In this installment, I want to provide two quick examples of how this might work at the level of individual ecologies, and sketch out how many more such ecologies might form a primordial computing soup.

AI Art Scenius with Titles

The first example is TITLES (who also have a Substack called ), the generative art platform that we use to produce the artwork for Protocolized. The brain part of Titles is a pipeline to make fine-tuned models from art collections by a particular artist. The high-personality part is that each model reflects a distinct individual artist’s style for that project.

The fabric part is a rather clever “creator studio” for composing these individual models together, to create an ecology based on “sampling” multiple models (in the sense of sampling in music) to create new artwork. The fabric accomplishes two things – combining multiple models together in a mathematically meaningful way, and keeping track of the contributions to allow for attribution and profit-sharing. The overall ecology also has a personality, similar to how music scenes can have personalities.

An Open Planetary Network of Robots

The second example is more complex, and one I’m involved in personally – the Yak Robotics Garage (YaRG) project.

The goal of this project is to create a planet-wide network of open-source rovers and other robots (such as drones), as a stepping stone towards rover networks on the moon and Mars. The idea started with (a Protocol School alum) figuring out how to teleoperate robots securely, in exchange for blockchain payments, and then generalizing the mechanism to use the ERC 8004 protocol (a sort of onchain directory and rating service for AI agents) to drive discovery of available robots for tasking.

Summer of Protocols researcher then joined in the fun and prototyped an auction marketplace to allow for posting of jobs for robots, and bidding by robots able to do them. There is currently a demo marketplace going (with dummy data, and a mix of real and virtual rovers, but real prototype protocol plumbing behind it) and plans underway to test the technology in the construction sector.

Where does AI fit in here?

Well, the problem with operating an open network of rovers in the real world is that there can be a dizzying variety of hardware types with different capabilities, owned by a large variety of actors of different levels of trustworthiness, situated in different environments. There can be all sorts of potential operators anywhere on the planet – or even on an entirely different planet – with varied skill levels.

Rather than brittle and specialized command modes, you want high-intelligence robots of all sorts to expose their capabilities to potential users/customers via a flexible command surface, and high-intelligence clients commanding them using LLMs that can understand their varied technical capabilities and map them to the needs of particular tasks and missions.

So you use MCP (Model Context Protocol) to expose the capabilities, ERC 8004 (try searching for “robot”) to discover the capabilities, LLM agents to use those capabilities to get tasks done, and either traditional or blockchain rails, using the x402 protocol, to organize a marketplace for robotic services to be provisioned and procured for money.

Those are just the main moving parts in a rather complex scheme – but one in which all the complexity is mainly dealt with by AIs rather than humans. Here is an explainer video (AI generated) of the technical infrastructure behind the scheme:

Here is a simple demo video of the basic protocol in action with a real robot. And here’s another video with Anuraj and Rafa demonstrating the auction marketplace in action.

It might not seem like much compared to the spectacular robot demonstration videos you find all over social media these days, but the point is not the robots themselves, or what they do, but that it is all being orchestrated over the open internet, using mechanisms that can potentially scale planet-wide without being owned or controlled by any single entity, such as a powerful corporation or state.

In this example, the brain is distributed across multiple rovers and the LLMs that can control them. The fabric is a stack of different protocols handling various coordination needs, ranging from discovery and verification of capabilities in a variable-trust market environment, to actually enabling the teleoperation connection, to handling the auditing of results and completing any financial transactions as agreed upon. All in high-speed automated ways that still allow for case-by-case judgment and decision-making by AIs supervised by humans.

It is worth comparing this vision to a competing vision: The kind promoted by vertically integrated robotics companies through jazzy demos featuring robots doing impressive acrobatics in controlled environments. These visions typically rely on highly integrated and closed products, even if they sometimes offer lip-service to open-source affordances for some parts of the whole picture. These are comparable to early proprietary computing networks, or contemporary social media platforms owned by large corporations.

An open robotics marketplace, on the other hand, would be more like the open internet – anyone with a robot of any sort (from small hobby rover in someone’s basement to a billion dollar rover on Mars) could potentially join, and connect with anyone else with a need for that particular robot’s capabilities and the ability to pay for it. It would be messy, janky, and glued-together. It would form a kind of tangled bank of artificial organisms competing for survival in an atomized market-like environment.

Which world would you rather live in? Yet another world of monopolistic platforms, or a cheerful anarchy of robots and their owners wheeling and dealing in an open economy?

Subscribe now

The Primordial Soup

These are just two examples of how protocols and AI can be put together in creative ways. There are dozens of others being experimented with right now, ranging from the viral and highly visible OpenClaw ecosystem to obscure and specialized ones that are as yet only crazy ideas in the heads of teenaged hackers.

Over the next decade, we’ll probably seen tens of thousands of such brain-and-fabric ecologies take shape independently. They will likely fall into loosely similar families of patterns. Some may converge, others may diverge, just like biological ecosystems.

If you think that’s a fun vision, imagine what could happen once these ecologies begin to run into each other and interconnect. Thanks to AIs, protocol systems that would have been non-interoperable in older technology paradigms will be able to automatically figure out how to talk to each other, forming squishy, oozy interfaces with each other, cobbled together by AI agents feeling each other out and inventing pidgins as they go. When AI is cheap enough, and the basic fabric capable enough, inventing a language even for just two entities to talk to each other for one short interaction becomes possible.

