I haven’t written in a while, but here I am, repenting my sins with a post that uses all of ’s capabilities.
In my introductory post on hyperproblems, I ended with Common Source - open source plus people sharing - as the circulatory system for collective intelligence. I hadn't said what exactly was circulating, but if you look at the practice of science today, we have:
Papers, but not only papers, since most papers are frozen objects.
Data, but not only data, which by itself doesn't solve anything.
Models, but not only models, since (as Box reminded us) all models are wrong and only some are useful.
Experimental techniques and apparatus’ - the least standadized of the lot IMHO.
And we are missing out on systems that hold papers, data, models and other knowledge artifacts together. What’s the word for all of these knowledge artifacts and more?
The missing word was one I had tried to coin a decade ago, in an essay called Epistrons.
Naming the category changes what you can see and plan for, the way "infrastructure" changed how socities thought about roads and power grids. I had completely set aside my work on Epistrons until I started thinking about hyperproblems last year. Common Source circulates epistrons - knowledge artifacts, products, devices or systems built to satisfy our knowledge desires.
Click on the link below to learn more about Epistrons
The Missing Word
I ended my introductory post on hyperproblems with an invocation of Common Source: open source plus people sharing. The idea is that the architecture of collective intelligence for 21st century science can't live inside single institutions. It has to be distributed and fluid. Code, data and people have to move - and I am using ‘move’ in both literal and metaphorical ways - and the system has to be designed for that movement.
Common Source is an example of system design, a circulatory system for….? For what? What’s the artifact that gets circulated?
Not only papers. Some papers move, but most papers are frozen objects. Preprints are more mobile, but they are still not decomposable. Not only data, for data by itself doesn't solve anything; the question is what we build with it. Not only models, for models are limited (as Box famously said - “all models are wrong, some are useful”). Not only experimental setups, for they are context dependent and don’t travel well.
What are the artifacts that embody systems of knowledge such as Common Source?
The thing Common Source circulates has a shape. A simulator has a shape. A protocol has a shape. A queryable Large Planet Model has a shape. An IPCC report has a shape. A graduate seminar has a shape. These are all objects of inquiry, they differ from each other in interesting ways, and we don't have a word for the category.
Except that I tried to name it about a decade ago, in an essay called Epistrons: The Design of Knowledge. I didn't realize at the time what I had a word for. I thought I was writing about design and philosophy. What I was actually doing was naming the unit of the architecture of Hyperproblems:
Epistron: the most general term for a knowledge artifact - product, device or system that has been created to satisfy our knowledge desires.Common Source circulates epistrons.
The COVID response isn't the exception; it's a case of science building the right epistron (mRNA platform plus commercial distribution plus public-health integration) for the scale of the problem it faced. It worked because everyone, briefly, agreed to circulate epistrons instead of defending them.
A word matters because it organizes what you can see and plan for. Before "software" was a word, people built programs and didn't quite have a category for the practice. Before "institution" was a word in the modern social-science sense, there were organizations and laws and customs, but no clean way of discussing what they had in common. Before "infrastructure" was a word, there were roads and power grids and water systems, but no way to ask whether a civilization was investing enough in the thing-underneath-the-things.
“Epistron" is a word of that kind.
It names a category that we have always been building but never named. Naming has its advantages: you can ask whether the inventory is well-stocked, whether the mix is healthy, whether the design capacity is growing or shrinking.
The rest of this piece is about epistrons: what they are, why most of science produces only two or three species of them, what goes wrong when you match the wrong species to a problem, and what it would look like to train people to design them well. If the previous Hyperproblems essay was about the what and the why, this one is about the unit. The molecule. The thing itself.
The second section is a preliminary count of the species of Epistrons (genus instead of species?). I am not trying to be Mendeleev; this is a first pass. Still, several species come through clearly, pairing into object and systemic forms: physical devices and physical institutions, mental devices and mental systems, computational epistrons and media forms.
Modern science produces papers en masse because we settled on the peer-reviewed paper as the canonical unit and built incentives around that unit. But that’s a monoculture of knowledge artifacts at a time when we need diversity: we can’t address hyperproblems with only one kind of knowledge artifact attracting social capital.
My hope is that computational epistrons, the youngest species, will help dissolve boundaries between the others - papers becoming interactive, provers generating theorems, universities reorganizing into platforms.
Click on the link below to learn more about the Epistron Zoo
The Epistron Zoo
I am not trying to be a Mendeleev who tabulates the periodic table of epistrons - this is just a preliminary count.
