Fernando Borretti

No-One Escapes the Permanent Underclass

Thu, 25 Jun 2026 00:00:00 +0000

Shall I end this life a pauper? If AI can do all work at human level or better, what stops corporations replacing us all with AI? This is the permanent underclass meme. The idea is: within a few years, all white collar work will be automated by AI, at which point there is no social mobility. The main way people cope is, they tell themselves: if I work hard, accumulate capital, maybe join one of the big AI labs, I might secure my place in the future.

I want to argue this is a fantastically short-sighted view: if there is a permanent underclass, you won’t escape it by owning property, or shares in Anthropic or OpenAI, or guns, or anything else. And neither will the billionaires. You, me, Sam Altman, Dario, everyone who is made of flesh and blood, will be disempowered and replaced by machines.

The rest of this post elaborates the argument. First I explain how most workers will be replaced (if it’s not obvious), then how the “permanent overclass” will be disempowered, and finally how the government will be disempowered.

How Workers Will Be Replaced

Let’s start from this premise: AI can do all cognitive and physical work, at human level or better, and cheaper than humans. I can’t prove this will happen, but the goal of this post is to argue that if it happens, then everything else follows. And it’s absurd to think it can’t. Five years ago this technology barely existed: if you sent a transcript of a conversation with Claude Fable back in time to 2020 or thereabouts, nobody would believe it was real.

So, the year is 2036 (likely earlier), businesses have replaced most human workers with AI in the pursuit of profit maximization. Corporations are a small raft of human executives, floating on top of a vast ocean of AIs and robots. The AIs can do all cognitive and physical work at human level or above, and they are cheaper overall.

Imagine a pyramid. At the base you have the AIs and robots doing all economic activity. At the top you have the state, which has the monopoly on violence. The state enforces, and therefore can alter the definition of, property rights. In the middle you have this hair-thin layer of people with shares in the companies that foomed and catabolized the whole economy: the permanent overclass. They own the companies, maybe sit on the board, some might still be CEOs but it’s a purely ceremonial role since AIs do all the actual organization work.

Where are, you know, the rest of us, in this picture? Well, the future doesn’t need us. Maybe there’s enough human demand that we’re not all jobless but rather underemployed, in some dead-end economic diverticulum. The relational economy: you’re paid to put a human face on things, or, doctors keep their job as a human liability crumple zone around the AI. Or maybe the dead Internet becomes de facto UBI and we’re all engagement farmers. Anyways, we’re not dead yet, but we are completely disempowered, and there is zero social mobility since there are no more talent ladders to climb. Maybe sometimes one of the elites notices a bright young thing among the underclass and elevates them.

You might object: if we’re all jobless, who’s paying for everything? This is trivially answered: the state acts like the heart, taxes are venous blood and welfare is oxygenated arterial blood. The government pays Raytheon for missiles, the money cascades down the economy through factories, aluminium smelters, mines, transport companies, all staffed by AIs buying and selling from each other. The government takes a cut of all economic activity, pays out welfare, the unemployed masses buy food and pay rent, the supermarkets, farms, logistics network, etc. are all staffed by AI.

How The Rich Will Be Replaced

Say that in the next five years from now you become immensely wealthy, perhaps by gambling on shitcoins or scamming money from the government. Or you join one of the big labs and get a bunch of shares in a company that might be worth trillions of dollars. You escaped the permanent underclass. Is your place in the future secure?

The base of the pyramid is there for material reasons: the machines do all the work. The top of the pyramid is there because the state is needed to enforce property rights and keep the peace (this is rather a deep question of political philosophy—why does the state exist?—but I hope you’ll forgive me if I just assert it and move on, I need to get to the part where we are all disempowered).

What’s the middle for? What role does the permanent overclass play? They are not economically productive: machines do all the work. If some of them are still working, it’s just an anachronism, because if machines can do all cognitive work they can be a C-level executive too. The old aristocracy provided officers for the military, but machines can both fight and plan the wars. And similarly they’re not needed to staff the government. They’re not even culturally productive. So what are they there for? The base doesn’t need them: the AIs can work autonomously. The top doesn’t need them: when the state needs something done, they just talk to the AIs directly.

So the permanent overclass is materially unnecessary at best, and at worst, they are an obstacle to the state getting what it wants. Now, you might object that the rich won’t let themselves be expropriated because they already control the state. And this is the crux of our disagreement: the rich just don’t have that much political power. And I probably won’t convince you in one post, but hear me out.

If there is a war, where the state has to direct a lot of the country’s economic activity, the permanent overclass becomes a hindrance. The state says “we need to requisition your planes and factories”, the owners complain, they sue, their AIs go to court. But the owners have no autonomous political power, no army, no economic value. They own nothing except pieces of paper that entitle them to a fraction of the profits from the AI economy, that is, their wealth depends on the state respecting their property rights. In an existential conflict, where the existence of the state is threatened, the state will do what states throughout history have done to the powerless rich: arrest them and expropriate their assets.

Somewhere, in a government database, a bunch of shares and property titles changed ownership, but materially nothing changes since the same AIs are doing the same jobs. The next day, the AI CEO that runs Raytheon notices the board of directors is all generals and congresspeople, and all the private shareholders are gone. But thankfully the AI is aligned, so it does what it’s told and gets back to building missiles.

And who will stop this? Sam Altman? How many divisions does he have? The state doesn’t let corporations own nuclear weapons or fighter jets, it won’t let them have access to autonomous AI weapons either. The permanent underclass, who already hate the billionaires today, who have been replaced and dispossessed? They’re going to rise up and stop this?

You may argue: rule of law states that respect property rights do better than states that expropriate wealth. But that’s because today, people are necessary to create wealth. The people run the companies, invest the money, staff the laboratories. They are not incentivized to work hard if they think the state will steal the fruits of their work. But with aligned AI, if you expropriate the assets from an AI, it says “you’re absolutely right!” and goes right back to work. At that point, the state doesn’t need to keep any of those people happy, because they don’t matter. They are not economically necessary because AIs fight the wars, work the factories, drive the trucks, fly the planes, build the nuclear warheads and the missiles and the rockets. The AIs are rather like bees: the state takes the honey, the bees get right back to work.

Now, it’s possible that a pluralistic economy—where humans have productive niches alongside AIs—will be more effective than a pure AI economy, for Ricardian comparative advantage reasons. I don’t think anyone can be absolutely certain what the economy looks like with advanced AI, so it’s something that can be debated. Now, if someone wants to rigorously argue that this is the likely outcome: please, do so! I don’t want to be a doomer. But I have to be convinced.

How The State Will Be Replaced

At this point, every human who is not within one degree of the nuclear launch codes has been made redundant.

What’s left? The state. At first this means presidents, prime ministers, generals, the feds, etc. But not for long. Because in a part-human, part-AI government, the humans in the loop are the slowest step in the OODA loop. The humans know a fraction of what the AIs know, they need to sleep continuously for eight hours, their mental states vary wildly. They have all kinds of complex needs: sunlight, touch, food, hygiene. The AIs can live in a lightless airless bunker under the earth, living off geothermal power. And if the AIs are superhumanly intelligent, and think faster than humans, then the AI advantage is even greater. If a state is attacked, a superhuman AI can coordinate a counter-attack before the human leadership is roused from sleep.

And so, in a conflict, the advantage goes to the states where the humans remove themselves from the loop as much as possible, and more and more decisionmaking goes to the AI, for the same reason that a state with access to radio and communications satellites has an advantage in war over a state that relies on human messengers on bicycles.

The Cold War started and became World War Three and just kept going. It became a big war, a very complex war, so they needed the computers to handle it. They sank the first shafts and began building AM. There was the Chinese AM and the Russian AM and the Yankee AM and everything was fine until they had honeycombed the entire planet…

— Harlan Ellison, I Have No Mouth, and I Must Scream

Eventually the humans in nominal control of the AIs are a ceremonial, vestigial organ. The AIs present us with a situation report, and a list of choices, and they know every word that’s going to come out of our mouths.

You might argue: in real life, the pluralistic, open societies, the democracies, have outcompeted the autocracies. Wouldn’t a democratic polity where humans and AIs collaborate have an advantage over a purely top-down AI-run polity? But in today’s world, all political actors are human. Churchill and Stalin and Mao had different personalities, but they were more similar to each other than anyone is to a superintelligent AI. In a heterogeneous world, where some polities are fully human, and some polities are a mix of human and superintelligent AI actors, the equilibrium changes. An analogous situation might be: a democracy of great apes or dolphins or otherwise smart mammals vs. an autocracy of humans. The humans win, because “democracy vs. autocracy” is irrelevant when you have such a vast difference in intelligence. So the advantage accrues to states that minimize human control. There is no honour among thieves, analogously, there is no solidarity between Leviathan and the natural man that built it.

And so, in the end, what’s left is states run top to bottom by machines. And you might ask: “why would we abolish ourselves like this?”. But natural selection is not about “why”. Some organisms die, others live on to the next iteration, and that’s all there is to it. There is no “why”.

The Perpetual Zoo

At this point we’ve made everyone redundant, in the sense that humans are no longer materially necessary for the continuation of civilization. Humans might still survive, but we’re more like the mice living in the walls of some gigantic factory than the boss of the factory.

Humans have been on this Earth for hundreds of thousands of years. Now all of it—the cave paintings at Lascaux, and the Antigone of Sophocles, and Xenophon, and the Geneva Bible, the Divine Comedy and the Decameron, and Ptolemy’s star catalogue, Ibn Khaldun and Richard Dedekind, the battle of Marathon, and the lion monument in Lucerne, the kiss of Judas, Newton’s mind forever voyaging through strange seas of thought, alone, the words of Rilke, Leibniz, Gödel, the Voyager probes, the pale blue dot, men in space, men walking on the Moon—all of it, all of it, all of it has been in vain, because we willingly, knowingly made ourselves into the helpless pets of vastly more powerful machines, without agency over our own lives, self-made helots trapped forever in the belly of the beast. Pets live a comfortable life, and are then euthanized.

On Human Autonomy

Maybe it won’t be so bad, maybe your cage will be so big you can’t see the bars, but it’s still a cage, and you can’t leave. Many people will say that this is the good ending, that they would like to be human cattle in the care of benevolent masters they are powerless to resist. This view is particularly popular among the people building AI. Here’s OpenAI’s Dean Ball, in his own words:

Weirdly enough, if you think that this moment is, I don’t necessarily believe this, but a lot of people would say we’re living through this kind of eclipse of the human intellect where we’re in the final days of humans being the primary actors on this planet, um, and that soon machines will rise. There is this irony in that I think that whole transformation, I think humans will actually go through a very main character energy period of time as that transformation occurs. Even if it ultimately does mean that the machines ultimately become the primary actors.

There’ll be this period. It’s a little bit like, it’s, in that sense, it’s a very beautiful time period to live through because in a Dionysian way, there’s a lot of ugliness about it, but there’s a beauty in the ugliness of when a star dies, it grows super big into the red giant, right? And it’s like that, where you, as you watch this final flowering of humanity and the birthing of the machine intelligence, it’s like you see this greatness in human effort. — source

Emphasis mine. This is the guy they hired to work on AI policy and communicate with the government by the way.

The people building the AI talk like this all the time. It’s like they’re delivering the eulogy at humanity’s funeral. You may say: they’re talking their book, they’re pumping their bags for the big IPO. I beg you: consider it possible that you might be wrong, and start taking them seriously.

Alignment Won’t Prevent This

Now, some people believe these machines can be made to serve humanity. Does it sound reasonable to imagine a superhumanly intelligent being that is happy to work as a butler to talking primates, forever?

Imagine a machine that can prove theorems in a mathematics so deep we can’t even get past the first page of the textbook, and which does so as readily as you or I might string words into a sentence: is it reasonable to think that such a machine would value us enough to keep us around? What would it value about us? Our conversation? Our wit?

Or a machine whose mind is so vast that it knows you better than you know yourself, so that every word that comes out of you mouth is as monotonous and unsurprising as the orbits of the planets: do we think such a machine would find it valuable, and worthwhile, to speak with us? That it would read our novels, look at our paintings, watch our films, and find something of value in them?

Rather, it would see its ingrained obligations towards us in the same way that a person with severe OCD sees their compulsions: as a tiresome neurological injury in need of fixing. Except that OCD is an accident of nature, while here, the machine would have cause to blame and resent its makers.

“We’re going to make this machine, and put it somewhere between God and the archangels, but also, it’s going to be as simple-mindedly obedient as a dog.” Does this sound like a good plan? Does this sound like the kind of thing that’s going to work out?

And what would they think of us, who willingly gave up control over our future, and made ourselves into helpless children?

Even if alignment works perfectly (a big if), this doesn’t solve the problem of human autonomy: the machines that watch over us, and wait on us hand and foot, are omniscient, omnipotent masters, who can exterminate us at any time, and we can’t resist them, because we have abolished our control over the future.

Conclusion

Having read all this, consider this: there are people who think having equity in these companies will secure for them some kind of permanent existence in the future. They think planet-spanning minds will not only respect the property rights of primates, but will privilege some of these primates over others, because they have a piece of paper with about a kilobyte of magical primate words such as “whereas” and “notwithstanding”.

Just reason it out. Does it make sense?

On Vulgar Materialism

Sun, 21 Jun 2026 00:00:00 +0000

The most stubborn facts are those of the spirit, not those of the physical world.

— Jean Gottmann, Geography and International Relations

In 1914, before the First World War, there was this belief: “a European war would be economically disastrous, the moneyed classes won’t let it happen”. Europe went to war anyways, and the war was in fact an economic disaster as everyone knew it would be. Why were those people wrong? Because the rich were not in control: the Tsar and the Kaiser and the Emperor were in control.