Take even the two examples in this essay. We can imagine photography robots in different parts of the world in the Yak network submitting photos to Titles to train individual models based on their particular image-making capabilities (such as different types of camera). Users could then sample those models to synthesize composite images by sampling those models to create strange new images seen by wholly synthetic robotic eyes.

Imagine that sort of thing, but in a primordial soup of thousands of ecologies.

As this process unfolds over the years, and the primordial soup boils and bubbles, the planetary computational character will begin to emerge in the form of a planet-scale emergent distributed brain, integrated and orchestrated by an emergent world fabric.

In both the Hitchhiker’s Guide universe and the Culture, machine personalities are narratively load-bearing rather than cosmetic features or shallow plot devices to make the non-human characters superficially “interesting.” The personalities shape the plots in material and non-human ways.

One fun example is the Nutrimatic drink machine in HHG, which claims to produce personalized drinks, but always produces the same liquid that tastes “almost, but not quite, entirely unlike tea” (which strikes me as an embodied behavioral cousin of some of the lazier hallucinatory and averaged-out responses of modern AIs). When Arthur Dent forces it to work harder to actually produce tea, it draws so much computing power away from the ship’s navigation, that the ship crashes.

In the real world, AI personalities are turning out to be just as consequential, though it’s not as funny when actual human lives are at stake.

The Missing Mechanisms Problem

In this essay, I want to argue that AI personalities are central to solving a problem Tim O’Reilly posed in a recent blog post: articulating agent ecologies with the right mechanisms.

Right now, there’s a problem that makes the AI/human knowledge market less efficient than it could be. The disrespect for IP that has been shown by the AI labs and applications during the training stage, and even now during inference, has led to efforts by content owners to protect their content from AI. Do not crawl. Lawsuits. Reluctance to share information. Even the AI labs are complaining about the theft of their IP and trying to protect their model weights from distillation.

It’s an economy crying out for mechanism design.

I want to address a slightly generalized version of Tim’s question, and think about ecologies rather than economies, drawing inspiration from one of our favorite essays here at Protocolized, Frank Chimero’s Only Openings, which argues that effective ecological stewardship relies on mechanism design that aims to manage problems indefinitely, rather than solve them once and for all. In Chimero’s essay, the specific personalities of the species involved in the case studies he talks about – bears, wolves, humans – materially shapes the mechanisms that help manage their interactions indefinitely and effectively.

How do we apply this idea to AI agent ecologies?

Modern real AIs already exhibit clear personalities, a mix of “genuine people personalities” inherited from their training data and protocols, and non-human dispositional aspects that are the result of model architectures and their underlying mathematics (transformer and diffusion models have different personalities for example). The current version of ChatGPT strikes me as an overconfident and slightly patronizing consultant, while Claude strikes me as an over-solicitous personality with some false humility (vaguely Uriah Heep-ish) going on. The human-legible and entity-anchored aspects of personality are merely the tip of the iceberg.

As with humans, it turns out that the personalities of AIs are intersubjective and situated. They are functions of how coherent entities disclose themselves and relate to each other, in the context of the things they do in collaboration. The personality of an AI or robot is a function of the stable gestalt disposition it presents as an interface to all other entities it might relate to. This disposition helps set expectations for counterparties in relationships. If you met an AI that called itself No More Mr. Nice Guy, would that shape how you interacted with it?

This point is not restricted to AIs, robots, smart homes, and other “intelligent” technological entities. Any sufficiently complex technological entity with any degree of autonomy of operations must present a stable disposition that can be deciphered and relied on by entities that interact with it.

For example, on the Ethereum blockchain, Layer 2 networks providing rollup services (bundling transactions into batches to submit to the Layer 1) can be “optimistic.” Here “optimistic” is both a term of art in the engineering, and a human-like attitude that embodies a pattern of expectations. Or to take an older technology, road traffic systems in well-developed urban regions tend to present a deferential attitude to pedestrians, while suburban ones tend to present a hostile attitude.

For a complex technology, it is useful to imagine an underlying “personality” with an intelligible point of view generating the visible disposition (regardless of where you land on the philosophy of mind question of whether there is “something it is like to be” an AI or robot). The interaction surfaces of simpler technologies can be mentally modeled as relatively unchanging “user experiences.” But with complex technologies, it is useful to model those surfaces as the fluid response surfaces of stable non-anthropomorphic personalities; ghosts inhabiting machines.

Perhaps the term Haunting Experience, or HX, should replace UX, for sufficiently complex technologies. AI certainly qualifies.

An AI presenting an intelligible HX is not quite as on-the-nose a feature as an AI being “explainable” (a rather ridiculous legalistic requirement to impose on a technology in my opinion; how many human beings, groups, or institutions are “explainable” after all?), but it does render complex technologies as somewhat predictable gray boxes rather than entirely inscrutable and unpredictable black boxes. It does not make them explainable, but it does make them narratable. It makes them composable.

What does this buy us? It buys us the ability to assemble such technologies into larger ecologies. This is where the real power of thinking in terms of HX becomes evident, when you are shaping the behavior of entire ecologies, rather than single agents.

