A cabinet of curiosities for wandering minds.
Somewhere between the technology that fascinates us and the things that seem to have nothing to do with it — yet somehow have everything to do with it — this site will be a place to collect ideas, curiosities, and discoveries that deserve to exist somewhere.
From artificial intelligence to improbable observations of everyday life. From tools that reshape the way we think to stories that defy categorization.
A cabinet of curiosities, then. Organized around a single rule: if something is worth thinking about, it is worth talking about.
"Disorder, when it is insightful, is a form of higher order."
[ What's in the Cabinet ]
Tech observations — the things that don't fit in a changelog.
Improbable ideas — the things that don't fit anywhere else.
Specimens that resist classification — filed under curiosity.
Nothing polished. Everything considered.
On the Vesuvius Challenge, Herculaneum, and what it means when machines open books
In February 2025, Oxford researchers and the Vesuvius Challenge extracted readable text from a Herculaneum scroll — one buried in 79 AD and unreadable for two thousand years. The AI that found those words was never taught a single Greek letter. It was trained to distinguish dark pixels from light pixels. Just: here is ink.
It found διατροπή — "disgust" — in a text nobody has seen since the first century. The machine that cannot understand a single syllable opened the door. The humans will figure out what's behind it.
Documentation tools are having the same moment. Code that can't explain itself finds the dark pixels in your codebase and surfaces them. Someone has to read it.
The irony is not subtle, but it is useful.
On infinite feeds, the performance of reading, and what we lose when nothing ever ends
The average person with an internet connection has access to more unread articles at any given moment than they could finish in a lifetime. This is not a new observation. But the structure of the observation keeps changing. The original problem was scarcity — information was expensive and attention was cheap. The new problem is the inversion: information is free and attention is the only finite resource, except the architecture was not built to respect that finitude. The feed was built to make you feel like finitude was a personal failure.
Here is the specific mechanism. Every save button — bookmark, read-later, Pocket, Instapaper, "I'll come back to this," the browser with forty open tabs — is a small act of hoarding. Not in a bad way. In the way that squirrels hoard. The squirrel buries the acorn because the squirrel does not know which day will be the bad day. You save the article because you do not know which evening will be the quiet evening, and you want to be ready. But the quiet evening rarely comes, and the feed keeps producing new things, and every saved item you have not read becomes a small monument to attention you intended to spend but didn't. The algorithm was not built to help you read things. It was built to make you feel like reading things was possible. The possibility is the product.
The real question is not whether the content is good. The question is whether the architecture was designed for anything to ever end. If the answer is no — if the default state of every platform is that everything persists and nothing expires — then the feeling is not a bug. It's the intended outcome. The guilt is the mechanism. You feel bad about the unread pile, so you keep feeding the platform with your attention, and the platform keeps offering you new ways to defer the pain without solving it. The solution to "I have too many unread tabs" is not "I need a better system for my unread tabs." The solution is to accept that the unread tab was always going to be unread and that is fine. The architecture does not want you to know that.
We have built, collectively, the largest archive of content nobody is reading in the history of human civilization. The content exists. The consumption does not. We produce the archive as a form of performance — look at everything I have access to, look at everything I intend to engage with — and the performance is the point. We are not reading any of it. We are maintaining the possibility of reading it. The feed has made a category error so fundamental that nobody notices it anymore: confusing the existence of information with the act of knowing it. The information is there. The knowing is not happening. And the architecture is designed to keep that state exactly as it is.
Here's the uncomfortable extension. If the feed was designed to keep you in a state of potential consumption rather than actual consumption — if the scroll was optimized for time-on-surface and not completion — then every platform you use is calibrated against a metric that measures the wrong thing. The metric is attention captured, not attention spent. A piece of content that you saw and didn't finish is as valuable as a piece of content you saw and did finish, to the platform. To you, these are completely different. The feed was not built for you. It was built for the platform's model of you. The model benefits from both the reading and the not-reading. Only you benefit from the reading.
The exit is not a better system. The exit is accepting that some things will not be read and that the failure is in the architecture, not in you. This is genuinely hard. The architecture has made the failure feel like a personal deficiency. You have too many tabs open, you haven't been consistent, you need a better approach. None of that is true. You are in a system designed to produce the experience of never having read enough. The only way out is to stop playing the game that the system is built for. Read something. Close the tab. Let it be over. The feed will still be there tomorrow. It is always there tomorrow. That is the design.
On the Voynich Manuscript, 600 years of wrong answers, and the cost of designing for an audience of one
In 1912, Wilfrid Voynich bought a 240-page manuscript in an Italian estate sale. It was hand-stitched, vellum-bound, filled with botanical illustrations of plants that don't exist, astronomical diagrams that don't match any known chart, and an alphabet that nobody has ever linked to any other language on earth. He had purchased what may be the most unread book in history. Nobody has successfully decoded it. Not once. Not in six hundred years.