Molecules have types, and so do epistrons. The knowledge artifacts people build to satisfy their knowledge desires fall into at least six broad species. The edges between the species are fuzzy, but the centers are distinct enough that naming them gives us a starting inventory of knowledge-making things.
Physical devices come first. Telescopes, microscopes, sequencers, mass spectrometers, oscilloscopes. Anything a laboratory buys or builds to extend the human senses. A physical device is an epistron with mass. You can ship it in a crate. You can break it. You can improve it in a workshop. Most scientific instruments of the last four centuries belong here, and most of them have been designed iteratively by the people who use them.
Physical institutions are the heaviest species of the lot. Schools, universities, laboratories, libraries, funding agencies, learned societies, archives. An institution is an epistron built at the scale of buildings and campuses, maintained across generations, and shaped by the rules of whatever society hosts it. Institutions are the slowest epistrons to build and the slowest to retire, which is why most of the ones we have were designed for problems we no longer have. Then again, conservatism might be a good thing - the oldest universities are centuries older than any nation state.
Mental devices have no mass. Concepts, axioms, theorems, principles, laws. A mental device has no moving parts, which is why it travels faster than any physical one. Bayes' theorem travels across the world without stopping at customs. A conservation law can get from a textbook in Göttingen to a classroom in Bangalore within a day. The best mental devices are astonishingly compact; a single equation can carry decades of insight in a form a student can hold in her head.
Mental systems are the most structural. Theories, constitutions, philosophies, notations, programming languages, mathematical symbol systems, musical notation, legal codes. A mental system lives as pure structure, but it is not a single mental device; it is a framework in which many devices can be composed. The difference between a theorem and a theory, or between a law and a legal code, is the difference between a device and a system. A programming language is a mental system. So is an accounting standard. So is the notation a physicist uses to write a Lagrangian. Mental systems take the longest to mature, and they have the longest half-life once they do.
You'll notice that epistron species come in pairs: there's the standalone 'object' form and the distributed 'systemic' form.
Computational epistrons are the youngest species. They too come in object and systemic forms: Simulators, models, for the first; AI systems, common-source platforms, queryable knowledge bases, etc for the second. These are executables. A simulator reenacts a phenomenon, which is a different kind of work from describing it. Computational epistrons barely existed as a recognizable species before about 1950, and through the 1990s they were a supplementary artifact to the other six. Today they are species in their own right, and also a solvent that is dissolving and reconstituting the others.
Media forms are the final species, and the inventory has widened over the last few decades. Papers, articles, journals, monographs, but also websites, dashboards, living ledgers, interactive notebooks, explorable explanations, and archives of every stripe (the systemic form). A media form is an epistron whose purpose is to hold and communicate other epistrons. The IPCC Assessment Report is a media form. A Jupyter notebook is a media form. A policy brief that updates when new data arrives is a media form. Much of the interesting design work in science communication happens here, usually outside the view of the disciplines whose work is getting repackaged.
That is the rough anatomy.
Modern science mostly produces papers, which account for the overwhelming bulk of the recognizable output. Theorems and concepts, which are mental devices, accumulate slowly and get compiled into papers, review articles and textbooks that are again media forms. Instruments get built because someone has to, but the people who build them are usually not the ones who get credit when the instrument enables a discovery. The other species - physical institutions, mental systems, computational epistrons - get produced badly or not at all, and the people who try to produce them do so against the grain of the career incentives that shape most scientific lives.
Over the twentieth century, science settled around the peer-reviewed paper as the canonical unit of work and built the rest of its machinery around that unit. Papers are trackable, distributable, archivable. Careers and Institutions can be constructed around them, funding agencies can allocate against them. Building a new journal, a new experimental protocol, a new piece of scientific software, a new research-community architecture is a career risk rather than a career move. A scientific culture that rewards one species ends up nearly illiterate in the rest.
That illiteracy is a catastrophe when it comes to addressing hyperproblems!
A modular experimental protocol that other labs can pick up, adapt, and improve is a media form designed for use rather than for reading. A shared simulation environment under permissive licenses is a computational epistron and a mental system at the same time. Common Source itself is an institutional form; it is a mental system taking a physical shape and circulating as infrastructure rather than being issued as a decree. These objects cannot be peer-reviewed the way a paper can be peer-reviewed. They cannot be counted, cited, or ranked by the apparatus we have. Evaluating them properly requires a different institutional metabolism, and for the most part that metabolism does not yet exist.
