Monolith!

Deconstructing the undeconstructable basic building blocks of reality

Sometimes, you sense that beneath a layer of conceptual battles and confusions there is a concept that is so obvious you forget to talk about it or even name it. One such concept is that of the monolith. It is the idea lurking beneath questions of monopolies and monopsonies in economics, monolithic architectures versus modular architectures in engineering, and vertical vs. horizontal integration in business models.

But what exactly is a monolith? And why do people have such a love-hate relationship with monoliths, torn between trying to destroy existing monoliths and creating new ones?

As is often the case with me, the impulse to think about monoliths started out with a throwaway shitpost. In my case, I came up with the faux-political slogan attack and dethrone monoliths! (a reference to the attack and dethrone god radical left slogan of the 60s). Then it struck me that I really don’t believe particularly strongly in that particular cause, and I came up with an alternative, the care and feeding of monoliths. But I don’t feel strongly about that either. Nor do I care much about the third obvious potential agenda: creating monoliths.

All three programmatic agendas — of destruction, preservation, and creation — only interest me to the extent they shed light on what monoliths actually are. This background thought was simmering when I noticed the cover design of There is No Antimemetics Division, which appears to depict one of the SCP entities in the book, and also inspired by the monoliths in Arthur C. Clarke’s 2001: A Space Odyssey series.

Monoliths also have a strange appeal in art. There’s that Clarkesque Utah monolith showing up in random places:

In Indian art history, monoliths seem to have been particularly prized. Ashokan pillars have a more monolithic construction than comparable Greco-Roman pillars, and entire rock-cut temples and viharas are scattered over South Asia.

As engineered artifacts, monoliths present different challenges from assembled or modular ones. Sometimes its easier to build a monolith, sometimes its easier to build non-monolithic things.

Let’s get a few examples of modern monoliths on the table before continuing. I’ll state my definitions before justifying them.

1. A monopoly is a business that has a monolithic production organization at the core

2. A monopsony is an aggregated demand source that has a a monolithic consumption organization at the core (typically a government, but also huge distribution aggregators)

3. A monolithic architecture in engineering is one that is hard or impossible to decompose reversibly into constituent parts, especially independently reusable/scavengeable modular ones.

4. A material monolith is any artifact that is grown bottom-up from an oozy process such as crystal growth (jet turbine blades), casting (any type of molding process), or deposition (3d printing).

5. To justify my meme image: a public works project like a monorail, large dam, or space mission, is monolithic in a historical sense, as in a unique set of circumstances generating a non-repeatable (and likely non-ergodic) outcome that is very hard to generalize from but makes a good case study. Here it is the memory of the thing that is a monolith. It is a story that cannot be easily decomposed into constituent abstract story-pieces.

6. Human brains and modern ML models can be monoliths, by the definitions I’ll be setting up, depending on how they grow/develop.

The Monolithicity Test

A good but not perfect test for monolithic character is the value of minimally destructive maximally deep disassembly and reassembly. If you take it apart in a way that can be maximally put back together, how deep can the decomposition go, and how far can the reconstitution recreate the monolith? And how much is lost in the loop?

If the answer is that destruction or disassembly is basically a one-way street to worthlessness, you have a monolith.

If the answer is that you can losslessly get back to the original or superior form with minimal material or energy losses, you have a non-monolith.

Some examples of the test at work:

1. Industrial age monopolies weren’t actually very monolithic. When Standard Oil was disassembled into parts, basically nothing broke, and the aggregate value of the pieces was ironically higher than the original, making Rockefeller very rich. But putting Standard Oil back together again in its original form would have been fairly trivial too. AT&T was similar.

2. Modern businesses that are on trial for potentially being monopolies are much more monolithic, and far harder to take apart for reasons we’ll get to. It is not obvious to me how I would break up Google, Meta, or Amazon for instance, while retaining significant value.