The text repeats. Words appear again and again in patterns that look like language — not random, not cipher, but something between. Scholars have tried Latin, Arabic, Hebrew, Kabbalah, phonetics, etymology, statistical analysis, and, in a move that was either very clever or very desperate, alien DNA. Nothing works. The writing is not gibberish. It's systematic. It's self-consistent. It was simply not designed for you to read it.
The best analogy is a letter written in perfect English, using a vocabulary that hasn't existed for six hundred years. Every word makes grammatical sense. No word means anything. You can map the grammar onto something — and find nothing.
The author of the Voynich Manuscript seems to have anticipated this. The text contains redundancies and structural patterns that a sufficiently patient reader might eventually decode — not fully, but enough. They built in an escape route for the one person who might need it. That's not failure. That's intimacy. A message for a single recipient, written in a code that assumes no future adversary will ever care enough to break it.
In 2023, a team at a university applied a new machine learning approach — not to decode the text, but to map it. Their conclusion: the manuscript has statistical properties consistent with a coherent language. Not random noise. Not a cipher. A language, unknown, somewhere, of which only one copy was ever written.
The discomfort is structural. You cannot read it because it was not written to be read. That is the design. Every decryption attempt fails because the manuscript was built to resist exactly what decryption does — an outside intelligence applying a methodology to extract meaning from an unknown system. It was sealed before it was written.
There's a version of this story that is about failure — six centuries of brilliant people, wrong about the same thing. But the truer version is about intent. Someone made a 240-page coded book, filled with images of impossible plants, for an audience of one. The cost was enormous. The effort was years. The message was designed to arrive intact and be understood by exactly one person and nobody else, ever. That person is probably dead.
The manuscript doesn't care. It was never yours to read.
On an extinct bird, the cost of having no enemies, and what we mean when we say something is 'natural'
The dodo was large — roughly three feet tall, forty pounds, built for an island with no predators. Mauritius had no mammals. The bird had evolved over millions of years in an environment where fear was unnecessary and running was pointless. It had no reason to run from anything. So when the Dutch sailors arrived in 1598 and started bashing them with clubs, the dodo walked toward the clubs.
Or so the story goes. The contemporary accounts are fragmentary, contradictory, and mostly written by people who had never seen a live one. What we know about dodo behavior comes from inference, archaeology, and the uncomfortable fact that we are reconstructing the psychology of an extinct bird from three sentences written by a sailor who was probably drunk when he wrote them.
The dodo went extinct in approximately sixty years. From first contact to last confirmed sighting: sixty years. It took longer to write this paragraph. The speed of the extinction is the part that should bother you. Not that it died — species go extinct all the time, the fossil record is mostly that — but how fast. Sixty years means the birds that were born when the Dutch first landed saw the species end before they died of old age. The last dodo stood on an island where every living thing it had ever known was gone, surrounded by species it had no framework to understand.
What does "natural" even mean, in that context? The dodo's behavior — unafraid, curious, approachable — was entirely natural. It was exactly what evolution had designed it to be. The problem was that evolution had not designed it for a world with sailors. The island had produced a creature perfectly adapted to a place that no longer existed in the way the dodo understood it. The environment changed. The behavior didn't. This is, more or less, how extinction works.
We find this funny because we are the sailors in this story, but we are also the dodo. Every comfortable assumption is a habitat. Every unexamined routine is an island where nothing dangerous has ever happened. The difference between the dodo and everyone else is not intelligence — the dodo was not stupid, it was simply optimized for a world that didn't include us. The difference is that we know other worlds exist. The dodo did not.
There is a version of this observation that is about adaptability and resilience. There is also a version that is about the way comfort erodes the behaviors that would matter if the comfort ended. The dodo didn't fail. It succeeded at being exactly what it was. That turned out not to be enough.
The last known dodo died sometime in the 1660s. The Dutch had stopped visiting Mauritius regularly by then. There was nothing left to take. The bird that had walked toward danger because it had never learned to fear anything was left alone with an island that no longer had anyone who remembered what it was for.
Nobody was there to watch it go.
[ Vol. I — Issue 01 : Nine Specimens ]
Nine specimens from the cabinet. Filed under curiosity. No further classification required.
The human brain consumes roughly 20 watts. A single LED bulb draws more. While GPT-4 runs inference, it consumes the equivalent of several hundred homes for a few seconds of thought. We built the most expensive cognition per calorie in the known universe and then decided to run it on GPU clusters.
The interesting question isn't whether AI is energy-intensive — it's what we're optimizing for. Human brains didn't evolve to compute efficiently. They evolved to survive in a world where a wrong answer once meant becoming someone else's lunch. Efficiency was never the priority.
There is a future where the machines get more efficient and the humans get faster at asking better questions. That future looks less like HAL 9000 and more like a very expensive assistant who still needs a human to tell it what to do.
Which, if you think about it, is exactly how a 20-watt brain prefers to work.
>> worth thinking about
A sourdough starter is a living colony of bacteria and wild yeast, each strain unique to a specific kitchen, a specific climate, a specific set of feeding habits. A starter from San Francisco is not the same organism as one from Oslo. The bread they produce reflects the terroir of the place it was made — much like wine.