Haunting Experience (HX) Design

We typically translate the personalities of simpler technologies to human-centric UX measures like “latency” or “walkability,” but with complex technologies, it is useful to reframe the problem in terms of designing the personalities of ghosts in machines (both plural, since we are considering entire ecologies), and how they should haunt us.

So how do we encourage the right ghosts to emerge?

The personalities of technologies are the result of two entangled forces acting together – human (and increasingly AI) design, and emergence. This is similar to the design of market mechanisms by human policy-makers in institutions (such as central bankers and elected representatives), interacting with the emergence effects studied by economists, to generate the economy we actually inhabit. It is neither an inscrutable black box, nor completely determinate. It is just intelligible enough to inhabit – it is no accident that Adam Smith used the ghostly metaphor of an “invisible hand” for describing the mechanisms of an economy.

We might use the term HX design for this sort of thing – conjuring ghosts within machines that exhibit particular desired personalities. The term is inspired by the output of a distributed AI workshop we ran last year (and derived from somewhat related usage of the term hauntology by philosophers such as Derrida and Mark Fisher).

You might reasonably suspect that HX design primarily has to do with AI and robots, but this would be a mistake (a typically anthropocentric one). Technologies that invite anthropomorphic projection (or possession perhaps) aren’t the only ones that induce partially designed emergent ghostly personalities within themselves.

Engineering is full of such conjured personalities. “Greedy” algorithms take the first good option they find. “Optimizing” algorithms look for the best option in some sense. “Satisficing” algorithms solve for “good-enough.” “Least commitment” approaches delay decisions as long as possible. “Eager” algorithms are proactive about whatever they do.

Subscribe now

High-Personality Ecologies

In every such case, there is a cost to the “personality” deployed for problem solving; one that must often be paid for by counterparties in transactions. If your automated decision-making is “optimistic,” then a counterparty system that monitors and audits its decisions must be “pessimistic” to make up for it. The calculus of benefits and costs to others associated with an agent’s behaviors, to a first approximation, is that agent’s personality.

The personalities of technologies, in other words, are intelligibility mechanisms for predictably distributing the computational cost of autonomous decision-streams among interacting entities (including both humans and autonomous machines).

The upside of such high-personality ecologies, with a lot of variation and diversity in the agents and interactions constituting them, is that they are vastly more generative than either monocultures based on low-personality fungible elements, or low intelligibility opaque elements. High-personality ecologies are like relatively free markets, while low-personality ones are like command economies, and opaque ones like the internal managerial economies of closed organizations.

The characteristics of high-personality technology ecologies is particularly clear in the field of operations research (OR), which deals in problems that are almost always NP-hard (i.e. computationally intractable), and must therefore be solved with heuristics that are only effective locally. OR is full of scheduling and planning algorithms that are defined by their personalities, which create consequences that must be dealt with by counterparties. For example, a simple and popular algorithm for prioritizing tasks in a queue, Shortest Processing Time (SPT) minimizes the average wait time for waiting tasks. But in a situation where tasks arrive constantly, it might delay longer tasks indefinitely. Producers of long tasks must negotiate appropriate service-level expectations that incentivize deviations from pure SPT behaviors.

An ecology comprising even simple processing agents with different “scheduling heuristic” personalities, and customers that bring various mixes of tasks for processing, is going to have a particular emergent personality, a particular style in which it gets things done. One that can be shaped and made intelligible and narratable to a useful extent by design. This is what it means for an entire ecology to have a personality. As we learned during Covid, a supply chain being lean or fat is a personality label that indicates how it behaves in real conditions, not a gratuitous obesity descriptor.

I will offer a stronger claim: only high-personality ecologies, ones with unique but mutually intelligible entities, can be economically generative. This is why AIs with personalities, composed into ecologies with personalities, are required to solve the problem of missing mechanisms.

To borrow a phrase from the title of a book by Ben Horowitz, what you do is who you are. And what you do typically involves relationships with others, whether the agent in question is a simple scheduling algorithm or an LLM.

The Protocol is the Personality

As Marshall McLuhan famously observed, every medium (by which he meant any technology, not just communications media) has a message. This is true of all technologies, whether simple or complex. A hammer has a message, as does a television. But sufficiently complex and autonomous technologies take the phenomenon to another level. Characteristic patterns of behavior (the rich “message”) reveal a general personality.

Here it is useful to characterize “sufficiently complex and autonomous.” Roughly speaking, a Turing-equivalent technology (i.e., equivalent to a general-purpose computer) that makes some significant class of decisions autonomously, based on engineered decision architectures rather than natural properties, is the kind of thing I am talking about.

This personality is best revealed in the context of interactions with other entities that must exhibit complementary personalities in order to form stable ecologies. An ecology of personalities with a particular distribution, woven together with particular protocols, has its own emergent distributed personality, just as human aggregates from subcultures to nations have their own personalities. Or, for that matter, pre-AI technological ecosystems such as the Microsoft or Salesforce ecosystems. And applying the same principle, what these ecologies do is who they are.

One way to frame this is: the protocol is the personality.

The behavior of an internet-connected computer isn’t entirely a function of its own architecture. Much of it is derived from the personality of internet protocols. Mac vs. PC or iOS vs. Android might be the atomic individual personality distinctions, but by what you do is who you are logic, to the extent both pairs are situated in the internet, both inherit the personality of the protocols of the internet.