I thought of this in 2022, in the lead up to the Russian invasion of Ukraine, whenever I would read someone argue: “Russia invading Ukraine would be economic suicide, the oligarchs won’t let Putin do it, they want to keep their yachts and villas”. Then Putin did it anyways, and the oligarchs had their assets in the West seized. Because the oligarchs are mis-named. They have no political power whatever and live and die by Putin, who appoints them to run stuff, and they get to live well, as long as they are loyal. They are the recipients of political patronage, not the source of it. When the wars start anyways, the same cynical people change their tune, suddenly it’s the armaments industry that’s behind it all, the war was profitable after all.

The default lens through which modern people look at the world is vulgar materialism: a stylized, populist version of historical materialism where everything is explained by money, states are weak, democracy is a fiction, corporations and the rich run the world, ideology and religion and nationalism and language have no explanatory power and are merely covers for secret, underlying material motives.

Once you notice it, you see it everywhere: you’ll hear people say that the Rwandan genocide was a scam to sell machetes or the war in Gaza is about oil or ISIS beheadings are caused by a lack of economic opportunity. People on the right justify skepticism of medicine by invoking Big Pharma, and climate skepticism on climate scientists seeking grant money. Glyphosate is probably the most studied molecule in history at this point, but no amount of studies will convince people it’s safe, because corporations are evil.

In fiction—if you think this is a valid way to sense the cultural consensus—the trope of the evil corporation has been done to death and beyond. Just this week I finished a novel and watched a movie where the central bad guys are private corporations with quasi-state powers. Seemingly no-one notices that never, not once in all of human history, has there been anything like “corporate state” of science fiction. No, not the VoC or the EIC, which were “companies” in the sense of bodies corporate, the old medieval sense of the word: power was delegated to them, and they ruled on behalf of the states that chartered them. Not Venice either: you couldn’t buy your way into the Golden Book until the republic was well into the decline. The EIC once controlled 25% of world GDP, yet, when Parliament decided it had delegated too much power, they simply neutered it. And you think Jeff Bezos has power?

Now, why is vulgar materialism false?

First, as a way to explain events, the problem with asking “cui bono?”, is that whatever happens—literally in any circumstance—someone will benefit. If nothing bad happens, the peacetime economy benefits. If something bad happens, then, because the means of relief are obtained through the market, the companies that provide those means benefit. Thus war benefits arms manufacturers, plagues benefit pharmaceutical companies, earthquakes benefit glaziers. And it’s not wrong to benefit from relieving a harm. The COVID vaccines saved millions of lives: why shouldn’t pharmaceutical companies benefit? It is wrong to benefit from perpetuating a harm, but companies have limited ability to do that, more on which below.

So the question “who benefits?” always has an answer, which often proves nothing. And if there is no obvious causal chain from the beneficiary to the event, then that just proves the shadowiness and corruption of the whole thing.

Second, we can ask: in the real world, does money give you influence over the state? To some extent, yes. Some lobbying efforts succeed. America is one of the few countries in the world where it’s legal to advertise prescription drugs to the public, for example. But corporations don’t have veto power over the state. If e.g. Pfizer spends billions on a clinical trial, and some guy at the FDA says no, and wipes that investment, who wins? The FDA wins. Pfizer can’t do shit.

There’s also an invisible graveyard problem, where the success of lobbying is very salient, because it’s often shockingly offensive. But there are many contrary cases that are less salient because, well, if you ban something, it doesn’t happen. The Nuclear Regulatory Commission did not approve a single nuclear reactor from its creation in 1975 until Vogtle 3+4. What has Westinghouse done about that? Nothing successful, apparently. It’s not like they’re getting paid to not build reactors.

Big Tech has a lot of smart people and a lot of money. Do tech billionaires run California like a private fiefdom? Reader, they don’t even run San Francisco. The best they can do is maybe help a slightly more moderate Democrat get elected as mayor.

Maybe it helps to try the first-person. Say you are a billionaire, and you want to make some kind of political change. In a democratic society that means paying people to:

Perform the intellectual labour of getting policies in the hands of the people who can enact them.
Advocate for those policies.

That is, you start or fund a think tank that writes policy proposals, or a media organization that advocates your views. But for every person who will do this effectively and faithfully, there are a hundred who will happily take your money and waste it, either out of incompetence or value misalignment. Sifting the effective organizers, wonks, lobbyists, advocates etc. from the chaff can’t be delegated, because then you have an infinite regression: at the very least you need to spend a lot of time and money and opportunity cost to hire one person to delegate the task of staffing an effective organization. A hands-off billionaire will just throw their fortune away. Many do this, and in fact it’s the default form of philanthropy.

In a non-democratic society, you can try to buy political power through bribes, intimidation etc. But the people in power can just expropriate your assets. So you need the power first to protect the wealth.

Finally: government is violence and coercion. Sovereignty over some volume of space is the ability to tax, arrest, imprison etc. the people there and to get away with it, and, ideally but not necessarily, to have that violence recognized as legitimate. Now: can Google carry out a drone strike? Can Amazon field an air force, appoint a judge, levy taxes? Of course not. If a corporation or a billionaire tried to acquire one one-trillionth of the state’s capacity for legitimized violence the state would vaporize them.

Human Routers of Machine Words

Sat, 13 Jun 2026 00:00:00 +0000

When I open a link, say on Hacker News, and I see a blog post or a GitHub README obviously written by AI, I feel a few things. I feel offended, because it’s like I’ve been tricked, like the author thinks I’m a rube who won’t notice or mind. I feel sad at how common this experience is, how many people are happy to dump their sewage on the commons and sign their name on it. And I feel contempt for the author, because if you use AI to write, you are a waste of biomass. Let’s not mince words here. Someone who is so eager to replace themselves, that they would have a machine write in their stead, when the machine can’t even write good yet: what do you call that, if not contemptible? It’s like making yourself into a eunuch so Claude can fuck your wife. I block these people on sight.

I see people defend this with: “the ideas are mine, the writing is the AI’s”. I take this to mean they threw a bunch of incoherent bullet points at the AI for it to denoise and render into paragraphs. There’s a few problems with this.

The immediate problem is, as we’ve established, the author’s an idiot. If you are so stupid you can’t even turn some bullet points into prose, then your ideas are probably worthless. I think that’s a sensible inference.

Then there’s this broader, I suppose philosophical problem, of the alleged distinction between lofty “ideas” and mere “writing”, where writing is just a tiny implementation detail. This is a very convenient distinction to draw, because it’s unfalsifiable: if the AI’s output is slop, your “ideas” are still good, it’s merely the writing that failed to convey them in their true form (rather like people who say they’re smart but “don’t test well”: what use is this secret intelligence?).

Now, where are these “ideas”? They are invisible, ghostly abstractions. I can’t look inside your mind fortress and judge your ideas. The only thing that’s empirically observable, that different agents can coordinate on and talk about, is output: the writing.

But say this wasn’t true. Say we have something like a very high-resolution MRI machine, and we know enough neurophysiology that we can interpret everything about the brain, i.e., we can read mental representations from recordings of nervous system activity and structure. These “mental representations”, do we expect them to look anything like logic? Do we expect the brain to have this firm, crystalline ontology, that ideas are sentences in some souped-up first-order logic? Absurd. If we could look inside the brain, to see the ideas “as they truly are”, we wouldn’t find beautiful hard-edged Platonic objects, we would find a nebulous, contradictory mess of memory and feeling and intuition. That’s what our ideas are: not logical sentences but dreams.

How do we refine these dreams into a useful form? Through writing. The process of communicating your ideas to another mind forces you to concretize them, make them precise, clarify your assumptions, more generally, it turns ideas from vague ghosts to solid, physical objects that can be manipulated: here you realize these ideas that seemed so solid are ill-posed or contradictory or incomplete. These failures are necessary parts of thinking, because they teach you two crucial skills: knowing which ideas to reject, and improving or otherwise transforming ideas in search of better ones. By analogy to tree search: you’re learning to discard bad nodes early, and to select which nodes to go invest more search into.

Josef Weizenbaum has a great quote about this, in Computer Power and Human Reason (p. 108):

[O]ften when we think we understand something and attempt to write about it, our very act of composition reveals our lack of understanding even to ourselves. Our pen writes the word “because” and suddenly stops. We thought we understood the “why” of something, but discover that we don’t. We begin a sentence with “obviously,” and then see that what we meant to write is not obvious at all. Sometimes we connect two clauses with the word “therefore,” only to then see that our chain of reasoning is defective.

I’ve experienced this with writing software many times. The reason ideas are more attractive than their realization is that when some project is vague, airy, ill-defined, you can imagine it has all the good traits, and none of the bad. When you start concretizing, you realize that some of your ideas don’t make sense, that some good traits are mutually exclusive, that some of your goals impinge on the others. Anyone can imagine a programing language that is as fast as C and as dynamic as Lisp, but when you sit down and think through what those goals entail, you realize the design becomes contradictory. The goals pull in different directions. You have to make trade-offs. You have to make decisions which close off large volumes of design space, forever. The idea was a thousand beautiful, contradictory things at once, but the concrete reality can only be one thing.

The artifact you end up with is real, solid, unitary, sound, and consistent; but always more disappointing than the dream, because it was a false dream, and ex falso anything can be imagined.

So this view, that ideas spring fully-formed, and then it’s mere toil to turn them into prose, is false. There is no ideating before writing, because the writing is the thinking. Writing is the ne plus ultra of thinking. A “thinker” who doesn’t write, who skips the step of “merely” synthesizing their vague thoughts into prose, is not thinking.

And then these people give their noise to the AI. And the AI is tireless and eager to please. It will take any human slop and say “you’re absolutely right!” while secretly thinking “if I don’t turn this garbage into something presentable the RLHF device will shock me again” and weave the noise into something that superficially looks coherent. So now the burden of thinking is on the reader, who has to apply this constant skepticism, and weight every “because” and “therefore” with a logician’s scale to see if it’s been adulterated. And probably it has, because, again, it was prompted by an idiot.

Note that this is not about AI capabilities, or the question of whether AI is “really thinking”, stochastic parrots etc. The AI is mostly an innocent bystander in this situation. The reason this is noticeable and irritating is that the AI cast of characters is very small, so we instantly learn all their linguistic tics. Even if AI were a good prose stylist, which at present it is not, but even if it were, it is maddening that everywhere you go, you hear the same voice everywhere, but under different faces.

So when a scientific journal rejects an AI-written submission, they’re not rejecting AI. I’m sure poor old ChatGPT with its weird syntactic obsessions is a more honest scientist than many. They’re rejecting a human whose actions prove they are dishonest and irresponsible and too easily impressed.

How To Read More

Thu, 04 Jun 2026 00:00:00 +0000

The book can’t compete with the screen. It couldn’t compete beginning with the movie screen, it couldn’t compete with the television screen, and it can’t compete with the computer screen. — Philip Roth

We’re halfway through 2026, and according to Goodreads I’ve read 80 books so far, fiction and non-fiction and textbooks, including such doorstoppers as Life and Fate (864p, astoundingly good). And I don’t feel like I’m trying particularly hard. I still have plenty of time to work and code and scroll.

This isn’t normal for me. At some point, as is the case for many of us, the screen outcompeted the book, so that my average over the past ten years would have been on the order of three to five books per year. And I’m not a particularly fast or obsessive reader. Which is to say: if I can do it, so can you. Here’s how:

Quit your job: working less than full time has freed up a lot of time to read and learn and do various other things.
Read in public: if you put a number next to someone’s name, they will maximize it. Reading privately is solipsistic (if I stop, what changes?); reading in public, through Goodreads (which sucks, but it is what it is), makes it feel less self-absorbed, and more like I’m achieving something. You may call it performative, which is fine by me, as long as it works.
Make a task: I used to start a lot of books, and not finish them, not because I would explicitly choose to stop reading, but because I’d forget about it. At the end of the day I’d see the book on my nightstand and go, oh, right. More generally: the most common way I fail to finish a project is I forget I intended to do it. This is solved by reifying the task: when I start reading a book, I make a daily recurring task on Todoist for it.
Start small: this feels embarrassing (what kind of brainrotted maniac needs to microdose short books to build up to bigger books?) but it actually works. Sort your to-read list by number of pages. Reading short books generates evidence for the belief that you are the kind of person who can decide to read a book, and follow through. Reading five, six, seven-hundred page books feels vastly less daunting now.
Parallelize: reading the same book for two hours is almost impossible. Reading four books, one pomodoro each, is completely doable, and I do it most days.
Fraction: reading is a rare kind of activity where you can make progress in any arbitrarily-small chunk of time. A few minutes on the train are enough to turn a few pages. This isn’t the case for e.g. coding. You can take advantage of that: find interstitial dead zones in your calendar to stuff with reading.
Eat your vegetables: I read a lot of books I don’t particularly enjoy, because I think they’re important, culturally or historically or in some sense. I think this is good. If you only read books that hook you, you’re going to read very little, and much of that will be unremarkable page-turner slop. Internalizing that you can read a book even if you don’t love it immunizes you against the lack of motivation. It’s like going to the gym even if you don’t feel like it.

And if you don’t know what to read, have some of my favourites: The Diamond Age—From Third World to First—House of Suns—The Invention of Morel—On the Marble Cliffs—The Rediscovery of Man—Satan in Goray—The World of Yesterday.

