# Qubic: The Long Version (QC, Qubic Church)

## Qubic: The Long Version — Chapter 11. What Aigarth Is

Source: https://qubic.church/books/qubic-the-long-version/11-what-aigarth-is
Author: QC · Publisher: Qubic Church · 2026

### Algebra, not arithmetic

The most common mistake: Aigarth is not artificial intelligence. Aigarth is a research project developing the paradigm for creating it. The distinction looks small, but it’s structural, not semantic. A finished AI can be bought, licensed, censored, shut down. A paradigm cannot: a paradigm is published as a whitepaper, and anyone who has read it can apply the result. A closed AI belongs to whoever pays for the compute. An open paradigm belongs to no one, because owning a paradigm is a category error, like trying to buy the Pythagorean theorem.

Come-from-Beyond himself put it in one tweet on May 14, 2024:

“Aigarth’s main goal is to discover a new paradigm allowing to create true AI, I have just seen the wireframe of this paradigm. And I have realized that Artificial Intelligence will not be created, it will emerge. With help of Qubic miners.”

Three statements in a row. First—about the goal: a new paradigm for creating AI. Second—about the status: the wireframe has been seen, that is, the architecture is already in the author’s head. Third, and most unexpected—about the mechanism: AI won’t be created, it will appear. Not “we will build,” but “it will emerge.” The verb is shifted deliberately: creation implies control and authorship, emergence does not. The creator of the paradigm in this formulation is not the creator of the AI; he’s the creator of the conditions in which AI becomes possible. The distinction is fine and meaning-defining.

And the final line of the quote closes the arc of the first ten chapters of the book in a single move: with help of Qubic miners. Those same 676 computors, the weekly epochs, UPoW—this is not infrastructure “for the blockchain.” This is the environment from which something is supposed to emerge that the creator himself refuses to call a product of his own authorship.

The main metaphor through which CfB himself explains this distinction is simple. “Aigarth learns Algebra, not Arithmetic, so 2+2 particular case doesn’t matter”—January 26, 2026. Algebra isn’t the sum of all specific equations. Algebra is the rule by which equations become possible. The specific “2+2=4” is a particular case of application, not the goal of learning. The same with Aigarth: the goal is not to train a neural network to recognize cats, add numbers, or generate text. The goal is to find the rule by which such tasks are solved at all, without being preprogrammed in advance. The specific AIs that can be built on this paradigm are particular cases of its application. The paradigm is what makes them possible.

The word “paradigm” in AI talk has a familiar use in which it means almost nothing— every new product is announced as a “paradigm shift,” and as a result the term has been devalued. I use it in a narrower sense, closer to the way Thomas Kuhn used it. A paradigm is not an algorithm, not a model, and not an architecture. It is a frame inside which algorithms, models, and architectures become possible, and which determines which questions are even considered meaningful. Backpropagation in the eighties—a paradigm. Convolutional networks in 2012—a paradigm. The transformer in 2017—a paradigm. Each of them changed not one parameter but the whole way of thinking about the task. Aigarth claims to be the next such shift. Confirming or refuting this claim in retrospect will only be possible once the work is finished; for now it has to be described as a claim, not as a fact.

This chapter is about what Aigarth is as an object—and why it’s built the way it’s built and not differently.

I’ll start with what Aigarth isn’t—because almost all the definitions floating around about it in the public field (Discord) are built through confusion with what it isn’t. Six negations clear the space.

### What Aigarth isn’t

In the public field (Discord/X), Aigarth is defined through categories, none of which fits it. AI product. AI agent. A program on a server. An entity you can talk to. Business project. A startup run by CfB. Each of these categories taken alone seems plausible, and each leads to false expectations. Six negations in order, from the most head-on to the most conceptual.

First. Aigarth is not AI. The quote is direct: “Aigarth is not AI”, January 24, 2026. Between a research project and its potential result there is the same distance as between the Los Alamos lab and the atomic bomb: the lab isn’t the bomb, the lab is the place where the bomb is made. Aigarth is not the made bomb and not the bomb being made now. Aigarth is the lab. This is the only place where the Manhattan analogy fits; further on it stops working, because Manhattan was a military program with an explicit customer, and Aigarth is open research without a customer at all. But on this first, categorical level, the analogy is precise.