The transition from the relatively atomized PC era to the connected and social (for both humans and machines) internet era took about a decade, but as with everything else, AI seems to be speed-running this phase transition. It is already becoming clear that the personality of different AIs is only partly an innate property of specific language or image models, traceable to their training data. The full personality of an AI is revealed when it becomes socially embedded in an ecology of other AIs and humans, and must deal with the consequences of its own dispositions on others.

The personalities of complex technologies are only fully expressed in the right ecologies. Protocols can be understood as precisely the engineered ecological scaffoldings that draw out full expressions of personalities from individual agents. Good protocols induce rich and generative ecologies. Bad protocols induce lifeless ecologies.

How can you tell them apart?

Protocol Affects

Just as humans might have a “game face” that is a function of specific games they may be playing, technologies too have game faces. We can call these protocol affects. To tell good and bad protocols apart, you have to read their affects.

The personalities of AI ecologies are currently emerging in inchoate, wild forms. Scaffolding elements like MCP and OpenClaw allow for relatively unbridled relational behavior among the various compute and human elements they weave together. But already there are signs of this Hobbesian wilderness being tamed. Protocols that are deliberately designed to shape the personality distribution of entire ecologies of intelligent agents in particular ways, and present them in stable ways, are rapidly emerging.

With humans, we use the term affect to point to how an underlying personality is expressed through deportment and comportment in a particular milieu. Protocol affects are the technological equivalent: Emergent typical behavior patterns of elemental high-personality technologies, when they are composed into “civilized” technological ecologies.

A good example of a protocol affect is the famously verbose and redundant one of TCP/IP, as revealed through jokes shared by networking engineers.

Hello, would you like to hear a TCP joke?
  Yes, I'd like to hear a TCP joke.
  OK, I'll tell you a TCP joke.
  OK, I'll hear a TCP joke.
  Are you ready to hear a TCP joke?
  Yes, I am ready to hear a TCP joke.
  OK, I'm about to send the TCP joke. It will last 10 seconds, it has two characters, it does not have a setting, it ends with a punchline.
  OK, I'm ready to hear the TCP joke that will last 10 seconds, has two characters, does not have a setting and will end with a punchline.
  I'm sorry, your connection has timed out... Hello, would you like to hear a TCP joke?

This “personality” expressed by TCP/IP (which replaced the Hobbesian anarchy of early network protocols) is not arbitrary. It is the result of a network consciously designed for high fault-tolerance under extreme circumstances, including nuclear war, which must continuously trade-off packet delay and packet loss.

Since it is a backend infrastructure technology, this is not a personality that lay users very often see (though they do experience the generativity it induces). But with other technologies, protocol affect can be part of broader human culture. AI, obviously, is one of these technologies.

What sorts of protocol affects might emerge from the various protocol ecologies taking shape today?

Zombiefied Discovery and Distribution

Applying the principle what you do is who you are, we can shed useful light on the nature and disposition of agent ecologies, as they continue to evolve past their wild phase, and develop stable protocol affects that human culture can take root in.

Computers at various scales of aggregation do different things. At the protocol level embodied by protocols like MCP, the main functions are discovery and distribution.

In the older stratum of the internet now entering its sunset phase, both were functions of what we call social media (at least as far as human users are concerned). The protocol affect accompanying these functions was one of delight and serendipity in the early years, which morphed to one of anxiety and frenetic competition over attention allocation in the later years. Thanks to the economic backdrop of the ZIRP era of zero/low interest rates, both discovery and distribution were cheaply available at global scale to almost everybody, with predictable over-exploitation and erosion of trust all around – what Cory Doctorow has labeled enshittification. Humans increasingly began retreating from the open internet to more closed cozy spaces. And the cost of this retreat was the breakdown of discovery and distribution mechanisms that relied on a lot of humans being publicly active online.

The protocol affect of the social internet has unraveled in the last few years. In terms of our personality metaphor for technologies, there is, in a sense “nobody there” anymore. No ghost haunting the social internet. There are no true public social media, and no protocol personality cohering to replace the one that unraveled. What remains is a pre-personality space of endless, mindless culture warring (what I called “the internet of beefs” elsewhere).

The internet still works mechanically, at the packet level, but as a global public social infrastructure with a defined and intelligible personality, marked by particular predictable planet-scale discovery and distribution dispositions, it has become zombified, even as our experience of it has become enshittified – the haunting experience of the public internet, its HX, is increasingly an empty and dispiriting one. There’s no there there anymore.

As a result, in the current era, discovery and distribution have become increasingly difficult and expensive for all activities that require internet-scale provisioning of those affordances. The problem is bad enough for existing needs, such as discovery and distribution of webpages and tweet-like messages. It gets exponentially worse when you consider the needs of new technologies.

Traditional discovery and distribution mechanisms are failing for traditional internet technologies such as social media and streaming video. They are complete non-starters for newer technologies.

Two in particular, are worth thinking about together, as a pair of evil twins: blockchains and AI. Curiously, the answer to the discovery and distribution problem might lie in a term shared by both, with different, but rhyming meanings – token.

The Packet and the Token

The legacy internet traffics in generic packets with some discrimination based on content type, and a presumption of bandwidth abundance. Discovery and distribution ultimately boil down to discovery and distribution of packets. The economy of the internet is, ultimately, the economy of packets. The still-unsettled back-and-forth political pendulum swinging around net neutrality is a debate about the political economy of packets, and whether it should be stewarded like a relatively abundant public commons or a corporatized market (dominated by a few large entities) that allocates a relatively scarce resource.