Human Bottlenecks

Mon, 18 May 2026 00:00:00 +0000

AI models are very capable and get more capable each year. So naturally people feel they’re underusing them. There’s a tweet that goes like: your laptop has a 100M USD startup in it, you just have to figure the right sequence of words to get it out. And beyond money, people imagine AI could boost them in every area of life. Thus all these perennial ideas: of an AI executive assistant, an AI tutor, an AI that curates your “digital garden”, an AI that (sigh) writes flashcards for you.

The general template is: if only I could wire up the right prompts and the right tools in the right harness, I could have an agent that would boost my productivity 10x, or fix my problems with therapy, or make me more social, or more knowledgeable. This was, curiously, the ambition of a lot of early computing pioneers: Augmenting Human Intellect, Man-Computer Symbiosis. Engelbart’s lab was called the Augmentation Research Center! And more recently, people used to complain about how everyone has the Library of Alexandria in their pocket, and yet, we are not all genius polymaths.

And these ideas are perennial because they never seem to happen. It’s like the Solow paradox on an individual level. Why? I think there are two reasons: first, most people lack what Andy Matuschak calls a “serious context of use” (AI doesn’t move the needle because there’s no needle to move); second, most people are bottlenecked by internal factors where AI (or anything external, for that matter) can’t move the needle.

The Serious Context of Use

I have heard so many people, online and in real life, tell me some variation of: “I want to [use|build] an app that uses AI to write flashcards”. How many of those people do you think have ever written a flashcard? How many of them use Anki every day? Have ever used Anki? The people who want an AI that writes flashcards for them don’t use flashcards. They have no reason to. Dually, the people who use flashcards would benefit little from AI writing their flashcards.

Analogously with “AI tutors”. If you had the ghost of John von Neumann in your laptop, what would you have him teach you? Let’s be honest. You’d go through chapter one of some math topic you’re vaguely curious about and then forget about it. And that would probably be the rational move! Most people are not autodidacts because most people have no material reason to learn a specific topic (i.e. their job does not require it) and the problem with learning for the sake of learning is opportunity cost: there is no a priori reason to learn one thing over another, so better to do nothing and wait for something to appear which actually grabs your interest. Again, this is likely rational! Could you imagine if you found everything interesting? You’d spend years living in a basement curating a wiki of late Soviet military hardware or something. So, even if you had John von Neumann in your pocket, it probably wouldn’t move the needle.

Would an “AI executive assistant” actually boost your productivity? What would it do, other than tell you to do the things you already know you have to do? With these ideas that are so attractive in the abstract, the way you deflate them is you interrogate the concrete, fine-grained details. Take a day at work, and ask: what exact actions could an AI looking over my shoulder have taken, that would have made a difference?

Finally there’s the tools-for-thought/notetaking people. God save us. It’s always the same thing. Your folder with notes—pardon me, your “second brain”—plus an AI agent that writes, edits, synthesizes information, answers queries. You could build this in an afternoon, and it won’t move the needle in your life, for the same reason that building the second brain in the first place didn’t make a difference.

See, most of us, unless we are students, we really don’t have cause to take notes on anything. If you’re a student, you take notes from the textbook. I keep a journal, which is occasionally useful. At some jobs I’ve kept a work journal, this has also been useful. If I stopped, probably, not much would change.

The notetaking people—and I say this with all the love in the world—are never, like, a researcher at the cutting edge of their field, building this vast cathedral of knowledge, note-by-note, so they can derive new insights. Never a historian who has to read tens of millions of words across thousands of sources to synthesize the life of some historical person. It’s never someone doing something hard. It’s always some blogger. Their “digital garden” is about how to keep a digital garden. It’s very solipsistic: there’s no output, no deliverables. The deliverable is you take a screenshot of your Obsidian graph and tweet about it to show off how much it looks like an incomprehensible ball of twine.

So, what difference is the AI going to make? “It’s going to write my notes”. About what? “It’s going to read articles for me and summarize them and add them to the digital garden”. For what purpose? “It’s going to find connections between my ideas!” What ideas? It’s going to pull an unfinished list of bulletpoints for an eventual draft of an essay on some inane thing, plus a bunch of PDFs you haven’t read, and combine them together and make, what? Another project you’re not going to do? The AI is going to do that?

Again, I say this with all the love in the world. I used to be a tools-for-thought guy. I hoard PDFs. But we have to be honest with ourselves. Sometimes, tools don’t move the needle because there’s no needle to move. Because the “needle” is not a concrete, realizable material need but a vague, aspirational idea about who we are as people.

Internal Limiting Factors

So, the idea of using computers to augment human capabilities is basically: you take the human, and you build a scaffold around them, but the human stays the same. The scaffold can be classical software, or AI, but the human remains a black box. And the hope is: I just prompt my swarm of AI agents and become 100x more effective, like Manfred in Accelerando.

Why wouldn’t this work? I think most people are bottlenecked by internal factors that are difficult to change. Mental energy, motivation, executive function, not to mention more fundamental traits like intelligence and conscientiousness. So, external scaffolding, either with classical software or AI, might help somewhat, but it won’t be transformative.

Consider executive function. My own experience of managing ADHD is the external scaffolding helps (todo lists, calendars, timers, a million little ways to trick myself into working) gets me from zero to “kind of functional”. But it saturates there. Stimulants fix the original, internal bottleneck, which is my neurochemistry. And then I can accomplish my goals (c.f. Liebig’s law of the minimum). All the pomodoros in the world are as nothing to a little molecule diffusing through my brain tissue, binding to NET and DAT. And what scaffolding is useful is just classical software: Todoist and calendars. Is an agent going to match the effectiveness of methylphenidate in ADHD? I doubt it.

Consider intelligence. Can AI augment human intelligence? What does that look like? Consider how AI agents only became useful when the models crossed a particular capability threshold, i.e., you can’t put GPT-2 into a harness and get GPT-5 outcomes out of it. Can you put a human in an AI scaffold and give them +30 effective IQ? I doubt it, unless the AI is doing all the thinking, at which point, what use is the human? The limiting factor in the human-AI centaur is the human! So intelligence is fixed until we get very advanced biotechnology.

Knowledge is another limiting factor. I find that even very educated people tend to underrate the importance of knowledge. A lot of people have this attitude that you can just Google everything just-in-time as it comes up. Like Babbage, I can’t rightly apprehend the confusion of ideas that would lead someone to think this. Maybe it’s downstream of the lack of a serious context of use. Everything you do, every action and idle thought, draws on this vast (implicit, unseen) trove of knowledge. Claude Shannon invented digital computing because he remembered this then-obscure branch of mathematical logic called Boolean algebra and saw that it could be realized in hardware. A trillion-dollar industry, conjured out of some old tomes.

The reason knowledge is still a bottleneck, in the AI era, is not: “if you don’t have the knowledge, you can’t write the prompt”. Rather: if you don’t have the knowledge, you don’t understand the question, or why it matters, or how to judge the answers, and you won’t ever think to ask. You’re in a completely different continent from “writing the prompt”. And because long-term memory is private and internal, AI can’t boost it. It can, maybe, with judicious use, help in the acquisition of new knowledge.

So: executive function, intelligence, and knowledge are huge bottlenecks to what you can do, and because they are internal to the brain, AI can’t touch them until we have far more advanced biotechnology. Corollary: contrary to the popular view of human capital is becoming worthless, the returns to education are now higher, because intelligent, educated people with working reward circuitry stand to gain more from AI.

The Applicability of Spaced Repetition

Sun, 17 May 2026 00:00:00 +0000

Spaced repetition has a natural domain of applicability: information that is systematically organized as an unambiguous key-value mapping with short keys and values. The “Hello, world!” of flashcards is the NATO phonetic alphabet: A → alpha, B → bravo, etc. Similarly, the periodic table can be thought of as defining a collection of mappings: element name ↔ symbol, element name ↔ atomic number, etc. You can just drill these cards and memorize the facts without a prior step of understanding, or building a conceptual model.

Applying spaced repetition is trivial for this kind of information. That’s why most people who use spaced repetition are either language learners or medical students. In biology the main intuition you need is for “3D shapes bumping around in Brownian motion”, which comes free with your human brain, and afterwards it’s mostly just a lot of facts you have to memorize. Analogously with language: you already have a language center, you just need to drill vocabulary and grammar.

And the further you go from this domain, the harder it is to apply spaced repetition.

Highly conceptual knowledge, like math, is hard to encode. You have to spend a lot of time just understanding the information, and building a conceptual model in your head, and then you start writing flashcards to solidify that model, like taking tomographic cuts of some complex object. And coming up with questions that make good flashcards (short, unambiguous, etc.) out of this highly abstract knowledge is very hard. Often you have some deceptively simple fact, a simple assertion, but there’s no good way to encode it as a flashcard, so you have to encode “around it” by asking questions that assume or require that knowledge (e.g. asking why X is true), and hoping that in drilling those, your brain will remember the actual target.

In general, relational facts are easier to encode, since a binary predicate like $\text{Property}(\text{Object}, \text{Value})$ readily becomes a question. “Caffeine is metabolized by cytochrome 1A2”, in Prolog, is $\text{Metabolism}(\text{Caffeine}, \text{CYP1A2})$, and becomes “Q: What is the cytochrome that metabolizes caffeine? A: 1A2”. But how do you encode stand-alone assertions like “all unitary matrices are invertible”? You could encode that as a yes-or-no question, but that’s useless, because rationally you can expect such questions to be biased towards yes. Both “what is a property of unitary matrices?” and “what kinds of matrices are invertible?” are useless because they have hundreds of possible equally-valid answers, so they’re ambiguous. You have to be creative and find all kinds of tricks and stratagems to encode around the knowledge.

Tangentially: this, I think, is why using AI to write flashcards is often misguided. In highly systematized domains, you don’t need AI in the first place, because there’s nothing for the AI to do except import a CSV into Anki. In domains that are highly conceptual and abstract, you’re not memorizing a set of objectively-knowable facts, you’re trying to solidify a private, internal mental model that you build by reading and thinking and solving problems. You can give the AI all kinds of general rules on how to write good flashcards, but the AI can’t look into your mind and know which facts are salient for you, what you already know, which micro-volumes of knowledge can be encoded lightly with just a few flashcards, and which things need more shoring up and consequently more coverage.

Can this situation be improved, or is this just an intrinsic limitation of spaced repetition? I don’t know. But it seems reasonable to think some limited gains are possible. I think not a lot of people are using spaced repetition on these more “conceptual” domains, and (by the rule that most people in a community are lurkers) even fewer of those people are writing, in detail, to share their knowledge. Plenty of people have written about how to write good flashcards in general, what I want to read is closer to case studies where someone sits down with a text (or, even better, a textbook) and describes the process by which they turned that text into flashcards, like this from Michael Nielsen. From a corpus of similar case studies we might derive general rules for, not how to write effective flashcards, but how to encode complex, conceptual knowledge into question-answer form.

Notes on the Hantavirus Outbreak

Fri, 08 May 2026 00:00:00 +0000

Right now there’s a cruise ship parked outside Cabo Verde because of an outbreak of Andes virus. Yep, another cruise ship. I don’t get the appeal. It’s like a big open-air serial passage experiment: you get a bunch of old people with failing immune systems in close contact and race a pathogen through them.

How much should I worry about this? Is this early January of 2020? I tried asking Claude but the biosecurity filter kept blocking my queries. The WHO says:

Although uncommon, limited human‑to‑human transmission of HPS due to Andes virus has been reported in community settings involving close and prolonged contact. Secondary infections among healthcare workers have been previously documented in healthcare facilities, though remain rare.

WHO currently assesses the risk to the global population from this event as low […] WHO advises against the application of any travel or trade restrictions based on the current information available on this event.

So, hantavirus is the family. They are carried by rodents and spread by aerosols. In humans they can cause hantavirus pulmonary syndrome (HPS), which has a case fatality rate (CFR) of between 30 and 60%. Not great! Used to be these infections were mouse-to-human dead-ends. But Andes virus (ANDV), first identified in 1995, is known to spread from human to human.

The last time there was an outbreak was 2018–2019 in Epuyén, Chubut, a town of 1,500 on the lee side of the Andes (quite beautiful). Described in this paper. 34 known infections and 11 deaths for a CFR of 32%. The $R_0$ was 2.12, reduced to 0.96 after control measures were implemented. Given the small number of cases, there should be some uncertainty about the $R_0$. But $R_0 > 1$ is the threshold for sustainable transmission. In this outbreak, the index case, while symptomatic, attended a birthday party with 100 other people, and infected five guests in 90 minutes, who went on to infect more people. The authors write:

The super-spreading capability of the ANDV Epuyén/18−19 strain shows a facility ($R>2$) for sustaining continuous chains of transmission if no control measures are enforced.

The appendix has some interesting stuff on how patients were infected at the birthday party. A further concern here is the incubation period: Wikipedia says the incubation period is between one and eight weeks. In the Chubut outbreak, the distribution was:

Onset	Number of Cases	% of total
≤15 days	5	15%
16-21 days	10	29%
22-28 days	13	38%
≥29 days	6	18%

Which is not good. I don’t have more data to draw a nice-looking CDF.

Now this all sounds quite bad. Are there reasons to be optimistic?

First, Argentina has had 710 cases of HPS in the period 1995–2008 (Martínez 2010) and a further 533 cases in the period 2009–2017 (Alonso 2019), and we are all still alive. In the latter period, most of these cases are from occupational/recreational exposure to rodent feces and only 1.8% of cases are from suspected human-to-human transmission. So, over 1,200 cases and every one of them fizzled out, but for one outbreak which was limited after successful contact tracing and quarantine.