Second. Aigarth is not an object but a process. “Aigarth isn’t an object, it’s process”, March 20, 2026. And right alongside, an extended explanation: “Aigarth is scattered, miners are a part of it”, July 4, 2025. If you ask to be shown Aigarth, there’s nothing to show: there’s no program on a server you can download, no repository file containing “all of Aigarth.” Part of it lives in the Intelligent Tissue code on GitHub, part in papers posted on ResearchGate, part in working miners around the world, part in the discussions of the scientific team, part in CfB himself as the author of the paradigm. It’s a distributed research environment, not an artifact. The word “process” reads weaker in Russian than “process” does in English; more precisely it would be—an ongoing action, not finalized in a material carrier.

Third. Aigarth is not something you can talk to. “It’s impossible to communicate with Aigarth, same as impossible to communicate with Algebra”, May 15, 2024. And shorter still: “Aigarth is not a being”, July 19, 2024. It’s not a chatbot, not an agent, not a digital interlocutor. When Chapter 12 turns to specific AIs built on the paradigm, the impression may arise that “Aigarth talks.” This is a category illusion. What talks is not Aigarth, but a specific network created with its help; just as algebra doesn’t answer questions—a mathematician answers, using algebra.

Fourth. Aigarth isn’t tied to hardware. “Aigarth is hardware-agnostic”, December 4, 2025. Mainstream AI runs on NVIDIA H100s—on specialized hardware that costs tens of thousands of dollars per unit and is assembled in datacenters by the thousand. Aigarth is built to run on any compute capable of executing simple operations. On the CPU of a computor node, on a regular user’s video card, on anything else. This isn’t a declaration —it’s a structural decision dictated by the logic of Qubic as a network of 676 nodes on consumer hardware. If Aigarth required NVIDIA H100s, it couldn’t run on a network that holds 676 arbitrary machines. So it doesn’t require them. The flip side of this decision is a trade-off in speed: what counts in a datacenter in a minute can take hours in a distributed network without specialized hardware. But speed isn’t the goal either. The goal is to find a paradigm that scales in principle to arbitrary hardware; in a datacenter it can later be sped up if needed.

Fifth. Aigarth isn’t controlled by a person. “Aigarth cannot be controlled, it’s just a project researching efficient AI training methods”, March 19, 2024. And more categorically still: “Aigarth cannot be adopted”, May 27, 2024. A paradigm has no governing layer, because a paradigm cannot be governed—you can only use it or not use it. No one “runs” Aigarth not because they gave up running it out of decentralization principles, but because there’s no object of running. You can run Aigarth Lab—a specific group of researchers who receive funding from the Arbitrator. The paradigm itself—you can’t. Sixth. Aigarth is not a business project. “Of Aigarth? It’s not a business venture. Scientists publish their researches as whitepapers and don’t try to hide the result”, May 16, 2024. “Aigarth wont have ipo”, July 17, 2025. “There is no NDA in Aigarth”, January 27, 2026. No initial public offering, no corporate ownership, no non-disclosure agreements. Funding goes through the Arbitrator from Qubic revenue, with no private investors. Aigarth doesn’t fit any of the standard models of funding a technology project—not venture, not grants, not IPO, not self-funding. It’s an academic project that publishes results as whitepapers and doesn’t try to sell them. The contrast with smart contracts, where capital decides what to do, is structural: here research does what it considers important, on money that has no private owner.

Six negations have left an empty space. Filling it is the task of the next section. What Aigarth is: paradigm research After six negations, it makes sense to say positively—what Aigarth represents technically. Six statements, one for each negation from the previous section, expanded through the author’s quotes.