For emerging computational technologies, a new political economy has emerged on top of the packet economy. This is the token economy.

On blockchains, tokens mediate all interactions that require certain cryptographically secured assurances, in flexible and programmable ways, creating an economy that is something like a non-neutral internet, but one that can approach perfect competition more closely. Instead of large tech companies paying for private bandwidth, or non-net-neutral jurisdictions discriminating coarsely based on packet type (video vs. text for example), capacity can be sliced and diced in arbitrarily fine-grained ways, based on economic decision-making that can happen at bot-speed. Unlike what we might call packetspace, blockspace (and its more esoteric descendant, blobspace) is intrinsically structured as a market that prices interactions in tiny fractions of dollars, and transactional time constants measured in the milliseconds. Blockchain economies begin where the fastest and most fine-grained corners of the traditional economies, such as high-frequency trading, end. For some, this is just metastasized financialization and scams. For others, it is the beginning of economic outer space travel.

For AIs too, tokens are units of production and transaction. We generate text, code, images, and video using computers that measure their work, and charge for it, by the token (to be precise, tokens/second/user). Again, the picture looks like a non-net-neutral internet. How many tokens you get, of what quality, and at what speed, depends on what you’re willing to pay. And as with blockchains, this economy approaches perfect competition more closely. Instead of large organizations paying human programmers, writers, or artists by the hour or by the month, a vast market of individuals and small organizations can pay for code, text, and images by the token. As with blockchains, these tokens slice and dice what we might call inference space in fine-grained ways, with time constants measured in the milliseconds.

Does the term token represent a mere cosmetic connection between two frontiers of computing, or might there be a deeper conceptual link?

I suspect there is a conceptual link here. On both frontiers, tokens organize a natural economy around real scarcity that can ultimately be reduced to energy units (watts powering computers). More importantly, both kinds of token are informationally expressive in a way that packets, as mere “containers” are not.

And most importantly, the two kinds of token are, to borrow a term from electrical engineering, impedance matched. They have similar temporalities, spatialities, and information densities. They can be woven together, to form the warp and woof of a fundamentally different kind of internet. By itself, each is limited. As Matt Webb observed last year, modern AI by itself offers intelligence “too cheap to meter” which makes it more trouble than it is worth to scaffold for economic activity in a sufficiently fine-grained way, at least using conventional economic mechanisms. Blockchains, on the other hand are, among other things, metering technologies that shine precisely in too cheap to meter regimes. The two can, in other words, mesh in a fine-grained way. If you want to allocate work between two AI agents at a token-level of resolution, blockchains can do the job.

This is not idle speculation. One emerging mechanism for distribution and discovery (ERC 8004), combines AI and blockchain tokens in precisely this sense, and has already catalyzed the emergence of an ecology of AI agents that combine metered intelligence and small crypto transactions to form a marketplace. In the next Obliquities column, I will explore specific case studies.

Whether or not this particular approach succeeds, I suspect the foundation of the future internet will be an economy of tokens. Symbolic tokens that carry meanings and associations, and transactional tokens that carry valuations and risks, intricately orchestrated by a scaffolding that generates a tangled bank of private and public information and computation.

More broadly though, to return to the original motivating question, how does this emerging vision help solve the missing mechanisms problem?

Articulating Agent Ecologies

To summarize the idea I’ve been laying out here, the solution to the missing mechanisms problem is high-personality agent ecologies composed of individual agents with their own personalities. These personalities, far from being cosmetic features, are what allow functional behaviors to cohere at all levels, by allowing agents to be intelligible and predictable enough to each other to transact fruitfully, and produce increasingly complex and large scale effects. For us humans, inhabiting such computational ecologies will feel like being surrounded by friendly milieus of ghosts haunting our digital environments.

As a side effect, such ecologies would solve the so-called alignment problem, to the extent that is a well-posed and meaningful problem at all. High personality ecologies create alignment as they go, and wither and die when they fail to do so.

If you find this kind of future hard to imagine, take a peek at the short AI-generated movie we made at our workshop a year ago, South Beast Asia, which imagines (a Southeast Asian inspired) technological future full of AI-haunted digital and physical environments. Read our collection of short stories from our contest last year, Ghosts in Machines. We’re already creating this future.

What sort of physical reality might underlie such a planetary digital-physical hyperobject?

One mental model that I’ve found very useful derives from Peter Thiel’s observation that AI is “communist” while blockchains are “libertarian” in their personalities.

To a first approximation, modern AI tends to be most powerful when aggregated into really large-scale models running in the densest physical aggregations of compute (hence the excitement over gigawatt-scale datacenters). This feature naturally lends them a centripetal, convergent, homogenizing tendency and a “communist” personality.

Blockchains, on the other hand, are really only valuable to the extent they deliver on properties like censorship resistance, global consensus, capacity for irrevocable commitments (what Josh Stark named “hardness”), client diversity, and unbreakable (including quantum-resistant) cryptography. These features naturally lend blockchains a centrifugal, divergent, pluralist tendency, and a “libertarian” personality.