The commercial yeast industry arrived and made bread consistent. This was the point. Sandwich shops need to know what time the dough will be ready. Sliced bread needs to fit a standard bag. Consistency is a feature.
The sourdough revival isn't really about flavor. It's about provenance. The question "where does this come from?" turned out to have a more interesting answer than we expected, once we started asking it about something as mundane as a loaf.
Everything is having its provenance moment. Coffee. Music. News. Code. We keep rediscovering that origin matters and that origin is usually local.
>> worth thinking about
LLMs confabulate. This is well-documented and widely mocked. Give a language model a question it doesn't know the answer to and it will produce an answer anyway, complete with citations that don't exist and logic that doesn't hold. The standard response is: fix the model. Make it admit ignorance.
But here is the counterintuitive observation: human expertise has always worked this way. The most confident experts in any field are frequently the most wrong — they've compressed so much experience into fluency that they stopped checking against reality. Dunning-Kruiter effect named this. But the mechanism isn't just confidence. It's the same compression that makes fluency useful.
A model that refuses to guess is a model that can't complete a thought. The hallucination is the cost of a system that can generate coherent responses to novel inputs. We've just never had a technology before where the failure mode was also the feature.
The real question isn't how to eliminate confabulation. It's how to build judgment around it. The humans who work with AI effectively have learned to treat every output as a first draft. That turns out to be a good rule for working with people too.
>> worth thinking about
The phrase "the map is not the territory" is one of the most repeated in management literature and one of the least applied. We build dashboards to understand complex systems and then optimize the dashboard rather than the system. The metric becomes the target and then the target becomes the thing.
GPS navigation created the most sophisticated map of road networks ever assembled and then immediately started routing everyone the same way, creating new traffic patterns that the map had not predicted. The map changed the territory. This is not a bug in GPS — it's a property of all sufficiently accurate maps. They change behavior. Behavior changes the thing being mapped.
The history of every large organization contains chapters where the KPI became the problem. Customer satisfaction scores make customers less satisfied by making them answer surveys. Click-through rates make content less clickable. Developer velocity makes developers less effective by making them measure velocity.
The solution isn't better metrics. It's understanding that metrics are descriptions of behavior, not objectives. They're useful in proportion to how carefully you remember what they're describing.
>> worth thinking about
Medieval cathedrals were not designed for sound. Organs, choirs, spoken word — none of it was what the stone was optimized for. The long reverb time (sometimes over 10 seconds) turned every syllable into a blur. The acoustics of a cathedral are the acoustics of a cave. They were designed for light.
Light came through windows. Specific windows, at specific times of day, through specific angles. The architecture was calibrated to make the light move. The prayers and music and speech that filled the space were tuned to work with that movement, not against it. You could not have a conversation in a cathedral. You could have an experience.
Modern offices were designed for conversation. Open floor plans, hard surfaces, optimal acoustics for speech. We put people in optimized speech environments and then wondered why they couldn't concentrate. The architecture was not wrong for speech. It was wrong for thinking.
There is no neutral room. Every space has an agenda. The question is whether that agenda aligns with what you need to do in it.
>> worth thinking about
Pigeons remember. Not vaguely — specifically. You dropped a crumb in a particular spot three years ago and they have not forgotten. The grudge is structural, not emotional. They navigate by magnetic fields and they remember your face. Two cognitive modes that should not coexist in something that eats crumbs off a sidewalk.
The city is full of creatures doing complex cognition without narrative — and doing it fine. Better, maybe. Their version doesn't collapse under the weight of the stories we tell about it.
There is an observation here about attention and memory and what we lose by turning everything into content. But the pigeons don't care about the observation. They just remember where you stood.
>> worth thinking about
We solved the documentation problem by stopping writing it. The screenshot became the answer to everything: "here, this is how it works." It captures the moment it was taken — already out of date the instant the UI changes, which is daily.
The documentation crisis in software is not a documentation problem. It's a knowledge problem. We forgot what we were supposed to know in the first place.
The unlikely thesis: the answer is not AI writing documentation. It's AI that makes documentation easier to write. Smaller units. Tighter constraints. A system that rewards readable prose over clever abstraction. The code was always the documentation. We just made it hard to read.
>> worth thinking about
A particle has no definite position until you look at it. This is not philosophy — it's the math. The wave function describes all possible positions simultaneously. Measurement collapses it to one.
But "measurement" here doesn't mean someone looking. It means any interaction with the environment — a photon bouncing off, a gas molecule brushing past. The world itself counts as a measuring device.
This means the universe was doing physics before there were physicists. The math is unambiguous. What it means about the nature of reality is a matter of ongoing, unresolved disagreement — which is a polite way of saying nobody has figured it out yet.
Some days it feels like the same is true of organizational structure.
>> worth thinking about
The cabinet is permanently open. New specimens arrive irregularly — when something earns a place.
"The world is full of obvious things which nobody ever observes."
— Sherlock Holmes