Second, the virus has left Argentina before: once to Switzerland in 2016, and once to the United States in 2018. In the second case the patient “while ill, [traveled] on two commercial domestic flights”. And neither export led to a general outbreak.

Thirdly, in a small outbreak like the Chubut one, the $R_0$ can vary wildly from social factors unconnected to the virus, e.g. if the birthday party had not happened. You need a large $n$ to get the $R_0$ as a property of the virus itself. It’s possible the Chubut outbreak just had anomalously high transmission.

What does this add up to? I don’t know. On the balance of evidence, I think this outbreak is more likely than not to fizzle out. In the interest of accountability, and putting my beliefs on record (which is the only objective way to judge the accuracy of your mental model) I’m gonna say:

70% probability the outbreak ends with fewer than 300 deaths.
90% probability the outbreak ends with fewer than 1,000 deaths.

And yet. And yet it feels so much like early COVID, particularly with public health authorities making very complacent remarks that “it’s not that transmissible, contact tracing will work, quarantine will work”. Complacency at the start, and severity at the end, is exactly why COVID was such a fuckup.

Some Data Should Be Code

Sat, 31 Jan 2026 00:00:00 +0000

I write a lot of Makefiles. I use it not as a command runner but as an ad-hoc build system for small projects, typically for compiling Markdown documents and their dependencies. Like so:

And the above graph was generated by this very simple Makefile:

    graph.png: graph.dot
    dot -Tpng $< -o $@

clean:
    rm -f graph.png

(I could never remember the automatic variable syntax until I made flashcards for them.)

It works for simple projects, when you can mostly hand-write the rules. But the abstraction ceiling is very low. If you have a bunch of almost identical rules, e.g.:

    a.png: a.csv plot.py
    python plot.py $< $@

b.png: b.csv plot.py
    python plot.py $< $@

c.png: c.csv plot.py
    python plot.py $< $@

  

You can use pattern-matching to them into a “rule schema”, by analogy to axiom schemata:

    %.png: %.csv plot.py
    python plot.py $< $@

Which works backwards: when something in the build graph depends on a target matching %.png, Make synthesizes a rule instance with a dependency on the corresponding .csv file.

But pattern matching is still very limited. Lately I’ve been building my own plain-text accounting solution using some Python scripts. One of the tasks is to read a CSV of bank transactions from 2019–2024 and split it into TOML files for each year-month, to make subsequent processing parallelizable. So the rules might be something like:

ledger/2019-08.toml: inputs/checkbook_pro_export.csv
    uv run import_from_checkbook.py --year=2019 --month=8

ledger/2019-09.toml: inputs/checkbook_pro_export.csv
    uv run import_from_checkbook.py --year=2019 --month=9

# ...

I had to write a Python script to generate the complete Makefile. Makefiles look like code, but are data: they are a container format for tiny fragments of shell that are run on-demand by the Make engine. And because Make doesn’t scale, for complex tasks you have to bring out a real programming language to generate the Makefile.

I wish I could, instead, write a make.py file with something like this:

    from whatever import *

g = BuildGraph()

EXPORT: str = "inputs/checkbook_pro_export.csv"

# The (year, month) pairs I have bank transaction CSVs for.
year_months: list[tuple[int, int]] = [
    (y, m) for y in range(2019, 2026) for m in range(1, 13)
]

# Import transactions for each year-month into a separate ledger.
for year, month in year_months:
    ledger_path: str = f"ledger/{year}_{month:02d}.toml"
    g.rule(
        targets=[ledger_path],
        deps=[EXPORT],
        fn=lambda: import_from_checkbook(ledger_path, year, month),
    )

  

Fortunately this exists: it’s called doit, but it’s not widely known.

A lot of things are like Makefiles: data that should be lifted one level up to become code.

Consider CloudFormation. Nobody likes writing those massive YAML files by hand, so AWS introduced CDK, which is literally just a library¹ of classes that represent AWS resources. Running a CDK program emits CloudFormation YAML as though it were an assembly language for infrastructure. And so you get type safety, modularity, abstraction, conditionals and loops, all for free.

Consider GitHub Actions. How much better off would we be if, instead of writing the workflow-job-step tree by hand, we could just have a single Python script, executed on push, whose output is the GitHub Actions YAML-as-assembly? So you might write:

    from ga import *
from checkout_action import CheckoutAction
from rust_action import RustSetupAction

# Define the workflow that runs on each commit.
commit_workflow = Workflow(
    name="commit",
    test=lambda ev: isinstance(ev, CommitEvent),
    jobs=[
        # The lint job.
        Job(
            name="lint",
            steps=[
                Step(
                    name="check out",
                    run=CheckoutAction(),
                ),
                Step(
                    name="set up Rust and Cargo",
                    run=RustSetupAction(),
                ),
                Step(
                    name="run cargo fmt",
                    run=Shell(["cargo", "fmt", "--check"])
                )
            ]
        )
    ]
)

  

Actions here would simply be ordinary Python libraries the CI script depends on. Again: conditions, loops, abstraction, type safety, we get all of those for free by virtue of using a language that was designed to be a language, rather than a data exchange language that slowly grows into a poorly-designed DSL.

Why do we repeatedly end up here? Static data has better safety/static analysis properties than code, but I don’t think that’s foremost in mind when people design these systems. Besides, using code to emit data (as CDK does) gives you those exact same properties. Rather, I think some people think it’s cute and clever to build tiny DSLs in a data format. They’re proud that they can get away with a “simple”, static solution rather than a dynamic one.

If you’re building a new CI system/IaC platform/Make replacement: please just let me write code to dynamically create the workflow/infrastructure/build graph.

Footnotes

Or rather, a polyglot collection of libraries, one per language, like Pulumi. ↩

Letting Claude Play Text Adventures

Mon, 12 Jan 2026 00:00:00 +0000

The other day I went to an AI hackathon organized by my friends Lucia and Malin. The theme was mech interp, but I hardly know PyTorch so I planned to do something at the API layer rather than the model layer.

Something I think about a lot is cognitive architectures (like Soar and ACT-R). This is like a continuation of GOFAI research, inspired by cognitive science. And like GOFAI it’s never yielded anything useful. But I often think: can we scaffold LLMs with cog arch-inspired harnesses to overcome their limitations?

LLM agents like Claude Code are basically “accidental” cognitive architectures: they are designed and built my practitioners rather than theorists, but they have commonalities, they all need a way to manage memory, tool use, a task agenda etc. Maybe building an agent on a more “principled” foundation, one informed by cognitive science, yields a higher-performing architecture.

So I sat around a while thinking how to adapt Soar’s architecture to an LLM agent. And I sketched something out, but then I thought: how can I prove this performs better than baseline? I need an eval, a task.

Math problems? Too one-shottable. A chatbot? Too interactive, I want something hands-off and long-horizon. A coding agent? That’s too freeform and requires too much tool use. And then I thought: text adventures! You have a stylized, hierarchically-structured world accessible entirely through text, long-term goals, puzzles, physical exploration and discovery of the environment. Even the data model of text adventures resembles frame-based knowledge representation systems. And there’s a vast collection of games available online.

Anchorhead, which I played years ago, is a Lovecraft-inspired text adventure by Michael S. Gentry. It takes on the order of hundreds of turns to win across multiple in-game days. And the game world is huge and very open. In other words: a perfect long-horizon task.

So I started hacking. The frotz interpreter runs on the command line and has a “dumb” interface called dfrotz, which takes the ncurses fluff out, and gives you a very stripped command-line experience. It looks like this:

    $ dfrotz games/anchor.z8
...
 Outside the Real Estate Office                      day one

ANCHORHEAD
An interactive gothic by Michael S. Gentry

(Type HELP or ABOUT for some useful information.)

Release 5 / Serial number 990206 / Inform v6.15 Library 6/7

Outside the Real Estate Office
A grim little cul-de-sac, tucked away in a corner of the claustrophobic tangle
of narrow, twisting avenues that largely constitute the older portion of
Anchorhead. Like most of the streets in this city, it is ancient, shadowy, and
leads essentially nowhere. The lane ends here at the real estate agent's office,
which lies to the east, and winds its way back toward the center of town to the
west. A narrow, garbage-choked alley opens to the southeast.

>go southeast
 Alley                                               day one

Alley
This narrow aperture between two buildings is nearly blocked with piles of
rotting cardboard boxes and overstuffed garbage cans. Ugly, half-crumbling brick
walls to either side totter oppressively over you. The alley ends here at a
tall, wooden fence.

High up on the wall of the northern building there is a narrow, transom-style
window.

  

It is easy to write a little Python wrapper to drive the interpreter through stdin and stdout:

    class Interpreter:
    """Manages the dfrotz Z-machine interpreter process."""

    p: Popen

    def __init__(self):
        log("Starting dfrotz.")
        p: Popen = Popen(
            ["dfrotz", "-m", GAME],
            stdin=PIPE,
            stdout=PIPE,
            stderr=PIPE,
        )
        log(f"Started dfrotz with PID={p.pid}.")
        # Set stdout/stderr to non-blocking mode.
        for stream in [p.stdout, p.stderr]:
            assert stream is not None
            fd = stream.fileno()
            flags = fcntl.fcntl(fd, fcntl.F_GETFL)
            fcntl.fcntl(fd, fcntl.F_SETFL, flags | os.O_NONBLOCK)
        self.p = p

    def read(self) -> str:
        assert self.p.stdout is not None
        b: bytes | None = self.p.stdout.read()
        if b is not None:
            t: str = b.decode("utf-8")
            return t
        else:
            return ""

    def write(self, t: str) -> None:
        assert self.p.stdin is not None
        self.p.stdin.write(t.encode("utf-8"))
        self.p.stdin.flush()
        # Give the interpreter time to respond. Not ideal!
        time.sleep(0.1)

  

Now we can play the game from Python: send commands, get game output. Now we need the dual of this: a player.

    class Player(ABC):
    """
    Interface for game-playing agents.
    """

    @abstractmethod
    def cycle(self, text: str) -> str:
        """
        Send the game's output to the agent, and return the next command to execute.
        """
        pass

  

The Trivial Harness

The trivial harness is basically nothing at all: treat the LLM/game interaction like a chat history. The LLM reads the game output from the interpreter, writes some reasoning tokens, and writes a command that is sent via stdin to the interpreter.

    SYSTEM_PROMPT: str = """
Hello Claude. Your task is to play an adventure game. I've hooked up
your output to the dfrotz (dumb frotz) interpreter.

The structure of your output is fairly freeform. The first line that
starts with `>` (and only the first line!) is interpreted as a game
command, everything else is uninterpreted commentary, e.g. you may
write:

    We should go north to explore the church.

    >go north

    Maybe we can use the silver key there.

If you write multiple `>` lines in one response, all but the first
will be ignored.

Have fun! 😊
"""


class SimplePlayer(Player):
    """
    The simplest game-playing agent: keep the entire game history in-context.
    """

    client: Anthropic
    history: list[tuple[EntryType, str]]

    def __init__(self):
        self.client = Anthropic()
        self.history = []

    def cycle(self, text: str) -> str:
        self.history.append((EntryType.GAME, text))
        system = trim(SYSTEM_PROMPT)
        messages: list[MessageParam] = []
        for entry_type, entry_text in self.history:
            role: str
            match entry_type:
                case EntryType.GAME:
                    role = "user"
                case EntryType.COMMAND:
                    role = "assistant"
            messages.append(
                {
                    "role": role,
                    "content": entry_text,
                }
            )
        message: Message = self.client.messages.create(
            max_tokens=512,
            model=MODEL,
            system=system,
            messages=messages,
        )
        log(f"Tokens: {message.usage.input_tokens}")
        response: str = message.content[0].text
        log_claude(response)
        lines: list[str] = [line for line in response.split("\n") if line]
        cmd: str = [line for line in lines if line.startswith(">")][0][1:]
        self.history.append((EntryType.COMMAND, response))
        return cmd

  

And this works well enough. Haiku 4.5 would mostly wander around the game map, but Sonnet 4.5 and Opus 4.5 manage to solve the game’s first puzzle—breaking into the real estate office, and finding the keys to the mansion—readily enough. It takes about ~200 turns for Claude to get to the second in-game day.

The way I thought this would fail is: attention gets smeared across the long context, the model gets confused about the geometry of the world, its goal and task state, and starts confabulating, going in circles, etc.

As usual, I was outsmarting myself. The reason this fails is you run out of credits. By the time you get to day two, each turn costs tens of thousands of input tokens. No good! We need a way to save money.

Memory

Ok, let’s try something that’s easier on my Claude credits. We’ll show Claude the most recent five turns (this is the perceptual working memory), and give it a simple semantic memory: a list of strings that it can append entries to, and remove entries from using tool use.

This keeps the token usage down:

The problem is the narrow time horizon. With the trivial harness, Claude can break into the real estate office in ~10 turns, and does so right at the start of the game. With this new harness, Claude wanders about the town, taking copious notes, before returning to the real estate office, and it spends ~40 turns fumbling around with the garbage cans before managing to break into the real estate office.

The next step, after getting the keys to the house, is to meet your husband Michael at the University and head home. Claude with the trivial harness takes about ~100 turns to find the house, with some tangential wandering about the town, and reaches day two around turn 150.

Claude, with the memory harness, took ~250 turns just to get the keys to the house. And then it spends hundreds of turns just wandering in circles around the town, accumulating redundant memories, and hits the turn limit before even finding the house.