3. A 3d-printed part can be melted down and reconstituted into filament, which can be used to reprint it. But there is a high energy cost to the loop, which is quite destructive and also materially lossy. And producing new filament is non-trivial.

4. The design of an Intel “monolithic” chip from the aughts or earlier is really hard to take apart into fragments that can be re-used in new designs. But newer SoC/IP (“system on a chip/intellectual property”) chips are designed to be highly reusable in precisely this sense. Intel is not very good at it, but other companies are. “Chiplets” are an extension in that direction. SoC/IP chips put together mostly reusable design elements (“IP” blocks) within a somewhat custom chassis.

5. A monolithic sculpture from antiquity is basically maximally monolithic. You can’t take it apart except by pulverizing it into rock dust, and reconstituting it (for eg. through some sort of sintering process using a 3d scan of the original) would be hard enough to be pointless.

6. If you took apart the story of a piece of public infrastructure (like a monorail!) that emerged in a historically unique way, using generic story patterns, you’d lose most of the story. The story is a monolith embedded in history. It is easier to “layer” new “infrastructure stories” on top, along the contours of existing infrastructure stories. This is also why you can’t “port” powerful stories, or replicate “Silicon Valley” elsewhere. That’s a narrative monolith.

7. Human brains and ML models are also maximally monoliths in that they can’t be put through this reconstitution loop at all, but the latter can be copied and are more monolithic. Human brains can become monolithic under some circumstances.

It is worth noting that “value” in this reconstitution loop test can vary. In disassembly/reassembly of economic monoliths, all you care about is something like GDP or gross profits. Nobody except perhaps founders and early employees is particularly attached to the ineffable nostalgic appeal of the entity. But at the other extreme, a monolithic sculpture pulverized and reconstituted would perhaps not retain its artistic merits and subjective value through the loop at all.

The case of the chip is particularly interesting. A finished chip, whether based on a monolithic or SoC/IP design, cannot be disassembled and reassembled at all. The physical thing is a pure monolith. All you can do is maybe pulverize it and recover some metal from the metal layers, which would be worth too little to bother. But then, the value of a chip lies almost entirely in the design (and in the design of the fabrication process), so it makes more sense to apply the monolithicity test to the design rather than the individual chip. There is too little material in the chips to worry about recycling anyway. The motherboards and plastic cases of devices matter a lot more.

But there are cases where the individual physical artifact is worth thinking about. For example, Apples airpods are monolithic. They have an integrated battery and electronics, and are headed straight for the landfill if they fail. Another example is composite blades of wind turbines. They’re basically impossible to recycle (except apparently into feedstock for coal-powered generation plants, which I find hilariously ironic).

What is a Monolith?

The monolithicity test doesn’t actually get us very far in characterizing what a monolith is. We have some sort of material intuitions based on sculptural monoliths and fictional artifacts like Clarke monoliths, but in what sense are Google or a modern chip like monoliths of those more obvious sorts?

Our intuitive idea of a monolith is a structural one, loosely characterized by a disassembly/assembly loop test (specifically, the amount of hysteresis or entropic irreversibility in such loops), but it’s easier to get at the essence of one, especially a modern one, by talking about its behavior.

A monolith is a system that is separated from its environment by a system boundary of some sort. It acts on the environment through output signals that cross the boundary, and responds to the environment by processing input signals that cross the boundary going the other way. These signals can be considered ongoing moves in a game between the system and the environment. In the simplest cases, we can think in game-theoretic terms.

In game theory there is the notion of a dominant strategy, where no matter what your opponent does, you have a winning counter move. There are various nuanced flavors of this. Depending on the flavor, you might be indifferent to your options (they all do the same thing), indifferent to what the opponent does (your move wins without having to be responsive or predictive in relation to theirs). If you enjoy any degree of dominance in a game, you enjoy two luxuries: being indifferent to what your opponent does and perhaps more subtly and importantly, being indifferent to what you do.