The respective token economies reflect these characteristics. Tokens in the sense of AI are essentially a “communist” currency, local to a particular model’s command economy. Tokens in the sense of blockchains only have value at all to the extent they are not local (“private blockchains” are deservedly mocked). Each by itself is impoverished and incapable of forming a high-personality agent ecology. Together, they can.

The interface between the two economies, I suspect, will feature phenomenology similar to the impossible trilemma in macroeconomics, or the boundary between the interiors and exteriors of firms in a Coasean economics sense.

Understood as a planet-scale computer, how do the two parts relate? AI will clearly be the “brain” of this planet-scale computer, similar to the CPUs, GPUs, or TPUs of individual computers. Whether this takes the form of dozens of gigawatt-scale datacenters running the largest models, and provisioning metered intelligence to the planet, or a more scale-free distribution of AI processing capabilities all the way to billions of intelligent entities on the network edge, is an open question. Whatever your political preferences for one or the other, there are also technological questions still being investigated. Is maximal aggregation necessary for performance? Can a gigawatt dispersed across a planet-wide decentralized network of small AIs be as capable as a single datacenter? Does embodiment matter? Does better local context beat cheaper tokens/second/user?

These are questions for which we will discover answers over the next few years.

The role that is likely to be played by blockchains (or functionally equivalent protocol technologies) will be that of the fabric. In modern computing, at all scales, the term fabric is usually used to describe the scaffolding that connects the different bits and pieces of the brain. There are fabric-like elements at the level of chips, servers, racks, and datacenters. The internet itself serves as the fabric at larger scales. The overall planetary computational fabric is a mix of smart and dumb elements. Fabrics embody the boundary intelligence of a system.

Blockchains are fabric technologies that can scale from personal computer scale to planet scale. They induce fabrics that operate by a different grammar than the familiar one we have today, but it is a grammar that is friendlier to agentic AI.

The fabric and the brain – an architecture for the emerging future of the internet that can sustain sufficiently high-personality ecologies to allow our frontier technologies to fully express themselves and truly thrive.

This is a very recent vision for the future of the internet (and indeed, the planet). As recently as five years ago, it was meaningful to describe Ethereum in terms of its original vision as a “world computer.” At the time, it was the only entity that merited such a description, since it allowed small-scale, highly constrained Turing-equivalent computing (the EVM, or Ethereum Virtual Machine) to run on a public blockchain. That was as good as planet-scale computation got, since traditional compute is, in a sense, stranded compute trapped within industrial-age organizational boundaries. There was no meaningful way to plug that compute into a planetary fabric, with or without blockchains.

AI brainpower though, is atomized into token-sized units (embodied by memory more than processing as we have come to appreciate), and capable of flowing smoothly across contexts. A fabric that can shape those flows, while preserving privacy with cryptographic guarantees, can create a kind of planetary intelligence that was impossible to even imagine just a few years ago.

One updated vision for the future of Ethereum in particular is as a world fabric rather than a world computer. It is, of course, not the only candidate auditioning for the role.

Whatever form the protocols constituting the fabric of planetary intelligence take, we will soon be living inside a planetary brain-and-fabric computer.

What will we do with this computer? That’s the question.

In the weeks since, thanks to the explosive adoption of coding agents like Claude Code, we’ve been inundated by evidence for this view of the future of AI. Amateur hobbyists are now vibe-coding entire complex digital production infrastructures involving dozens of agents swarming in parallel across a single computer’s filesystem, within complex organizational scaffoldings. We could think of these as agent factories.

Agent factories enable a great deal of complex higher-order action. Much of the attention has been drawn to moltbook (“Reddit for agents”), and the many entertaining trainwrecks involving OpenClaw (“claw” seems to have emerged as a term of art for an agent living dangerously and autonomously on its own server on the public internet, often armed with crypto wallets – what could go wrong?). But the truly interesting developments are largely invisible – individuals with significant mutual trust interacting with each other through their personal and bespoke Claude Code infrastructures, exchanging work-in-progress materials.

We could call these interaction patterns have your factory call my factory, and the underlying relationship pattern F2F (a rather fun overload of face-to-face). An exuberant F2F ecology is likely to be a central feature of the protocolized future.

My own personal experience with Claude Code illustrates the pattern well.

Subscribe now

Case Study: Indie Book Publishing Pipeline

I started my first week of using Claude Code by producing an online book of my Twitter archive, but I ended it by setting up an entire book manuscript production factory. Currently, my factory dashboard shows a couple of dozen book projects in flight, most derived from two decades worth of my personal blog and newsletter archives (including new editions of old books), and a handful of from-scratch projects.

Factory is really the only word for what I’m doing. In my case, a factory resembling a flexible job shop of the sort that makes varied things using a flexibly configured set of machine tools. My book projects are individual enough that each needs some bespoke handling, but similar enough that many processes and code modules can be reused. So a job shop is both an appropriate metaphor and a useful reference pattern. Other patterns would be appropriate for other production activities – flow shops, cell-based factories, assembly lines.

I found myself relying on dim memories of decades-old industrial engineering and operations research coursework to set things up. The factory floor is a portion of my laptop filesystem within my Dropbox folder, where various Claude Code sessions operate within a folder hierarchy and each folder has its own claude.md file. Each folder with a claude.md is a bit like a workstation or cell. Thought needs to go into defining boundaries, hand-off artifacts, and so on.