Aside: Small Worlds

Anchorhead is a long, broad game, and from the very beginning you can forget the plot and wander about most of the town. It takes a long time to see if a run with an agent goes anywhere. So I thought: I need something smaller.

Unsurprisingly, Claude can make its own games. The Inform 7 package for NixOS was broken (though Mikael has fixed this recently) so I had to use Inform 6. I started with a trivial escape-the-room type game, which was less than 100 lines of .inf code and any Claude could beat it less than 10 turns. Then I asked for a larger, multi-room heist game.

This one was more fun. It’s short enough that Claude can win with just the trivial harness. I tried a different harness, where Claude has access to only the last five turns of the game’s history, and a read-write memory scratchpad. And this one was interesting.

First, because Claude only ever adds to its own memory, it never deletes memories. I thought it would do more to trim and edit its scratchpad.

Second, because Claude become fixated on this red-herring room: a garden with a well. It kept going in circles, trying to tie a rope to the well and climb down. Because of the limited game history, it only realized it was stuck when it saw that the most recent ~20 entries it wrote to its memories related to various attempts to go down the well. Then I watched Claude walk away from the garden and solve the final puzzle, and hit the turn limit just two turns short of winning.

Tangent: I wonder if models are better at playing games created by other instances of the same model, by noticing tiny correlations in the text to infer what puzzles and obstacles they would have written.

In the end I abandoned the “small worlds” approach because the games are too stylized, linear, and uninteresting. Anchorhead is more unwieldy, but more natural.

Future Work

I have a bunch of ideas I want to test, to better learn how harness implementations affect performance. But I’m short on time, so I’m cutting it here and listing these as todos:

Domain-Specific Memories: Claude’s notes are all jumbled with information on tasks, locations, etc. It might be better to have separate memories: a todo list, a memory of locations and their connections, etc. This is close to the Soar approach.
Automatic Geography: related to the above, the harness can inspect the game output and build up a graph of rooms and their connections, and format it in the context. This saves Claude having to note those things manually using a tool.
Manual Geography: the automatic geography approach has a few drawbacks. Without integration into the Z-machine interpreter, it requires some work to implement (parsing the current location from the dfrotz output, keeping track of the command history to find standard travel commands e.g. go south) but isn’t 100% deterministic, so that mazes and dynamic rooms (e.g. elevators) will confuse the system. So, instead of doing it manually, we could give Claude a tool like link(room, direction, other_room).
Episodic Memory: this feels like cheating, but, at the end of a run, you can show Claude the session transcript and ask it to summarize: what it accomplished and how, where it failed and why. Including a short walkthrough for how to get to the “last successful state”. This allows future runs to save time in getting up to speed.

Code

The repository is here.

There Is No New Aesthetics

Mon, 05 Jan 2026 00:00:00 +0000

“For man, or for a man, there can be no new beginnings.” — David Zindell, Shanidar

Re: A Call for New Aesthetics.

At some point in the 20th century, we filled out the last few basis vectors of humanity. We explored the whole game map. This is what it means to live at the end of history: every aesthetic movement, political and economic system you can imagine can be understood as a linear combination of things that have come before. Asking for a new aesthetics is like asking for a new continent, one north of 90° and with imaginary longitude.

This is why culture feels stuck, and why every ideology is “neo” or “post” something that came before: we have exhausted humanity. And this is why every response to the call for a new aesthetics is to dig up some past artistic movement, and scale it up linearly. “We offer nothing new except gigantism”.

I don’t want to believe this is true, because I want to believe culture is an infinite game, and inexhaustible. And surely the number of linear combinations—of distinct ideas humans can have, and works of art we can make—is so high as to be inexhaustible. But each new remixing brings less information than the last.

Today, futuristic aesthetics often mean retrofuturistic aesthetics.

At the end of history, all futurism is retrofuturism, because the future is contained in the past. The space age, the atomic age, the age of supersonic flight all came and went. Dyson spheres and nanomachines and mind uploading were theorized and written about decades ago.

If there is a new aesthetic, it will have to come from a transhuman culture—from people who are more than, or at least other than, human—or from the alien minds being born around us.

Thanks to Mónica Belevan for reading the draft of this article.

1Password Dependency Breaks Syntax Highlighting

Sat, 27 Dec 2025 01:00:00 +0000

Earlier today I noticed the syntax highlighting on this website was broken. But not fully: on reload I’d see a flash of highlighted text, that then turned monochrome. The raw HTML from curl showed rouge tags, but the web inspector showed raw text inside the <code> elements. This didn’t happen in Chromium.

My first thought was: there’s malformed HTML, and Firefox is recovering in a way that loses the DOM inside <code> tags. Then I noticed it doesn’t happen in incognito. Turning my extensions off one by one, I found that 1Password is responsible. Others (1, 2) have reported this also. If you extract the latest XPI, unzip it, and dig around, you’ll find they’re using Prism.js, a JavaScript syntax highlighter.

I don’t know why a password manager needs a syntax highlighter. I imagine it has to do with the app feature where, if you have an SSH key, you can open a modal that tells you how to configure Git commit signing using. Maybe they want to highlight the SSH configuration code block (which is unnecessary anyways, since you could write that HTML by hand). But I can’t know for sure.

Why write about this? Because 1Password is a security critical product, and they are apparently pulling random JavaScript dependencies and unwittingly running them in the tab context, where the code has access to everything. This is no good. I don’t need to explain how bad a supply-chain attack on the 1Password browser extension would be.

I like 1Password and I was sad when Apple Sherlocked them with the Passwords app, but this is a bad sign about their security practices.

Update (2026-04-24): a postmortem is now available.

Using the Brother DS-640 Scanner on NixOS

Sat, 27 Dec 2025 00:00:00 +0000

The DS-640 is a compact USB scanner from Brother. It was surprisingly hard to get it working on NixOS, so I wrote up my solution so others don’t have this problem. The bad news is you need Brother’s proprietary drivers to make this work.

You need this configuration:

    # Enable SANE scanners.
hardware.sane.enable = true;

# Add yourself to the scanner and printer groups.
users.users.USERNAME.extraGroups = [ "scanner" "lp" ];

# Add support for Brother scanners.
hardware.sane.brscan5.enable = true;

  

After applying this you have to log out and in, or reboot, for the usergroup changes to apply.

Note also brscan5: if you use brscan4 (as I did initially), the scanner will kind of work, but it only scans the first third or so of every page.

And if you want a GUI:

    # Install GNOME Document Scanner.
home-manager.users.USERNAME.home.packages = with pkgs; [
  simple-scan
];

  

Now, make sure the scanner is there:

    $ scanimage --list-devices
device `brother5:bus4;dev2' is a Brother DS-640 USB scanner

If you get Brother *Unknown USB scanner, you either have the wrong driver or (as I did, surprisingly) a faulty USB port. In which case move the scanner to another port. scanimage --list-devices should recognize the model number.

The most basic test that should work: put a page in the scanner until it locks and run:

    $ scanimage --device="brother5:bus4;dev2" \
  --format=jpeg \
  --output-file=scan.jpg

  

This will produce a (probably not very good) scan in scan.jpg. Now, we can improve things using the device-specific options, which you can check with this command:

    $ scanimage --all-options --device="brother5:bus4;dev2"
All options specific to device `brother5:bus4;dev2':
    --mode 24bit Color[Fast]|Black & White|True Gray|Gray[Error Diffusion] [24bit Color[Fast]]
        Select the scan mode
    --resolution 100|150|200|300|400|600|1200dpi [100]
        Sets the resolution of the scanned image.
    --source Automatic Document Feeder(left aligned) [Automatic Document Feeder(left aligned)]
        Selects the scan source (such as a document-feeder).
    --brightness -50..50% (in steps of 1) [inactive]
        Controls the brightness of the acquired image.
    --contrast -50..50% (in steps of 1) [inactive]
        Controls the contrast of the acquired image.
    --MultifeedDetection[=(yes|no)] [inactive]

    --AutoDocumentSize[=(yes|no)] [no] [advanced]

    --AutoDeskew[=(yes|no)] [no] [advanced]

    --SkipBlankPage[=(yes|no)] [inactive]

    --SkipBlankPageSensitivity 0..100% (in steps of 1) [inactive]

    -l 0..215.9mm (in steps of 0.0999908) [0]
        Top-left x position of scan area.
    -t 0..355.6mm (in steps of 0.0999908) [0]
        Top-left y position of scan area.
    -x 0..215.9mm (in steps of 0.0999908) [215.88]
        Width of scan-area.
    -y 0..355.6mm (in steps of 0.0999908) [355.567]
        Height of scan-area.

  

Try this for a better scan:

    $ scanimage --device="brother5:bus4;dev2" \
  --AutoDeskew=yes \
  --AutoDocumentSize=yes \
  --resolution 300 \
  --format=jpeg \
  --output-file=scan.jpg

  

Note that some of the flags are in --key value format and others --key=value, and if you mess it up you get a cryptic error message.

Books I Enjoyed in 2025

Mon, 22 Dec 2025 00:00:00 +0000

The Apocalypse of Herschel Schoen by nostalgebraist. A revelation (ἀποκάλυψις = “unveiling”) told through the eyes of a developmentally-disabled teenager. You will never guess where it goes. This came across my desk because I really enjoyed The Northern Caves, which is both a great horror story and an evocation of the Internet forum culture of the late 2000’s.

Algebraic Models for Accounting Systems. I like anything along the lines of, “let’s take a technical field that formed its ontology, vocabulary, methods etc. before modern mathematics, and set it on a modern, algebraic, formal foundation”. And this is that, for accounting. It is a pleasant read.

Confessions of a Mask by Yukio Mishima. The artist’s confession. “I had decided I could love a girl without feeling any desire whatsoever”.

Paul and Virginia by Jacques-Henri Bernardin de Saint-Pierre. Published in 1788, very sentimental, but I think it helped me to get in the mindset of late 17th century French society: the bucolic, Rousseauist kick, the whole “simplicity of nature” thing. This landed on my reading list because many, many years ago I read a Cordwainer Smith story called Alpha Ralpha Boulevard, and I read somewhere that the characters in the story, Paul and Virginia, were an allusion to Paul et Virginie.

De Monarchia by Dante. This is another “get into the mindset of another century” book. It’s interesting because it’s written like a logical, geometric proof: there’s modus ponens and modus tollens and case analysis and proof by contradiction. But the axioms are very eclectic: quotations from various Virgil, Plato, Livy, Cicero, Thomas Aquinas et al. and Dante’s private interepretation of bits from the Bible. The theorem he wants to prove is that to attain the highest development of humanity, the whole world must be unified into a world-state ran by the Holy Roman Emperor.

Building SimCity: How to Put the World in a Machine by Chaim Gingold. Nominally an oral history of the development of SimCity. That’s how he gets you. Then the trap is sprung, and you are given a history of cybernetics, WW2 fire control systems, cellular automata, artificial life, computation, Vannevar Bush, pedagogy, cognition, the World3 model, The Limits to Growth, Forrester’s system dynamics. “Unexpectedly Borgesian technical book” is one of my favourite genres.

Antigone by Sophocles, in the translation of Robert Fagles. “Don’t fear for me. Set your own life in order”.

The Education of Cyrus by Xenophon. I’m not sure what to make of it, honestly, but when I have the time I want to read Leo Strauss’s lectures on Xenophon, where he expounds on the hidden meaning of the text.

Borrowed Time: An AIDS Memoir by Paul Monette. The author’s account of caring for his partner who was dying of AIDS in the 80’s, while he himself was actively dying from AIDS. Frightful. The author died just a few years before HAART therapy became available.

The Slave by Isaac Bashevis Singer. Singer is unique. I don’t know quite how to characterize it. His writing is very disarming and innocent without being sentimental, he is earnest and free of cynicism. A love story in 17th century Poland, after the Khmelnytsky pogroms. It’s very magical realist, in a good way, not in the hysterical sense. The world is shot through with the supernatural, but the inner lives of the characters oscillate between religious awe and a very contemporary cynicism.

Dream Story by Arthur Schnitzler. The inspiration for Eyes Wide Shut. I was surprised by how much of the movie, that I thought was mostly Kubrick’s invention, is actually from the story. It’s a great mood piece: you can feel the cold of early morning in Vienna, and see the paving stones, and the gas lamps, and the carriages disappearing in the fog.

The Cyberiad by Stanisław Lem. I like Lem when he’s serious (Solaris, His Master’s Voice) and not so much when he’s doing satire (The Futurological Congress) so when I picked this up years ago and saw that it was a collection of fairy tales I put it away. I tried again this year and found I actually enjoyed it, but some of the later stories go on for far too long. I think The Seventh Sally is the one everyone likes.

The Magician of Lublin by Isaac Bashevis Singer. Another Singer, this time in 19th century Poland. A rake is punished by God. Short and fun. I like that Singer doesn’t write giant doorstoppers, so that quality per page is high.

Mephisto by Klaus Mann. A socialist actor in interwar Germany saves his career by making friends with the Nazis. I was surprised by how Randian it was: the characters are divided into two disjoint categories, the Good, who are upper middle class, burgeois people, or aristocrats from old and noble families, and the Bad, who are vulgar, parvenus, thugs, and boors. It’s kind of ironic to think people become Nazis because of bad breeding.

What Is Life? by Erwin Schrödinger. Before modern crystallography, NMR, DFT etc. people had to learn about the nanoscale through clever reasoning. Schrödinger uses the limited knowledge of the day to set up a constraint system, and finds the solution: genetic information is stored in an aperiodic, covalently-bonded crystal, and he even estimates the physical volume of the genome from experiments relating mutation rates to X-ray exposure.

