000

Corporations reframe systemic crises as consumer morality plays. From plastic bags to five-minute showers, how the public ends up feeling guilty while industrial actors avoid regulation.

Advice to My Younger Self

Mon, 25 May 2026 00:00:00 +0000

Fifty things I would tell my younger self, from pensions and emergency funds to sunscreen, boundaries, and uncomfortable conversations.

The Books I Recommend to Friends

Fri, 22 May 2026 00:00:00 +0000

The thirteen books I actually push on friends, grouped by the kind of friend I'd press them into the hands of.

Why Nobody Thanks the Person Who Stopped the Disaster

Thu, 21 May 2026 00:00:00 +0000

We celebrate the people who fix disasters and take preventive work for granted. That's a bug in how we think, and it costs us.

When I Ended Up in Italy Because of Mistaken Identity

Tue, 19 May 2026 00:00:00 +0000

How a mix-up over an email address sent me to the northern Italian mountains to photograph a tech event I had no business being at.

How to Write a Blog (or: What I Learned By Doing It)

Wed, 13 May 2026 00:00:00 +0000

What I learned about writing a blog: why the word itself put me off, how I found a process, and what to do when people hate your work.

The Hidden Cost of Cobalt

Wed, 13 May 2026 00:00:00 +0000

The batteries in our phones, laptops, and EVs are largely made of cobalt mined in the Democratic Republic of the Congo, often by hand and often by children. A reflection on a presentation I gave, and what we can actually do about it.

ISO 8601 and the Date Cult I Happily Joined

Tue, 12 May 2026 00:00:00 +0000

An ode to ISO 8601 (YYYY-MM-DD) as the only sortable, unambiguous, international date format worth using.

Accessibility-First Design in Data Science

Mon, 11 May 2026 00:00:00 +0000

Data science outputs are often inaccessible by default. Accessibility-first design in plots, dashboards, and tables improves understanding for everyone.

The Life and Death of the Early Internet

Wed, 06 May 2026 00:00:00 +0000

From CRT monitors and Comet Cursor to MSN Messenger, Kazaa, Warcraft 3 forums, and Bebo: a personal history of how the early internet shaped who we are.

From Scotland to Michigan

Mon, 04 May 2026 00:00:00 +0000

From Scotland to Michigan: a personal account of culture shock, free education, healthcare, xenophobia, weather, and what it means to belong somewhere new.

The Ethics of LLM Use (Not LLMs)

Wed, 29 Apr 2026 00:00:00 +0000

An AI researcher's unvarnished take on LLM ethics: the energy numbers, the cobalt hypocrisy, the Jevons paradox, and why telling people not to use AI on environmental grounds is a privilege position.

Dungeon Crawler Carl and the Strange Dignity of LitRPG

Sat, 25 Apr 2026 00:00:00 +0000

A book snob's confession: how Dungeon Crawler Carl used absurd LitRPG mechanics to deliver one of the most devastating emotional payloads in modern fantasy.

Fifteen Years of Silence: Patrick Rothfuss, The Doors of Stone, and the Architecture of Creative Friction

Wed, 22 Apr 2026 00:00:00 +0000

Fifteen years since The Wise Man's Fear and still no Doors of Stone. A personal reflection on the Kingkiller Chronicle, the long wait, the charity debacle, and why writing at this level is genuinely hard.

No Idea? No Problem: A Beginner's Guide to Building Your Data Science Portfolio

Mon, 20 Apr 2026 00:00:00 +0000

A practical guide to building a data science portfolio when you have no idea where to start. Covers project types, documentation, novelty, collaboration, and working in the GenAI era.

Evolutionary Computation's Identity Crisis in the Age of GenAI

Sat, 18 Apr 2026 00:00:00 +0000

Evolutionary Computation is having an identity crisis in the GenAI era, despite delivering major real-world wins in engineering, logistics, and constrained optimization.

In Defense of Audiobooks

Fri, 17 Apr 2026 00:00:00 +0000

The stigma against audiobooks is rooted in bad assumptions about learning, intelligence, and what counts as reading. The research doesn't support any of it.

Why It Still Matters to Learn to Code in the Age of AI

Thu, 16 Apr 2026 00:00:00 +0000

Coding still matters because the value is not syntax memorization. It is learning to break down problems, reason across systems, and build reliable solutions.

Your Professional Second Brain for Local LLM Work

Wed, 15 Apr 2026 00:00:00 +0000

A practical system for project documentation, knowledge management, and local LLM workflows that improves delivery, reviews, and project re-entry.

Snowball vs Avalanche: The Science of Paying Off Debt

Tue, 14 Apr 2026 00:00:00 +0000

Snowball and avalanche both work, but for different reasons. A practical guide with simulator-backed charts on interest, timing, and motivation.

Book Ratings Are Broken

Sun, 12 Apr 2026 00:00:00 +0000

569 books rated, and the star system fails at all of them. A data-driven look at my own Goodreads ratings, review lengths, genre bias, and the silence problem.

Why Most Self-Help Books Are Trash (And Which Ones Aren't)

Sat, 11 Apr 2026 00:00:00 +0000

The bad ones are blog posts stretched to 300 pages. The good ones cite their sources. Here's how to tell the difference.

Building a Photo Tagging System with CLIP

Fri, 10 Apr 2026 00:00:00 +0000

How I used OpenCLIP to automatically classify 500+ photos into 10 categories in under a minute, zero API cost, running entirely on CPU.

[ref-1] Pashler, H., McDaniel, M., Rohrer, D., & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119. https://doi.org/10.1111/j.1539-6053.2009.01038.x

[ref-2] Riener, C., & Willingham, D. T. (2010). The myth of learning styles. Change: The Magazine of Higher Learning, 42(5), 32–35. https://doi.org/10.1080/00091383.2010.503139

[ref-3] Kirschner, P. A., & van Merriënboer, J. J. G. (2013). Do learners really know best? Urban legends in education. Educational Psychologist, 48(3), 169–183. https://doi.org/10.1080/00461520.2013.804395