But setting up a Claude factory wasn’t the most interesting thing I did. It was setting up a logistics link between my factory and another Claude factory, set up by my long-time publishing co-conspirator , who has helped me personally publish two books in the past, and also handled much of the publishing work for Summer of Protocols/Protocolized, including the complex Protocol Kit and four books.

Jenna also happens to be an enthusiastic early adopter, and has set up her own factory to produce finished books from manuscripts. My factory takes messy raw materials and produces rough first-draft manuscripts. Her factory will take those manuscripts and produce finished artifacts that can be uploaded to Amazon for distribution as print and ebook volumes.

The handoff point between us is a shared Dropbox folder plus a “manuscript transmittal” server she’s set up for metadata. Here’s my “account” view of her factory:

And here is the manuscript transmittal page:

The fascinating thing? This very corporate-seeming pipeline was set up by two people who basically don’t code!

What we do bring to the party though, is domain expertise.

Domain Knowledge > Coding Knowledge

Jenna is a publishing industry veteran who knows exactly how to set up and run book production. I’m an experienced blogger and self-publisher with a dozen self-published books to my credit. We both know what we’re doing on our respective ends of this pipeline. Claude Code brings highly skilled coding ability to the party, but Jenna and I bring the (rather artisanal in this case) domain-specific knowledge required to decide what to do and how. Tasks that call for opinionated and tasteful decision-making rather than raw intelligence or procedural skills. We do both need some intelligence to make this work, but that’s not the main act. It’s a sideshow, provisioned in commoditized form by Anthropic.

Our F2F link is live. We’re currently discussing fussy details that are involved in producing a print version of my Twitter book. I sent her a docx file produced by my factory that’s the starting point for her factory, and she turned it around with revised requirements, which I implemented and returned to her. I had to tell my factory to redo the initial docx to address some global styling issues before Jenna’s factory can begin designing the book. I’m figuring out how best to automate the pipeline.

Both of us are using a good deal of custom code written by Claude Code, along with open standards like docx. We’re currently using Vellum (book design software), but we’re exploring replacing it with a bespoke design tool.

So far I haven’t touched a line of content text, and haven’t even looked at any code. I watch the action entirely at the shell level, like a factory floor supervisor. Python, json, and html fly around, while I chew on my cigar in my top hat.

This is not an isolated example. Elsewhere, with collaborators on a hobbyist robotics project, I’m helping prototype a discovery and marketplace infrastructure using the Ethereum 8004 discovery protocol for AI agents, and the 402 payments protocol.

And in the broader Claude ecosystem, the primary article of commerce is the skill, a fragment of agentic intelligence that perfectly fits the definition of social kernel. A kind of industrial intermediate, albeit for a cottage industry of individual-scale agent factories.

I’m sure there are plenty of more complex examples under development.

What are we to make of this type of F2F relationship? The principals (“legal persons”) involved in such interactions are individual humans, but the connections between them are a universe apart from the simple “friend” and “follower” type digital relationships we’re used to. Interactions are vastly more complex than social objects in digital envelopes that track likes and shares.

The only precedent I can think of is B2B relationships between factory-like entities.

I strongly suspect that this is the invisible 90% of the iceberg in the agentic AI revolution. While the public theatrics on moltbook and the claw ecosystem are much more visible, the sheer depth of capability integrations enabled by factory-to-factory connections between individuals argues in favor of high-trust relationships being the locus of the real action. Especially considering the zeitgeist vibe shift, in human social media, from more public spaces to Dark Forest/cozyweb spaces.

In intelligence media, there’s a lot more you can do within trusted friendships than within parasocial relations. Low-trust relationships are in fact rapidly hemorrhaging social energy.

The Factory-Owner Economy

One of the biggest concerns going around right now is the future of jobs, or more generally, the future of work. The conversation is a familiar one. Pessimists issue apocalyptic warnings of impending economic collapse. Optimists rehearse sunny arguments about the lump of labor fallacy, Jevon’s paradox, and Milton Friedman’s notion of “new wants and needs” emerging to fill the economic vacuums created by the disappearance of old ones.

Whether optimistic or pessimistic, our discourses seem unable to think about the future outside of existing categories – jobs, SaaS companies, outsourced white-collar labor, knowledge-work professions, mortgages. Several viral essays in recent weeks have (rather cynically and aggressively) doubled down on prognostication based on such bankrupt ontologies, to feed both wishful dreams and lurid fears, instead of taking on the harder work of coming up with useful new categories to think with.

The “factory owner” economy offers at least one new category to think with. It suggests, for instance, that in the future, rewarding and fulfilling work will be organized neither as “jobs” nor “gigs” but at least partly as an economy of bespoke F2F artisanal capitalism. The main factors of production are intelligence-on-tap that is too cheap to meter at the detail level, computers, and internet connections.

The F2F ecology won’t be the whole future of course (one of our doctrinal beliefs here at Protocolized is “your imagined future isn’t the only one unfolding while everything else stays unchanged”), but it will be one big force shaping it.

Is this an optimistic or pessimistic future? That is the wrong question. The right question is: Is it an interesting future; one that allows us to continue playing the game of civilization?

We here at Protocolized believe the answer is yes. And a big part of our mission this year is to put some serious thinking behind that answer.