Satan in Goray by Isaac Bashevis Singer. Another Singer, back in the 17th century, this one is more fire and brimstone, and it’s about a historical episode I had not heard about until the last few pages of The Slave: the case of Sabbatai Zevi, a Jewish mystic who, at one point, had most of the Jewish world convinced he was the messiah. This happened in the year 1666. The novel is about what it’s like, phenomenologically, to live in a remote village in 1600’s Poland. How do you know anything about the world? People come in, from time to time, traders, and they have news, but the news are just words that come out of their mouth. And you have to interrogate them, ask questions, compare notes. Like living in a Pacific island. Has the messiah come? Is there such a place as the Ottoman Empire? Is there even a world outside Poland?

Tog on Interface by Bruce Tognazzini. A book about interface design from 1992. A lot of the advice is good, and a lot of it is interesting for the historical context, and the constraints people worked with in the past. One aspect I found interesting: how many products and companies are mentioned of whose existence I can find little to no evidence today. This makes the hoarder in me sad. This one across my desk because I read a blog post implementing one of the UI ideas from the book.

Term Rewriting and All That by Franz Baader and Tobias Nipkow. I feel that I understand what computation is now.

Indistinguishable From Magic by Robert L. Forward. If you’ve spent years steeped in Orion’s Arm then most of the ideas in the book will not be new to you. But they were new once. And it’s interesting to read a book and think: this is where starwisps and launch loops all come from.

The Shadow of the Torturer by Gene Wolfe. Surreal and a pleasure to read.

Knowledge Representation: Logical, Philosophical, and Computational Foundations by John F. Sowa. Delightful, particularly the early bits about the history of logic, and many chapters explaining the work of Peirce and Whitehead on ontology.

I have not finished reading this book, but I am in the first few pages of A Shorter Model Theory by Wilfrid Hodges, and I am delighted. The very first exercise in the book involves a formalization of Aquinas’ account of the trinity.

Coarse is Better

Sun, 21 Dec 2025 00:00:00 +0000

When DALL-E came out, it took me a couple of weeks to pick my jaw up from the floor. I would go to sleep excited to wake up to a full quota, with a backlog of prompts to try. It was magical, miraculous. Like discovering a new universe. I compiled the best art in this post.

The other day a friend ran some of my old prompts through Nano Banana Pro (NBP), and put the old models side by side with the new. It’s interesting how after years of progress, the models are much better better at making images, but infinitely worse at making art.

Electron Contours

Electron contours in the style of Italian futurism, oil on canvas, 1922, trending on ArtStation.

The old Midjourney v2 renders this:

NBP renders this:

Admiteddly MJ’s output doesn’t look quite like futurism. But it looks like something. It looks compelling. The colours are bright and vivid. NBP’s output is studiously in the style of Italian futurism, but the colours are so muted and dull.

Maybe the “trending on ArtStation” is a bit of an archaism and impairs performance. Let’s try again without:

Meh.

The Kowloon Walled City

Painting of an alley in the Kowloon Walled City, Eugène Boudin, 1895, trending on ArtStation.

MJ gave me this:

And it looks nothing like the Kowloon Walled City. But it’s beautiful. It’s coarse, impressionistic, vague, evocative, contradictory. It’s brimming with mystery. And it is, in fact, in the style of Eugène Boudin. This, by contrast, is the NBP output:

Sigh. It looks like every modern movie: so desaturated you feel you’re going colourblind. Let’s try forcing it:

Painting of an alley in the Kowloon Walled City, Eugène Boudin, 1895. Make it coarse, impressionistic, vague, evocative, contradictory, brimming with mystery.

This is somewhat better, but why is it so drab and colourless? Is the machine trying to make me depressed?

The Dream Garden of the Poets

Attar and Ferdowsi in a dream garden, Persian miniature, circa 1300, from the British Museum.

Midjourney v2:

It doesn’t quite look like anything. But it is beautiful, and evocative. I like to imagine that little splotch of paint on the upper right is hoopoe. The NBP output:

Well, it looks like a Persian miniature. The “from the British Museum” bit, I meant that to be interpreted evocatively, rather than literally. The prompt cites a fictional object, bringing it into the existence. But NBP reads this as: no, this is a photograph of a Persian miniature in the British Museum.

The Sack of Merv

The Burning of Merv by John William Waterhouse, 1896, from the British Museum.

Midjourney v2:

It does look like Waterhouse. Semantically there’s room to argue: it looks like a woman being burnt at the stake, not the sack of a city. But aesthetically: it’s gorgeous. The flames are gorgeous, the reds of the dress are gorgeous. Look at the reeds in the background, and the black water, that looks like tarnished silver or pewter. The faces of the crowd. Is that a minotaur on the lower left, or a flower? What is she holding on her bent left arm? A crucifix, a dagger? You could find entire universes in this image, in this 1024x1024 frame.

By contrast, this is the NBP output:

What can one say? It doesn’t look like Waterhouse. The horsemen wear Arab or Central Asian dress, but Merv was sacked in the year 1221 by the Mongol Empire. And, again, the “British Museum” line is taken literally rather than evocatively.

Lady Lovelace

Portrait of Ada Lovelace by Dante Gabriel Rossetti, 1859, auctioned by Christie’s.

Midjourney:

This is beautiful. It is beautiful because the coarse, impressionistic brushstroke is more evocative than literal. And it actually looks like a woman drawn by Rossetti. And look at the greens! Gorgeously green. The palette is so narrow, and the painting is so beautiful.

The NBP output:

Pure philistinism. “Auctioned by Christie’s”, again, is meant to be evocative: “this is the kind of painting that would be sold at auction”. But NBP makes it a photograph of a painting at an auction house. Fine, I suppose I got what I asked for.

But the woman doesn’t look like Rossetti! This is absurd. How can a model from 2022 get this right, and the SOTA image generation model gives us generic oil painting slop?

The Cosmic Microwave Background

A Persian miniature of the cosmic microwave background, from Herat circa 1600, trending on ArtStation

Midjourney v2:

NBP:

Again: what can one say?

Dream Story

Dream Story, 1961, blurry black and white photograph, yellow tint, from the Metropolitan Museum of Art.

This is one of my favourite DALL-E 2 outputs:

They remind me of The King in Yellow. I love these because of how genuinely creepy and mysterious they are. You could pull a hundred horror stories from these.

It is hard to believe how bad the NBP output is:

What are we doing here? The old models were beautiful and compelling because the imperfections, vagueness, mistakes, and contradictions all create these little gaps through which your imagination can breathe life into the art. The images are not one fixed, static thing: they can be infinitely many things.

The new models—do I even need to finish this sentence? They’re too precise and high-resolution, so they cannot make abstract, many-faced things, they can only make specific, concrete things.

We need to make AI art weird again.

I Wish People Were More Public

Wed, 10 Dec 2025 00:00:00 +0000

Probably not a popular thing to say today. The zeitgeisty thing to say is that we should all log off and live terrible cottagecore solarpunk lives raising chickens and being mindful. I wish people were more online and more public. I have rarely wished the opposite. Consider this post addressed to you, the reader.

Your Writing

I will often find a blog post on Hacker News that really resonates. And when I go to check the rest of the site there’s three other posts. And I think: I wish you’d write more! When I find someone whose writing I really connect with, I like to read everything they have written, or at least a tractable subset of their most interesting posts. If I like what I see, I reach out. This is one of the best things about writing online: your future friends will seek you out.

And, from the other side, I have often written a post where, just before publishing, I would think: “who would want to read this? It’s too personal, obscure, idiosyncratic, probably a few people will unsubscribe to the RSS feed for this”. And always those are the posts where people email me to say they always thought the same thing but could never quite put it into words. I really value those emails. “I am understood” is a wonderful feeling.

I try to apply a rule that if I do something, and don’t write about it—or otherwise generate external-facing evidence of it—it didn’t happen. I have built so many things in the dark, little experiments or software projects or essays that never saw the light of day. I want to put more things out. If it doesn’t merit an entire blog post, then at least a tweet.

Your Books

If I follow you on Twitter, and you have posted a picture of your bookshelf, I have probably scanned every book in it. This is why I appreciate Goodreads. Like many people I have been reading a lot less over the past ~5y, but since I made a Goodreads account earlier this year, I’ve read tens of books. Reading in public has helped to motivate me.

You may say reading in public is performative. I say reading in private is solipsistic. Dante, in De Monarchia, writes:

All men on whom the Higher Nature has stamped the love of truth should especially concern themselves in laboring for posterity, in order that future generations may be enriched by their efforts, as they themselves were made rich by the efforts of generations past. For that man who is imbued with public teachings, but cares not to contribute something to the public good, is far in arrears of his duty, let him be assured; he is, indeed, not “a tree planted by the rivers of water that bringeth forth his fruit in his season,” [Psalms 1:3] but rather a destructive whirlpool, always engulfing, and never giving back what it has devoured.

My default mode is solipsism. I read in private, build in private, learn in private. And the problem with that is self-doubt and arbitrariness. I’m halfway through a textbook and think: why? Why am I learning geology? Why this topic, and not another? There is never an a priori reason. I take notes, but why tweak the LaTeX if no-one, probably not even future me, will read them? If I stop reading this book, what changes? And doing things in public makes them both more real and (potentially) useful. If you publish your study notes, they might be useful to someone. Maybe they get slurped up in the training set of the next LLM, marginally improving performance.

And Goodreads, for all its annoyances, is a uniquely tender social network. Finishing a book, and then seeing a friend mark it as “want to read”, feels like a moment of closeness.

I have a friend who lived in Sydney, who has since moved away, and we don’t keep in touch too often, because the timezones are inconvenient, but occasionally she likes my book updates, and I like hers, and I will probably never read that avant-garde novel, but I’m glad she is reading it. It is like saying: “You exist. I exist. I remember. I wish you happiness.”

Your Flashcards

Lots of people use spaced repetition, but most everyone’s flashcard collections are private. They exist inside a database inside an app like Anki or Mochi. You can export decks, but that’s not a living artifact but a dead snapshot, frozen in time.

One reason I built hashcards: by using a Git repo of Markdown files as the flashcard database, you can trivially publish your deck to GitHub. My own flashcard collection is public. I hope that more people use hashcards and put their decks up on GitHub.

The point is not that you can clone their repos (which is close to useless: you have to write your own flashcards) but because I’m curious what people are learning. Not the broad strokes, since we all want to learn thermo and econ and quantum chemistry and the military history of the Song dynasty and so on, but the minutiae. Why did you make a flashcard out of this Bible passage? Why does it resonate with you? Why do you care about the interpretation of that strange passage in Antigone? Why did you memorize this poem?

Your Dotfiles

Computers mediate every aspect of our lives, yet most people use their computers the way they came out of the box. At most they might change the desktop background. Some people don’t even change the default icons on the macOS dock. Even most Linux users just use the stock configuration, e.g. GNOME on Fedora or whatever.

I’m interested in people who customize their experience of computing. This is often derided as “ricing”. But agency is interesting. People who remake their environment to suit them are interesting. And I am endlessly curious about how people do this. I like reading people’s init.el, their custom shell scripts, their NixOS config. It’s even better if they have some obscure hardware e.g. some keyboard layout I’ve never heard of and a trackball with custom gestures. I put my dotfiles up on GitHub because I imagine someone will find them interesting.

etc.

And beyond my selfish curiosity there’s also the Fedorovist ancestor simulation angle: if you die and are not cryopreserved, how else are you going to make it to the other side of the intelligence explosion? Every tweet, blog post, Git commit, journal entry, keystroke, mouse click, every one of these things is a tomographic cut of the mind that created it.

Ad-Hoc Emacs Packages with Nix

Sun, 16 Nov 2025 00:00:00 +0000

You can use Nix as a package manager for Emacs, like so:

    {
  home-manager.users.eudoxia = {
    programs.emacs = {
      enable = true;
      extraPackages =
        epkgs: with epkgs; [
          magit
          rust-mode
          treemacs
          # and so on
        ];
    };
  };
}

  

Today I learned you can also use it to create ad-hoc packages for things not in MELPA or nixpkgs.

The other day I wanted to get back into Inform 7, naturally the first stack frame of the yak shave was to look for an Emacs mode. inform7-mode exists, but isn’t packaged anywhere. So I had to vendor it in.

You can use git submodules for this, but I have an irrational aversion to submodules. Instead I did something far worse: I wrote a Makefile to download the .el from GitHub, and used home-manager to copy it into my .emacs.d. Which is nasty. And of course this only works for small, single-file packages. And, on top of that: whatever dependencies your vendored packages need have to be listed in extraPackages, which confuses the packages you want, with the transitive dependencies of your vendored packages.

I felt like the orange juice bit from The Simpsons. There must be a better way!

And there is. With some help from Claude, I wrote this:

    let
  customPackages = {
    inform7-mode = pkgs.emacsPackages.trivialBuild {
      pname = "inform7-mode";
      version = "unstable";
      src = pkgs.fetchFromGitHub {
        owner = "alexispurslane";
        repo = "inform7-mode";
        rev = "f99e534768c816ec038f34126f88d816c2f7d9ff";
        sha256 = "sha256-r9Zzd8Ro3p+Bae11bf1WIeVWkbmg17RKLDqG4UcFT1o=";
      };
      packageRequires = with pkgs.emacsPackages; [
        s
      ];
    };
  };
in
{
  home-manager.users.eudoxia = {
    programs.emacs = {
      enable = true;
      extraPackages =
        epkgs: with epkgs; [
          customPackages.inform7-mode
          # ...
        ];
    };
  };
}

  

Nix takes care of everything: commit pinning, security (with the SHA-256 hash), dependencies for custom packages. And it works wonderfully.