A particularly subtle flavor of this is where there is some notion of a margin of victory. Let’s say you have 2 options, and the opponent has 2 options, but all 4 combinations result in you winning. The only difference is how much you win by.

Let’s say you win by 1, 10, 100, or 1000 points. Do you care? Probably. Unless you have a very weird marginal utility curve, 1000 points is probably worth more than 1 point. Maybe not 1000x more, but enough more that you’d care (though there are weird cases where the marginal utility drops to zero after 1 unit).

Let’s say you win by 997, 998, 999, or 1000 points. Do you care now? Probably not. If the win margins are all high enough, then the differences among the margins are likely in your indifference band. Regardless of the semantics of the scoring system or the shape of the utility curve.

It’s obvious how to apply this idea to a textbook business monopoly. You put out shoddy products (you don’t care what you do), you have little to no customer service (you don’t care what the opponent does), and you enjoy huge rents to the point that you don’t care about minor differences in margins at all. Whether your margins are 69% or 70%, you don’t care. You can be a lazy billionaire either way, allowing a bloated bureaucratic organization to sprawl on your watch either way.

It is less obvious how to apply this idea to things like say chip architecture. One example is the “monopoly” (technically a duopoly) of the x86 architecture. When the entire rest of the industry is set up to fit around the x86 architecture, you really don’t care what makers of other supporting chips, compiler makers and OS makers think. You’re going to do what you like to the architecture and they’ll deal. The utter dominance of x86 was quite something until Arm and RISC-V began making inroads (there were Motorola and DEC chips in the 90s, but they never really threatened x86).

Here’s another one, this time from software. If you talk to AI people, they’ll all tell you they hate Nvidia’s Cuda (low-level software AI programs use to make GPUs do things) but they really have no choice but to use it (full disclosure: I consult for a company, TensTorrent that makes a competing product). This is because while making GPUs is challenging, making a software stack that makes GPU programming accessible to mortals is even more challenging. The layers between atomic matrix-multiplication instructions at the level of the chip, and PyTorch or similar frameworks used to build ML models, is a zone of Dark Magic that Nvidia owns in a dominant way. In part because of a long-term incumbency, and in part because it employs an army of engineers maintaining the Dark Art codebase. The closest thing to an actual shoggoth in ML is the Cuda layers, not the models themselves.

Game theory reading: Cuda is such a monolith, Nvidia doesn’t really have to care that you hate using it, or even worry too much about what kinds of GPUs it offers. It is a dominance advantage.

This relates to business models by the way. Modern monopolies and monopsonies are more monolithic because they are built on top of computing hardware and software, both of which can get radically more monolithic than industrial age machines. Making modular hardware is hard, though SoC/IP and chiplets are getting there slowly. Making modular software is far harder.

And finally, this is also why AI software, or “Software 2.0” as Andrej Karpathy calls it, is even more monolithic. Because the “software” as such isn’t really in the code (either at the Cuda level, or the few hundred lines of Python code). The intelligence is almost entirely in the trained model weights.

You can apply my reconstitutability test at the level of AI models. You can train AI models on the input/output behavior of traditional software 1.0 programs and produce illegible, inscrutable, monolithic equivalents. The reverse problem of capturing the input/output behavior of a model in the form of an “explainable” construct, ie a set of algorithms with legible control flows, is basically the problem of turning it into a modular software 1.0 program (which by definition is explainable). Though both problems are at early stages of research, my sense is that the latter is harder. Both are hard enough that effectively model-weight sets constitute monoliths. Except they’re monoliths that can be released open-source (but not scrutable-source — you can’t back out the training protocol from the weights) and copied.