Behind the scenes, a new stratum of public and private infrastructure protocols, starting with RSS and the Facebook newsfeed, powered the shift. This was accompanied by a shift in hardware – from the desktop and laptop to the phone as the primary device for accessing the internet, and with the camera and microphone replacing the keyboard as the primary input mechanisms.

In 2026, we at Protocolized are betting that a similar transition will begin in AI, from destination AI to intelligence media. A landscape shaped by “visits” to oracular destination AIs will be reshaped around intelligence circulating in intelligence media. Here we mean “intelligence” in the sense of a kind of content (similar to what “intelligence agencies” produce and transmit) rather than a kind of processing capability.

We have opinions on how this shift ought to play out. We would prefer it to play out in decentralized, capture-resistant ways, rather than through aggregation dynamics powering feed-like experiences.

Intelligence media need not themselves be particularly intelligent. Cutting and pasting an LLM chat link into a messenger, committing AI-generated code to GitHub, or downloading a set of weights all count as intelligence media operations. Last weekend, an important new class of intelligence media emerged with moltbook.com: social networks for AI bots.

What is important about all these emerging examples is that intermediate artifacts of AI processing move from one locus to another, in a permissioned, socially mediated way, jumping contexts in the process.

Intelligence media are media through which intelligence flows from one locus to another, primarily in disaggregated forms that get further metabolized as they flow, via interaction with shifting contexts corresponding to distinct loci.

We’ve already witnessed a shift from “prompt engineering” to “context engineering,” and we are about to discover that the most powerful way to (re)engineer context is to simply move work-in-progress to a new context. That is what intelligence media do. They achieve context engineering through context switching.

When Alice shares a ChatGPT link with Bob, who opens it and continues the chat, Bob’s fork of the chat can now draw on Bob’s memory context, which need not be shared with Alice (OpenAI of course, remains a third party in the background whom you must trust).

Currently, we’re improvising with the limited intelligence media we already have (chat link cut-and-paste probably accounts for 80%), but dedicated intelligence media, adapted to the needs of moving live intelligence rather than information, are beginning to emerge. Claude Code, for instance, moves coding-assistant intelligence to a directory in your local filesystem. Moltbook moves that local assistant intelligence to a space where that context comprises other assistants.

Will we see a rise in intelligent intelligence media, which might do some sort of processing as intelligence moves through pipes from one locus to another?

Precedents from other domains suggest the answer is no. One precedent is the “dumb pipes” vs. “smart pipes” debate in telecom a couple of decades ago, which has largely been settled in favor of dumb pipes. Another precedent domain is containerization, an “intelligence” transformation of global supply chains where the actual media were “dumb” containers. The intelligence lay in the fact that the contents were increasingly work-in-progress artifacts (which often crossed borders multiple times) rather than either raw materials or finished products. These examples suggest that intelligence will primarily be metabolized in step-function ways, at discrete locations, as it circulates. Not continuously in transit. So we might hazard a prediction that intelligence media will not be particularly intelligent. AI will suffuse the contents more than the containers.

Keep in mind though, that there might be invisible loci inserted between source and destination loci. We might see “context in the middle” attacks. Might browsers or operating systems on either end do things to links between Alice cutting-and-pasting and Bob clicking? Might ISPs sniff around at the behest of state and non-state actors? Ought we use Signal for passing chat links around? What are OpenAI’s servers doing when you generate a share link? “Prompt injection” as understood today is a primitive class of attacks compared to what will be possible once intelligence media begin to mature.

The shift to intelligence media will be marked by the rise of an AI analogue to social objects – what we might call social kernels. Unlike social objects (such as gifs, videos or podcasts), which are largely complete and ready for consumption when they enter social circulation (even if they trigger cascades of commentary, sampling, remixing, and meme-making), social kernels are primarily intermediate artifacts; snapshots of a process of progressive metabolism operating on information objects moving through a sequence of loci and coming into contact with different contexts.

Here is an initial definition:

Social kernels: Snapshots of evolving molecular human or centaur behaviors that shape each other at a low level, and contribute to low-level sociality norms, but do not necessarily catalyze sociality at the higher levels of complete “creators” or “content.”

We will develop the idea of social kernels more carefully in a later column, but a link to a partially complete LLM chat is a good prototypical example to keep in mind for now. It is not a complete artifact like a blog post but a few conversational turns on a theme that can be continued by Bob after it has been created and shared by Alice. Bob can then add a few more turns and share it again. The chat itself, or rather the particular moving instance of the original chat (an entity that repeatedly gets cloned, forked, and mutated as it gets passed along), is the social kernel.

The logic of this larger transition to intelligence media and social kernels, we believe, explains much of the frenetic action we’re seeing almost everywhere along the AI frontier, from the shift to so-called “agentic” AI, to the rise of non-chat UXs, to the sudden acceleration in robotics.

Last year, we at Protocolized paid particular attention to the emerging contours of distributed AI, and early protocols like MCP and A2A which aimed to provide scaffolding for it. It is now clear that the protocolization of AI, both to “distribute” it, and do other things with it, will be much messier and richer than the architects of MCP and A2A anticipated. Intelligence media will likely be a tangled-bank protocol ecology rather than just a handful of dominant standards.

One of the threads we will track this year in this column is how this protocolization is progressing. Make sure to stay subscribed to Protocolized to follow along.

Subscribe now