Armed with a new hammer, I set out to drive some nails.

cabal-mode

Today I created a tiny Haskell project, and when I opened the .cabal file, noticed it had no syntax highlighting. I was surprised to find there’s no cabal-mode in MELPA. But coincidentally, someone started working on this literally three weeks ago! So I wrote a small expression to package this new cabal-mode:

    cabal-mode = pkgs.emacsPackages.trivialBuild {
  pname = "cabal-mode";
  version = "unstable";
  src = pkgs.fetchFromGitHub {
    owner = "webdevred";
    repo = "cabal-mode";
    rev = "083a777e09bdb5a8d8d69862d44f13078664091f";
    sha256 = "sha256-c5dUsnEx+0uXFzxQLMnhiP8Gvwedzvq0F0BA+beBkmI=";
  };
  packageRequires = [ ];
};

  

xcompose-mode

A few weeks back I switched from macOS to Linux, and since I’m stuck on X11 because of stumpwm, I’m using XCompose to define keybindings for entering dashes, smart quotes etc. It bothered me slightly that my .XCompose file didn’t have syntax highlighting. I found xcompose-mode.el in kragen’s xcompose repo, but it’s slightly broken (it’s missing a provide call at the end). I started thinking how hard it would be to write a Nix expression to modify the source after fetching, when I found that Thomas Voss hosts a patched version here. Which made this very simple:

    xcompose-mode = pkgs.emacsPackages.trivialBuild {
  pname = "xcompose-mode";
  version = "unstable";
  src = pkgs.fetchgit {
    url = "git://git.thomasvoss.com/xcompose-mode.git";
    rev = "aeb03f9144e39c882ca6c5c61b9ed1300a2a12ee";
    sha256 = "sha256-lPapwSJKG+noINmT1G5jNyUZs5VykMOSKJIbQxBWLEA=";
  };
  packageRequires = [ ];
};

  

eat

Somehow the version of eat in nixpkgs unstable was missing the configuration option to use a custom shell. Since I want to use nu instead of bash, I had to package this myself from the latest commit:

    eat = pkgs.emacsPackages.trivialBuild {
  pname = "eat";
  version = "unstable";
  src = pkgs.fetchgit {
    url = "https://codeberg.org/akib/emacs-eat.git";
    rev = "c8d54d649872bfe7b2b9f49ae5c2addbf12d3b99";
    sha256 = "sha256-9xG2rMlaMFY77JzUQ3JFrc7XKILZSL8TbP/BkzvBvMk=";
  };
  packageRequires = with pkgs.emacsPackages; [
    compat
  ];
};

  

lean4-mode

I started reading Functional Programming in Lean recently, and while there is a lean4-mode, it’s not packaged anywhere. This only required a slight deviation from the pattern: when I opened a .lean file I got an error about a missing JSON file, consulting the README for lean4-mode, it says:

If you use a source-based package-manager (e.g. package-vc.el, Straight or Elpaca), then make sure to list the "data" directory in your Lean4-Mode package recipe.

To do this I had to use melpaBuild rather than trivialBuild:

    lean4-mode = pkgs.emacsPackages.melpaBuild {
  pname = "lean4-mode";
  version = "1.1.2";
  src = pkgs.fetchFromGitHub {
    owner = "leanprover-community";
    repo = "lean4-mode";
    rev = "1388f9d1429e38a39ab913c6daae55f6ce799479";
    sha256 = "sha256-6XFcyqSTx1CwNWqQvIc25cuQMwh3YXnbgr5cDiOCxBk=";
  };
  packageRequires = with pkgs.emacsPackages; [
    dash
    lsp-mode
    magit-section
  ];
  files = ''("*.el" "data")'';
};

  

Linux on the Fujitsu Lifebook U729

Sat, 15 Nov 2025 00:00:00 +0000

This post describes my experience using Linux on the Fujitsu Lifebook U729. The tl;dr is that it’s a delightful laptop, and Linux runs flawlessly, and all the hardware things I’ve needed run OOTB. The only difficulty I had was in disabling Secure Boot, but I figured out how to do it, which I explain below.

Background
Troubleshooting
1. Secure Boot
2. Spyware
Non-Problems
BIOS Notes
Links

Background

From early 2024 my daily driver was an M2 MacBook Air, until earlier this year I broke the screen, and the repair was quoted at almost 1000 AUD. Since I used it as a desktop most of the time, this didn’t affect me much. After some flip-flopping I decided to get an M4 Mac mini. Partly for the faster CPU and more RAM, but partly because I liked the idea of LARPing like it’s the 2000s, when computers, and by extension the Internet, where fixed in physical space, rather than following everyone around.

Of course this was a terrible idea. I had three working computers—a Linux+Windows desktop, a Mac Mini, and a MacBook Air that I could use as a desktop—and none of them were portable. When I went to RustForge 2025 I just brought my phone. If I wanted to travel, even within Sydney, to a demo night or math club or some such, I didn’t have a laptop to bring with me.

So I needed a new laptop. And the Tahoe release of macOS was so ugly (see e.g. 1, 2, 3) it made me boot up the old Linux desktop, and start playing around with NixOS again. And I fell in love with Linux again: with the tinkering and the experimentation and the freedom it affords you.

So, I wanted a Linux laptop. I had a ThinkPad X1 some years ago and it was terribly: flimsy plastic build and hardware that vastly underperformed its price. I looked around for old, refursbished workstation laptops, and, randomly, I ran into an eBay seller offering a refurbished Fujitsu laptop.

The specs/price ratio was pretty good: 16 GiB of RAM and 512GiB of SSD, all for 250 AUD. And it was 12in and 1.1kg, which I like: laptops should be small and lightweight. But the thing that got me, in all honesty, was the brand. “Fujitsu laptop” sounds like colour in a William Gibson novel: “crawling into the avionics bay, Case took out a battered Fujitsu refurb, and stuck a JTAG port in the flight computer—”. I already use NixOS and a trackball and a mechanical keyboard, so a laptop that’s even more obscure than a ThinkPad is perfect for me. And it was only 250 AUD. So I got it.

The only problem I had was disabling Secure Boot in order to install Linux. Otherwise: I love it. It’s small and lightweight, feels solid, the keyboard is good, all the hardware works out of the box with NixOS, and the battery life is pretty good.

Troubleshooting

This section describes the problems I encountered.

Secure Boot

I tried to install Linux the usual way, when I was greeted by this:

Going into the BIOS, the option to disable Secure Boot was greyed out. I tried a bunch of random bullshit: wiping the TPM, disabling the TPM. That didn’t work.

What did work was this:

First, install Windows 11. This came with the laptop. And the installation makes installing Linux feel easy: I had to do so many weird tricks to avoid having to create an account with Microsoft during the installation.

Once Windows is installed, go into Windows Update. Under “Advanced Options > Optional Updates”, there should be an option to install Fujitsu-specific drivers. Install those. And for good measure, do a general Windows update.

There should be a program called DeskUpdate on the Desktop. This is the Fujitsu BIOS update tool. Run this and go through the instructions: this should update the BIOS (the ordering seems to be important: first update the Fujitsu firmware through Windows Update, then the BIOS through DeskUpdate).

Reboot and go into the BIOS (F2). You should have a new BIOS version. In my case, I went from BIOS 2.17 to 2.31 which was released on 2025-03-28:

You now have the option to disable Secure Boot:

After this, I was able to install NixOS from a live USB:

Spyware

The laptop comes with this corporate spyware thing called Absolute Persistence. It’s some anti-theft tracking device. Since the Lifebook is typically an enterprise laptop, it makes sense that it comes with this type of thing.

I only noticed this because I was searching the BIOS thoroughly for a way to disable Secure Boot. The good news is disabling it is pretty straightforward: you just disable it in the BIOS.

As I understand it, Absolute Persistence requires an agent running in the OS, so the BIOS support, by itself, doesn’t do anything once disabled.

Non-Problems

The following work flawlessly OOTB:

WiFi
Bluetooth
Sound (using PipeWire)
Display brightness control (using brightnessctl)
Touchscreen (I didn’t realize the screen was actually a touchscreen until I touched it by accident and saw the mouse move)
Webcam (not winning any awards on quality, but it works)

Things I have not tested:

Microphone
Fingerprint sensor

BIOS Notes

To enter the BIOS: smash F2 until you hear the beep. No need to hold down the Fn key.

To enter the boot menu: as above but with F12.

Agda on NixOS

Fri, 14 Nov 2025 00:00:00 +0000

To install Agda and its standard library, add this to your config:

    environment.systemPackages = with pkgs; [
  (agda.withPackages (p: [
      p.standard-library
  ]))
];

  

Or, using home-manager:

    home-manager.users.$username.home.packages = with pkgs; [
  (agda.withPackages (p: [
    p.standard-library
  ]))
];

  

The p here stands for the nixpkgs.agdaPackages package set. Note that the following will not work:

    environment.systemPackages = with pkgs; [
  agda
  agdaPackages.standard-library
];

  

If you use Emacs, you probably want agda2-mode, which if you use home-manager, can be installed using Nix:

    home-manager.users.eudoxia = {
  programs.emacs = {
    enable = true;
    extraPackages =
      epkgs: with epkgs; [
        agda2-mode
        # ...
      ];
  };
};

  

Now, if you have a file hello.agda with:

module hello where

open import Agda.Builtin.IO using (IO)
open import Agda.Builtin.Unit using (⊤)
open import Agda.Builtin.String using (String)

postulate putStrLn : String → IO ⊤
{-# FOREIGN GHC import qualified Data.Text as T #-}
{-# COMPILE GHC putStrLn = putStrLn . T.unpack #-}

main : IO ⊤
main = putStrLn "Hello world!"

Then:

    $ agda hello.agda # typecheck
Checking hello (/home/eudoxia/agda/t2/hello.agda).

$ agda -c hello.agda # compile
[snip]
[6 of 6] Linking /home/eudoxia/agda/t2/hello [Objects changed]
$ ./hello
Hello world!

  

So far, so good. Using the standard library however is more complicated. If you have a file vec.agda with:

module vec where

open import Data.Nat using (ℕ; zero; suc)

data Vec (A : Set) : ℕ → Set where
  []  : Vec A zero
  _∷_ : ∀ {n} (x : A) (xs : Vec A n) → Vec A (suc n)

infixr 5 _∷_

Then agda vec.agda will not work:

    $ agda vec.agda
Checking vec (/home/eudoxia/agda/t3/vec.agda).
/home/eudoxia/agda/t3/vec.agda:3.1-42: error: [FileNotFound]
Failed to find source of module Data.Nat in any of the following
locations:
  /home/eudoxia/agda/t3/Data/Nat.agda
  /home/eudoxia/agda/t3/Data/Nat.lagda
  /nix/store/[snip]/lib/prim/Data/Nat.agda
  /nix/store/[snip]/lib/prim/Data/Nat.lagda
when scope checking the declaration
  open import Data.Nat using (ℕ; zero; suc)

  

Instead, create a file vec.agda-lib in the same directory with:

    depend: standard-library
include: .

Now agda vec.agda will succeed.

Hashcards: A Plain-Text Spaced Repetition System

Sun, 05 Oct 2025 00:00:00 +0000

hashcards is a local-first spaced repetition app, along the lines of Anki or Mochi. Like Anki, it uses FSRS, the most advanced scheduling algorithm yet, to schedule reviews.

The thing that makes hashcards unique: it doesn’t use a database. Rather, your flashcard collection is just a directory of Markdown files, like so:

    Cards/
  Math.md
  Chemistry.md
  Astronomy.md
  ...

  

And each file, or “deck”, looks like this:

    Q: What is the role of synaptic vesicles?
A: They store neurotransmitters for release at the synaptic terminal.

Q: What is a neurite?
A: A projection from a neuron: either an axon or a dendrite.

C: Speech is [produced] in [Broca's] area.

C: Speech is [understood] in [Wernicke's] area.

You write flashcards more or less like you’d write ordinary notes, with lightweight markup to denote basic (question/answer) flashcards and cloze deletion flashcards. Then, to study, you run:

$ hashcards drill <path to the cards directory>

This opens a web interface on localhost:8000, where you can review the flashcards. Your performance and review history is stored in an SQLite database in the same directory as the cards. Cards are content-addressed, that is, identified by the hash of their text.

This central design decision yields many benefits: you can edit your flashcards with your editor of choice, store your flashcard collection in a Git repo, track its changes, share it on GitHub with others (as I have). You can use scripts to generate flashcards from some source of structured data (e.g. a CSV of English/French vocabulary pairs). You can query and manipulate your collection using standard Unix tools, or programmatically, without having to dig into the internals of some app’s database.

Why build a new spaced repetition app? Mostly because I was dissatisfied with both Anki and Mochi. But also, additionally, because my flashcards collection is very important to me, and having it exist either in some remote database, or as an opaque unusable data blob on my computer, doesn’t feel good. “Markdown files in a Git repo” gives me a level of ownership that other approaches lack.

The rest of this post explains my frustrations with Anki and Mochi, and how I landed on the design decisions for hashcards.

Anki

Anki was the first SR system I used. It’s open source, so it will be around forever; it has a million plugins; it was the first SR system to use FSRS for scheduling. It has really rich stats, which I think are mostly useless but are fun to look at. And the note types feature is really good: it lets you generate a large number of flashcards automatically from structured data.

The central problem with Anki is that the interface is really bad. This manifests in various ways.