Applying the dominance test for a monolith here leads to a surprising conclusion. AI models

1. Don’t care what form the input takes (they can take in multimodal, unstructured, noisy data) and

2. Don’t care about accuracy, provability, justifiability or explainability in their outputs.

This is a weird reading of AI models by the way, since the desiderata in 2 are typically viewed as “feature engineering” problems that are part of “productizing” models, not the output-indifference of a game-agent with a dominant strategy. But I think this is in fact the correct reading. Behaviorally, a powerful AI model is indistinguishable from a processing monopoly with an utterly dominant strategy.

Making ChatGPT care about accuracy is at least as hard as making the old AT&T care about customer service, or the old Standard Oil care about worker conditions in oil fields, and for rhyming if not identical reasons.

But this is somehow unsatisfactory. “Indifference” and “apathy” are very anthropocentric readings of the behavioral profiles of monoliths. Let’s translate to more impersonal terms.

Boundaries vs. Event Horizons

In the last section, I described a monolith as a system with a boundary across which I/O signals go. This is not quite right. Indifference and dominance effects mean that input signals can be safely ignored and output signals can be non-responsive, arbitrary, or dysfunctional.

The boundary is really an event horizon of sorts. The black hole information paradox rhymes strangely well with the behavior of monoliths of other sorts. As best as I am able to understand the black-hole information paradox, information exists only at the surface, and evaporates away through Hawking radiation as the black hole ages, with the black hole only retaining information about the mass, charge, and angular momentum of the collapsing body. A blackhole basically clears the “market” of mass in its neighborhood, throws away most of the information, and collapses into a solipsistic state defined by its genesis information environment, and decoupled from its live information environment.

So a monolith, in more general terms, is a system that is sufficiently massive inside its boundary that it has a tendency to collapse into itself, with the boundary being turned into an event horizon with information-decoupling properties. It can suck things in but you’ll have no idea what happens to those things. It can emit radiation that is in some illegible sense an information-preserving function of the input, but it effectively bounces off the surface and returns in an “encrypted” form. I’ll stop there since I’m out on a limb here at the limit of my physics analogy abilities, but you get the idea. A black hole is a pure physical counterpart to the perfectly non-responsive and oppressive bureaucracy.

We can now get away from anthropocentric terms like “dominance”, “indifference”, and “apathy.” Whether or not you attribute such attitudes to a blackhole like entity, what matters is that you have to interact with an event-horizon type boundary with the following properties:

1. You cannot peer through it at the interior

2. You can put things in but cannot expect to know what happens to them

3. Things come out but they don’t make sense and don’t seem responsive to your concerns

4. It evolves to the point where all it embodies is some minimal information about its historic initial conditions

Bureaucracies, traditional monopolies, modern platform monopolies, and finally, brains and AI-like systems all have these black-hole like features, and feature an “event horizon.” To the extent there is a meaningful idea of a singularity in the picture, it is much closer to the notion of a black-hole type physics singularity than any scrutable notion of “runaway self-improvement.” Relative to the last point, note that big, historic public-work projects like monorails often end up in states where all they embody is some information about the unique historical circumstances that gave rise to them.

“Runaway self-improvement” and notions of “goals” attributed to the entity behind the event horizon constitute classic anthropmorphic projection errors, no different from New Age sci-fi spiritualities (ironic or not) thinking of black holes as gods.

Understood this way, a monolith is a generalized black hole concept that is exemplified by many kinds of systems, which lend themselves to anthropomorphic projection to varying degrees: Rock sculptures, semiconductor chips, certain layers of software, the interiors of monopolistic businesses and government agencies, AIs, brains, and so on. Or for that matter, primitive experience of nature that has not yet yielded to any kind of reductionist science, so everything appears to be inscrutable and impenetrable divine agency.

From the outside, a monolith will always present an inscrutable, impenetrable, low-responsiveness, cryptic appearance that acts like an event horizon. It’s a function of behavioral dominance in relation to a local information environment that leads to “indifference” to inputs and “apathy” towards outputs.

What about the inside?