The hope is that computational epistrons will help dissolve some of the boundaries between the others. Papers are turning into interactive documents. Automated provers are starting to generate and verify theorems. Executable domain-specific languages are displacing hand-written notations. Even laboratory instruments are increasingly the software wrapped around the hardware rather than the hardware itself. Universities are being reorganized into platforms. The computational species is not only adding objects to the inventory; it is restructuring the others from inside. Some of the most consequential epistrons of the next decade will be hybrids that the older taxonomy could not have anticipated - an institution that is partly a queryable model, a mental system that is partly executable, a media form that is partly a simulator.
Don't get me wrong: universities have lasted longer than empires, and we should be very careful about replacing a system that's worked for such a long time; but caution and innovation can go hand in hand...
The third section is about level-mismatch. Epistrons come in rough sizes that correspond to the scale of the problem they address. A thermometer is small; a periodic table larger; a university larger still; a constitution larger again. The levels are not linear. A university is several orders of magnitude removed from a thermometer in design complexity, institutional weight, and maintenance burden. Mismatch shows up most sharply at the interface between an epistron and the world. My canonical case is the IPCC: planetary-scale content shipped as an Assessment Report that a farmer in Indonesia cannot query, a municipal planner in Chennai cannot parameterize, a trader in Tokyo cannot stress-test. What we need is something closer to a Large Planet Model with a multimodal UI. Mismatches run the other way too - a university's ethics apparatus can smother a routine experiment - and in the most dangerous direction we aim a tiny epistron at a civilizational problem, as when PageRank made a single algorithm the custodian of human knowledge. The mismatch persists because small epistrons are legible in ways large ones are not: countable, fundable, career-bearing. Can we do better at matching? Yes. But our incentive systems still pay for smallness.
Level Mismatch
Knowledge is an artisanal product that has mostly been transmitted within traditions. Institutions that make knowledge transmission easy - universities, for example - have lasted for centuries because (traditionally if not now) there's alignment between the various epistrons that go into the making of a university - the habits of mind, teaching rituals, libraries etc - and the purpose of the university, which is to pass robust, valuable knowledge from one generation to the next.
The institution and the problem are well-fitted.
Epistrons come in levels - rough sizes that correspond to the scale of the problem they are built to address. A thermometer is a small epistron, fit for a small well-defined measurement. A periodic table is larger; it organizes a whole branch of inquiry. A university is larger still; it binds research, instruction, credentialing, and community across generations. A national constitution is larger again; it negotiates the terms on which a few hundred million people live together.
Sometimes there's mismatch between the problem and the institutions we have on hand. Planetary level governance, say, of climate change, is a good example - the scale of the problem is larger than any current institution can handle. An epistron designed for a problem at one scale cannot be applied to a problem at a different scale without something breaking.
The levels are not linear. A university is not an additive multiple of a thermometer; it is several orders of magnitude of design complexity, institutional weight, and maintenance burden. You cannot get from one to the other by doing a better job on the smaller object. A better thermometer is still a thermometer; a university is a categorically different object. Mismatches can also happen at the interface between an epistron and the world - for example, the scientific content the IPCC summarizes is planetary in scale, but it ships to the world in the form of an Assessment Report: a long-form technical document, that's inaccessible to most. A farmer in Indonesia cannot query it. A municipal planner in Chennai cannot parameterize it. A trader in Tokyo cannot stress-test it. The right interface is more like a queryable, locally adaptable, continuously updatable computational epistron that plugs into decision-making. Large Planet Model with a multimodal UI is a lot more like what we need.
The opposite direction happens too. A large epistron pointed at a small task can smother it. A university's ethics apparatus is built to adjudicate questions with generational reach, and it is invaluable when those questions arrive; the same apparatus, brought to bear on a routine experiment with no novel ethical content, produces a hundred-page approval for work that should have taken a conversation (or outright denial). Sometimes, we aim a tiny epistron to civilizational problems. PageRank turned Google into the custodian of all human knowledge, a single algorithm with generational consequences.
Why does the mismatch persist? Partly because small epistrons are legible in ways that large ones are not. A peer-reviewed paper, a PhD thesis, a lab report - each has well-understood quality criteria, evaluation rituals, and career consequences. You can count them, rank them, fund them. A planetary-scale epistron has no peer-review equivalent yet. There is no established way to evaluate a Large Planet Model the way you would evaluate a paper. There is no tenure committee that knows what to do with a well-designed knowledge commons. So we keep producing small epistrons, even though they are inadequate to their tasks.
Can we get better at matching epistrons to scale?
Yes. But. We mostly don't do it because our incentive systems reward the production of small-scale epistrons regardless of whether a large-scale one is what the problem actually requires.
The rest of the essay sketches what large-scale epistrons might look like, what I think an epistron designer should be trained to do, and a proposal for a common-source catalog of epistrons.
this essay is a work in progress, and I will keep adding pages over time