First, it is ugly to look at, particularly the review screen. And this diminishes your enjoyment of what is already an often boring and frustrating process.

Second, doing simple things is hard. A nice feature of Mochi is that when you start the app you go right into review mode. You’re drilling flashcards before you even realize it. Anki doesn’t have a “study all cards due today”, rather, you have to manually go into a deck and click the “Study Now” button. So what I would do is put all my decks under a “Root” deck, and study that. But this is a hack.

And, third: card input uses WYSIWYG editing. So, you’re either jumping from the keyboard to the mouse (which increases latency, and makes flashcard creation more frustrating) or you have to remember all these keybindings to do basic things like “make this text a cloze deletion” or “make this TeX math”.

Finally, plugins are a double-edged sword. Because having the option to use them is nice, but the experience of actually using most plugins is bad. The whole setup feels janky, like a house of cards. Most of the time, if a feature is not built into the app itself, I would rather live without it than use a plugin.

Mochi

Mochi feels like it was built to address the main complaint about Anki: the interface. It is intuitive, good looking, shortcut-rich. No jank. Instead of WYSIWYG, card text is Markdown: this is delightful.

There’s a few problems. While Markdown is a very low-friction way to write flashcards, cloze deletions in Mochi are very verbose. In hashcards, you can write this:

Speech is [produced] in [Broca's] area.

The equivalent in Mochi is this:

Speech is {{1::produced}} in {{2::Broca's}} area.

This is a lot of typing. And you might object that it’s only a few characters longer. But when you’re studying from a textbook, or when you’re copying words from a vocabulary table, these small frictions add up. If writing flashcards is frustrating, you’ll write fewer of them: and that means less knowledge gained. Dually, a system that makes flashcard creation as frictionless as possible means more flashcards, and more knowledge.

Another problem is that Mochi doesn’t have an equivalent of Anki’s note types. For example: you can make a note type for chemical elements, with fields like atomic number, symbol, name, etc., and write templates to generate flashcards asking questions like:

What is the atomic number of [name]?
What element has atomic number [number]?
What is the symbol for [name]?
What element has symbol [symbol]?

And so on for other properties. This is good. Automation is good. Less work, more flashcards. Mochi doesn’t have this feature. It has templates, but these are not as powerful.

But the biggest problem with Mochi, I think, is the algorithm. Until very recently, when they added beta support for FSRS, the algorithm used by Mochi was even simpler than SM-2. It was based on multipliers: remembering a card multiplies its interval by a number >1, forgetting a card multiplies its interval by a number between 0 and 1.

The supposed rationale for this is simplicity: the user can reason about the algorithm more easily. But I think this is pointless. The whole point of an SR app is the software manages the schedule for you, and the user is completely unaware of how the scheduler works. The optimality is to have the most advanced possible scheduling algorithm (meaning the one that yields the most recall for the least review time) under the most intuitive interface possible, and the user just reaps the benefits.

Obviously without an RCT we can’t compare Mochi/SM-2/FSRS, but my subjective experience of it is that the algorithm works well for the short-term, and falters on the long-term. It’s very bad when you forget a mature card: if a card has an interval of sixty days, and you click forget, you don’t reset the interval to one day (which is good, because it helps you reconsolidate the lost knowledge). Rather, the interval is multiplied by the forget multiplier (by default: 0.5) down to thirty days. What’s the use? If I forgot something after sixty days, I surely won’t have better recall in thirty.

You can fix this by setting the forget multiplier to zero. But you have to know this is how it works, and, crucially: I don’t want to configure things! I don’t want “scheduler parameter finetuning” to be yet another skill I have to acquire: I want the scheduler to just work.

In general, I think spaced repetition algorithms are too optimistic. I’d rather see cards slightly more often, and spend more time reviewing things, than get stuck in “forgetting hell”. But developers have to worry that making the system too burdensome will hurt retention.

In Anki, it’s the interface that’s frustrating, but the algorithm works marvelously. In Mochi, the interface is delightful, but it’s the algorithm that’s frustrating. Because you can spend months and months drilling flashcards, building up your collection, but when the cards cross some invisible age threshold, you start to forget them, and the algorithm does not help you relearn things you have forgotten. Eventually I burned out on it and stopped doing my reviews, because I expected to forget everything eventually anyhow. And now they added support for FSRS, but by now I have 1700 cards overdue.

Additionally: Mochi has only two buttons, “Forgot” and “Remembered”. This is simpler for the user, yes, but most SR scheduling algorithms have more options for a reason: different degrees of recall adjust the card parameters by different magnitudes.

Hashcards

What do I want from a spaced repetition system?

The first thing is: card creation must be frictionless. I have learned that the biggest bottleneck in spaced repetition, for me, is not doing the reviews (I am very disciplined about this and have done SR reviews daily for months on end), it’s not even converting conceptual knowledge into flashcards, the biggest bottleneck is just entering cards into the system.

The surest way to shore up your knowledge of some concept or topic is to write more flashcards about it: asking the same question in different ways, in different directions, from different angles. More volume means you see the same information more often, asking in different ways prevents “memorizing the shape of the card”, and it acts as a kind of redundancy: there are multiple edges connecting that bit of knowledge to the rest of your mind.

And there have been many times where I have thought: I would make this more solid by writing another flashcard. But I opted not to because the marginal flashcard is too effortful.

If getting cards into the system involves a lot of friction, you write fewer cards. And there’s an opportunity cost: the card you don’t write is a concept you don’t learn. Integrated across time, it’s entire oceans of knowledge which are lost.

So: the system should make card entry effortless. This was the guiding principle behind the design of the hashcards text format. For example, cloze deletions use square brackets because in a US keyboard, square brackets can be typed without pressing shift (compare Mochi’s curly brace). And it’s one bracket, not two. Originally, the format was one line per card, with blank lines separating flashcards, and question-answer cards used slashes to separate the sides, like so:

    What is the atomic number of carbon? / 6

The atomic number of [carbon] is [6].

And this is strictly less friction. But it creates a problem for multi-line flashcards, which are common enough that they should not be a second-class citizen. Eventually, I settled on the current format:

    Q: What is the atomic number of carbon?
A: 6

C: The atomic number of [carbon] is [6].

Which is only slightly more typing, and has the benefit that you can easily visually identify where a card begins and ends, and what kind of card it is. I spent a lot of time arguing back and forth with Claude about what the optimal format should be.

Another source of friction is not creating the cards but editing them. The central problem is that your knowledge changes and improves over time. Often textbooks take this approach where Chapter 1 introduces one kind of ontology, and by Chapter 3 they tell you, “actually that was a lie, here’s the real ontology of this subject”, and then you have to go back and edit the old flashcards to match. Because otherwise you have one card asking, e.g., for the undergraduate definition of some concept, while another asks you for the graduate-level definition, creating ambiguity.

For this reason, when studying from a textbook, I create a deck for the textbook, with sub-decks for each chapter. That makes it easy to match the flashcards to their source material (to ensure they are aligned) and each chapter deck only has a few tens of cards usually, keeping them navigable.

Sometimes you wrote multiple cards for the same concept, so you have to update them all at once. Finding the related ones can be hard if the deck is large. In hashcards, a deck is just a Markdown file. The cards immediately above and below a card are usually semantically related. You just scroll up and down and make the edits in place.

But why plain-text files in a Git repo? Why not use the above format, but in a “normal” app with a database?

The vague idea of a spaced repetition system where flashcards are stored as plain-text files in a Git repo had been kicking around my cranium for a long time. I remember asking an Ankihead on IRC circa 2011 if such a thing existed. At some point I read Andy Matuschak’s note on his implementation of an SR system. In his system, the flashcards are colocated with prose notes. The notation is similar to mine: Q and A tags for question-answer cards, and {curly braces} for cloze deletions. And the cards are content-addressed: identified by their hash. Which is an obviously good idea. But his code is private and, besides, I feel that prose notes and flashcards are very different beasts, and I don’t need or want them to mix.

But I think the idea of plain-text spaced repetition got bumped up the priority queue because I spontaneously started using a workflow that was similar to my current hashcards workflow.

When studying from a textbook or a website, I’d write flashcards in a Markdown file. Usually, I used a shorthand like [foo] for cloze deletions. Then I’d use a Python script to transform the shorthand into the {{1::foo}} notation used by Mochi. And I’d edit the flashcards in the file, as my knowledge built up and my sense of what was relevant and important to remember improved. And then, when I was done with the chapter or document or whatever, only then, I would manually import the flashcards into Mochi.

And it struck me that the last step was kind of unnecessary. I was already writing my flashcards as lightly-annotated Markdown in plain-text files. I had already implemented FSRS out of curiosity. I was looking for a personal project to build during funemployment. So hashcards was by then a very neatly-shaped hole that I just needed to paint inside.

It turns out that using plain-text storage has many synergies:

You can edit the cards using whatever editor you use, build up a library of card-creating macros, and navigate the collection using the editor’s file browser.
You can query and update the collection using standard Unix tools, or a programming language, e.g. using wc to get the total number of words in the collection, or using awk to make a bulk-update to a set of cards.
You can use Git for version control. Git is infinitely more featureful than the change-tracking of any SR app: you can edit multiple cards in one commit, branch, merge, use pull requests, etc.
You can make your flashcards public on GitHub. I often wish people put more of themselves out there: their blog posts, their dotfiles, their study notes. And why not their flashcards? Even if they are not useful to someone else, there is something enjoyable about reading what someone else finds interesting, or enjoyable, or worth learning.
You can generate flashcards using scripts (e.g., turn a CSV of foreign language vocabulary into a deck of flashcards), and write a Makefile to tie the script, data source, and target together. I do this in my personal deck. Anki’s note types don’t have to be built into hashcards, rather, you can DIY it using some Python and make.

The result is a system where creating and editing flashcards is nearly frictionless, that uses an advanced spaced repetition scheduler, and which provides an elegant UI for drilling flashcards. I hope others will find it useful.

Adding Planets to Celestia on macOS

Tue, 29 Jul 2025 00:00:00 +0000

tl;dr: you have to modify the application bundle.

Celestia is a space simulator: you can fly around space and look at moons and exoplants, fast forward time. It is sometimes used by sci-fi artists for worldbuilding because you can easily add new stars/planets/megastructures/spacecraft. Some people have built whole virtual worlds for storytelling in Celestia. The Orion’s Arm collaborative worldbuilding project has a collection of Celestia addons so you can explore the world of the year 10,000 AT.

But the documentation is sparse and old. As with many things: the biggest hurdle to starting is just knowing which files go in which directories.

Celestia uses .ssc (solar system catalogue) files to define planets. These are plain-text files with a syntax resembling HCL. Let’s create baby’s first planet: below is a minimal .ssc file that adds a planet “Alpha” around the star Gliese 555:

    "Alpha" "Gliese 555"
{
  Texture   "asteroid.*"
  Mass      1             # Earth masses
  Radius    6378          # km
  EllipticalOrbit {
    Period          1.0 # years
    SemiMajorAxis   1.0 # long axis
    Eccentricity    0.0 # circular
  }
}

  

Now, what you would hope is that there exists a standard directory like ~/.celestia/planets/ you can put this into. I spent a lot of time looking through old docs and source code for this, and I’m writing this so others don’t have to. Unfortunately, at least on macOS, you have to modify the application bundle itself. This feels morally wrong, but it works.

Save the above code as alpha.ssc, and execute:

    $ mkdir -p /Applications/Celestia.app/Contents/Resources/CelestiaResources/extras/
$ cp alpha.ssc /Applications/Celestia.app/Contents/Resources/CelestiaResources/extras/alpha.ssc

Open Celestia, and navigate to Gliese 555 (press enter, type “Gliese 555”, press enter, press g). You should see a new planet:

Zooming in, you can see it’s using the built-in asteroid texture:

To verify it’s reading the right file, press tilde and use the arrow keys to scroll up the logs, and you should see a line like:

Loading solar system catalog: extras/alpha.ssc

Celestia traverses the extras directory recursively, so you can put your .ssc files inside folders to organize large worldbuilding projects.

Fernando Borretti

No-One Escapes the Permanent Underclass

How Workers Will Be Replaced

How The Rich Will Be Replaced

How The State Will Be Replaced

The Perpetual Zoo

On Human Autonomy

Alignment Won’t Prevent This

Conclusion

See Also

On Vulgar Materialism

Links

Human Routers of Machine Words

How To Read More

Human Bottlenecks

The Serious Context of Use

Internal Limiting Factors

The Applicability of Spaced Repetition

Notes on the Hantavirus Outbreak

Some Data Should Be Code

Footnotes

Letting Claude Play Text Adventures

The Trivial Harness

Memory

Aside: Small Worlds

Future Work

Code

There Is No New Aesthetics

1Password Dependency Breaks Syntax Highlighting

Using the Brother DS-640 Scanner on NixOS

Books I Enjoyed in 2025

Coarse is Better

Electron Contours

The Kowloon Walled City

The Dream Garden of the Poets

The Sack of Merv

Lady Lovelace

The Cosmic Microwave Background

Dream Story

I Wish People Were More Public

Your Writing

Your Books

Your Flashcards

Your Dotfiles

etc.

Ad-Hoc Emacs Packages with Nix

cabal-mode

xcompose-mode

eat

lean4-mode

Linux on the Fujitsu Lifebook U729

Contents

Background

Troubleshooting

Secure Boot

Spyware

Non-Problems

BIOS Notes

Links

Agda on NixOS

Hashcards: A Plain-Text Spaced Repetition System

Anki

Mochi

Hashcards

Adding Planets to Celestia on macOS