Into the Monolith

Supposedly, when you fall into a blackhole, you get really stretched out in space and your sense of time slows to a crawl. If you had the magical constitution to make it through alive, you’d probably be a very different person, thanks to the spatio-temporal distortions. “Falling” into a cult, past the event horizon of indoctrination, probably feels something like this.

But that’s not how you truly experience the inside of a monolith. Any monolith young enough and small enough to not kill things falling through is too young and small to convey the full experience. Cults are known for dramatically cutting off or controlling the interactions between members and the world beyond, but ultimately they are toy spaces. To truly experience monoliths, you have to turn to large corporations or government agencies.

You experience the inside of a full-blown monolith in one of two ways:

1. By already being inside the boundary before the collapse begins and it turns into an event horizon.

2. By being informationally “empty” enough (young enough) when you enter that the entry stress doesn’t kill you and you can continue development entirely inside.

The first experience is common with people who join startups that get big very early. The second experience is common with people who join old monolithic organizations very young and spend decades growing old within them (and also people raised in cults from childhood).

The most important part of the experience is solipsism. Your entire information environment is dominated by internal signals, a kind of blooming, buzzing confusion of total internal reflection. You hardly ever see meaningful information from the outside that is not heavily processed and distorted into internal frames. Your actions hardly ever emanate as signals visible to the outside world.

If the monolith is wrapped by an event horizon for outsiders, the outside world is wrapped by an equally impenetrable event horizon for insiders.

There is a strange duality here, but with a slight difference. The inside of a monolith is at least temporarily the winning side, in signal-based game theory terms. So where outsiders care about the event horizon and want to penetrate it, the insiders don’t have to care. If the interior environment is sufficiently rewarding and pleasant, who cares about the world you can’t see beyond the event horizon?

You’ll often get a taste of it when you talk to a long-term veteran of a company who has only ever known the world through that company, and has perhaps been let go in a middle-aged layoff, or has retired. Some symptoms:

1. They’ll talk about familiar universal things in odd, idiosyncratic ways derived from “insider” language

2. They’ll be used to very idiosyncratic tools, instruments, and infrastructure for getting things done (talk to an ex-Googler about the difficulty of adapting to non-Google commodity IT environments for example)

3. They’ll often be surprised to hear that ideas they thought were unique to their monolith’s interiors actually exist in multiple other, better forms outside.

4. They will often be surprisingly ignorant of basic things, but often “know” seemingly indistinguishable versions.

5. Their epistemology will be oddly monocultural, fitting a little too coherently together (even if inscrutably for you).

The human experience of being a monolith insider is of course trivially hackable and counter-programmable. Unless you’re shut away beyond a physical boundary without an internet connection, all you have to do is consume enough outside information. Information that is not from the funhouse mirror maze that is the interior of a monolith.

An example: when I was at Xerox (2006-11), we were handed a copy of the book Copies in Seconds, about the invention of xerography (which post-antitrust competitors call electrophotography — a notable case of insider/outsider language) and the founding of Xerox. I read that, then went ahead and read two books about the most famous chapter of Xerox history, PARC (Fumbling the Future, and Dealers of Lightning). When I tried to talk to my colleagues about all the fun historical stories, it turned out hardly any of them had read any of the books. They simply had some vague second hand notion of the history based on listening for 5 minutes in some orientation session. The average employee of Xerox had far less awareness of the fascinating history of Xerox than the average interested technologist outside.

This surprised me at the time, but it doesn’t anymore. Even when the difference between being captured by the interior information environment of a monolith and being connected to the outside world is as simple as picking up a book, most people don’t bother. It is simply easiest to just do your job and default to only using the trusted channels of input/output that are part of your internal environment.

The point of this discussion is to note that the interior of a monolith isn’t actually inscrutable. It’s merely highly solipsistic, shaped by almost entirely internal terms of reference and the lazy lack of curiosity that marks the psyche of a (temporary) winner.

Again, let’s make the move away from anthropocentric terms of reference. When you’re talking monoliths that are not inhabited by humans, you see a similar solipsism: idiosyncratic, non-standard designs, weird ways of referring to industry terms, idiosyncratic “physics” based on internal metaphors and mental models, and so on. The net result is that a monolithic artifact, if you manage to open it up, will look like it’s composed in an alien idiom to an outsider. Written not just in a different language, but with reference to a different reality. The inner and outer realities only converge in the most tenuous ways, many degrees removed from individual actions.

Different is not better. It’s often worse. Talk to a Googler. The average Googler, despite being an insider, typically has no better explanations for why search quality sucks so much, and often misses insights obvious to outside observers. The only real difference is they talk a different language of helplessness.

When you’re talking about monolithic products created by a monolithic organization, often you’ll see a mirrored inscrutability. The internal structure of the human organization will make no sense to outsiders, and the internal structure of the design will make no sense to people who attempt a teardown. This is the monolith version of Conway’s Law, which states that product structure mimics organizational structure. In a monolith, inscrutable organizations produce indecipherable products.

“Integration” Horizontal and Vertical

A sidebar on some related language. In business contexts, you’ll often hear of horizontal versus vertical integration. How do these ideas relate to monoliths?

If you’re not familiar with the terms, a “vertically” integrated company, like Intel, Apple or Tesla, puts complicated things together starting from scratch, with the most basic raw materials, through many layers of transformation and processing, to produce things that are used in black-box ways by customers. Vertically integrated companies tend to be monolithic and produce monolithic products.

“Horizontally integrated” is a bit of a misnomer since there is aggregation rather than integration. Horizontally integrated companies tend to aggregate demand for a particular kind of component or input needed by other businesses and get very good at producing just the one class of things. Due to the need to interface with many upstream and downstream “value-chain partners,” and the fact that they specialize in simpler subsystems or parts, such companies tend to be fundamentally less monolithic. But this is not always true, and need not stay true even if it starts out true. For example, both Intel and Microsoft started out as “horizontal” companies provisioning “horizontal” layers to IBM for PCs. But complexity quickly accrued, they acquired enough internal layers they began to behave more “vertically.” The overall market they were part of, the PC market, began to loosen and commoditize.

By contrast, Apple started out vertical and got more vertical over time. Now it makes its own chips and its own flavor of the Unix operating system.

In general, industries slowly cycle through horizontal and vertical organizational patterns as they evolve and mature, as different players gain and lose control of the “stack,” eventually turning into a set of monolithic islands in stormy oceans of commoditized complements. The commoditized complements often act as newer kinds of natural “raw” materials.

One monolith-friendly metaphor for understanding this is as an metal-working process. There are two ways to produce a metal monolith. The first way, if you can generate high enough temperatures, is simply to melt the metal and pour it into a mold. The second way is to work with pieces of metal that you can heat to softening, but not quite melt.

Metal can be cold-worked by hammering, which knots up inner layers into complicated tangles. It can also be hot-worked by heating and cooling through annealing and quenching cycles. Forge-welding combines these processes.

Such a process can be used to create a monolith from non-monolithic parts. The famous Iron Pillar of Delhi, for instance, appears to have been created this way. A bunch of hot lumps of iron hammered together into a monolith. The pillar was something of a mystery until the 1970s, since it is remarkably rust-free and the construction process was not understood. Crackpot conspiracy maven Eric Von Daniken thought it was of extraterrestrial origin, which is a nice illustration of the inscrutability of monoliths.

Mature organizational landscapes often come to resemble such an Iron Pillar. If you diagram the structure of real landscapes, facile mental models of horizontal/vertical start to fall apart. What you’re left is more monolithic parts of landscape shaping the evolution of less monolithic parts.

Anti-, Pre-, and Post-Monoliths

Monoliths are best considered ossified chunks of raw collapsed history. They exist not just in space (structural, behavioral, functional), but in time, as congealed memories of the indispensable information about initial historical conditions. They emerge during unique historical circumstances as sui generis islands of solipsisms, interacting over their lifetimes with their environment through an event horizon. A curious duality holds across this event horizon, where the “outside” can appear monolithic to the inner reality. To the extent there is an asymmetry, it has to do with temporary winning conditions (you can go wild with this and conclude there is no reality, only a topology of connected solipsisms).

To center monoliths in accounts of phenomena is to take a particularlist, historicist approach to analysis, rather than a general approach based on atemporal, abstract theories of structure, behavior, and function.

One way to do that is to treat the entire space of non-monolithic things via historical spatio-temporal relationships to monoliths around them. We can distinguish three kinds of non-monolithic entities:

1. Anti-monolithic entities are those that do not have enough mass within their boundaries to collapse into monoliths, but are likely in the region of influence of one or more monoliths.

2. Pre-monolithic entities are ones defined by system boundaries that are not event horizons, but contain enough mass that they could collapse into monoliths

3. Post-monolithic entities are ones defined by systems that emerge through the energetic destruction, evaporation, or collapse of monoliths.

Approached this way, many ontological questions appear in a very different light. Take for instance, tensions between centralization and decentralization, or between hierarchies and networks. All these are merely anti-monolithic structures that can exist when collapse into monoliths is not imminent. When these frames apply, it means historically non-unique conditions are prevailing, and generic theories might be useful.

Or consider why it’s harder for monolithic human organizations to benefit from generalist sources of knowledge, such as academic research or consultant knowledge. When the internal information environment is both monocultural and idiosyncratic, and non-interoperable with external epistemologies, there is no way for external knowledge to truly penetrate. Outsiders have to learn the internal language and become insiders first. If the process doesn’t kill them, it is likely to domesticate them so they forget or lose the information that made them worth bringing in in the first place. Even individuals entering in highly empowered ways, such as new external CEOs entering companies known for internal “lifer” career tracks, can be rendered entirely helpless in short order. In the show Yes, Prime Minister, this process is known as a new minister getting “house-trained” by the Whitehall bureaucracy.

The same logic explains why the “Not Invented Here” (NIH) syndrome exists in engineering design. It is remarkably hard for big monolithic companies to adopt obviously powerful innovations from competitors. Almost everything gets lost in translation. So they typically have to reinvent rather than imitate.

And finally, the same logic explains why it’s so hard to put AIs into a proper feedback loop with users, with local-contextual memory and a genuine understanding of the user’s unique needs. An LLM is effectively all the information on the internet collapsed into a highly idiosyncratic and solipsistic data monolith. Knowing and relating to it is as hard as knowing and relating to other human minds. For both sides.

Monolith Governance

But all this also suggests that monoliths, despite the analogy to black holes, are neither indestructible, nor incorrigible (I’ll reluctantly accept the term “corrigibility” from the AGI cult as a helpful one; I don’t want to turn into my own cultish NIH blackhole here).

Monoliths are neither intrinsically good, nor intrinsically bad. Nor are they omnipotent or omniscient. A monolith is merely a little bubble of solipsism trapped behind an event horizon of its own making, with a region of influence that is limited in space-time by their initial conditions (mass, charge, and angular momentum in some suitably generalized senses). It’s a non-factorizable “prime number” in the structure of reality.

The regions of influence can be huge, but are never infinite. The universe beyond the monolith is always bigger than the universe within the monolith. There is always more mass/charge/angular momentum outside than inside. Which means any victory enjoyed by the monolith is necessarily temporary.

What makes them difficult to work with is not just that you can’t understand them from the outside. It is that they can’t understand you either. The incomprehension is mutual.

Which means governance is rather uniquely constrained.

How do you Attack and Dethrone Monoliths? What is involved in the Care and Feeding of Monoliths? How do you create monoliths?

I will tackle these questions in a sequel.