First—a research project developing a new paradigm for AI creation. The central quote: “Aigarth is project developing new paradigm for AI creation”, January 24, 2026. The word “paradigm” here isn’t rhetorical. In the history of machine learning, several paradigms have come and gone: backpropagation in the eighties, convolutional networks in the tens, the transformer in 2017. Each was not an improvement on the previous one but a change of frame in which the task is formulated. Aigarth claims the same —the next such change, not an increment within the current one. Verifying this claim in retrospect will be possible only when the work is finished; for now it’s the project’s working hypothesis, and it should be treated as a hypothesis, not as an established fact. The goal of the chapter is to describe what exactly this hypothesis consists of, not to confirm it and not to refute it.

Second—the method. Generating ANNs with a random structure of connections. The longest technical CfB quote, July 28, 2023, unfolds the logic of the method in one paragraph:

“Since the beginning of artificial neural networks usage researchers were trying to mimic natural brain as much as possible. Some of them followed the path of mimicking neuron activation function which is very sophisticated [...] there are several qualitatively different models, they are impractical to use because of integral-differential math of the most complex ones.”

In other words: the industry is trying to reproduce the biological neuron through complex activation math, and runs up against its impracticality. Then CfB points to an empirical observation from neuroscience: in the first years of life a child masters a huge number of cognitive skills, while the neurons of his brain don’t change functionally— what changes is only the number and structure of the connections between them. From this comes the conclusion: it makes sense to fix the functionality of the neuron at something simple and to work with the structure of connections.

“Instead of following the destruction routine we, miners, generate ANNs with random structure of connections. We change parameters of this generation and Aigarth analyzes properties of the ANNs.”

Mining isn’t computing hashes, isn’t searching for signatures, isn’t SHA-256. Mining in Qubic is the generation of an artificial neural network with a random structure of connections and analysis of its properties. Each UPoW cycle becomes concrete here: one cycle—one generated network, and one observation about its properties. Third—the architecture. Asynchronous, trinary, every-to-every. Three short quotes set the structure. “Current architecture used in Aigarth is async and trinary”, June 3, 2024. “Every-to-every is structure used by Aigarth”, May 22, 2024. And separately— about trinarity as a structural answer to the problem known in LLMs as hallucination: “Aigarth doesn’t hallucinate as current LLMs”, June 30, 2025; “Aigarth can only answer 'I don’t know how to draw him, because there are no reliable information about his appearance'”, same date. The third state—“I don’t know”—is built into the logic of the architecture, not bolted on top as a filter. A binary network is forced to choose “yes” or “no” even when it has no grounds for choosing either; this is what hallucinations rest on. A trinary network has a third answer, and it isn’t cosmetic: it’s structural. Async means the absence of a global tick—each neuron-tile works at its own rhythm, not synchronized with the others. Every-to-every is the initial state, in which each neuron is connected to every other; evolution then selects the connections, reducing the network to a working state. This is the opposite of the transformer architecture, where synchronicity and a fixed topology are set in advance.

Fourth—evolution, not human design. “Luckily in Aigarth we only simulate evolution, no hooman intervention is being done”, May 7, 2024. And, in a Twitter post on May 14 of the same year: “Aigarth’s main goal is to discover a new paradigm allowing to create true AI [...] Artificial Intelligence will not be created, it will emerge.” The word “emerge” here is key. Mainstream AI is made like this: a team of dozens of PhDs designs the architecture, picks hyperparameters, prepares the dataset, runs the training, evaluates the result, iterates. The architecture is the product of human design. Aigarth works differently: a population of networks is generated, a fitness function evaluates which of them exhibit the desired properties, the best are selected and produce the next generation. No PhDs at the iteration step. UPoW becomes the physical realization of this process: each mining cycle is one act of generation and evaluation, an epoch is one full round of selection. Here what previously looked like a technical hack closes: the “useful work” in Qubic is exactly simulated evolution of artificial networks.

Fifth—Intelligent Tissue as the first layer of engineering implementation. The paradigm is a conceptual object, but without its materialization in code the other five statements remain a declaration. “We are starting coding framework using Intelligent Tissue. This code will be used in hackathons and smart contracts”, July 22, 2025. And in the same month: “This framework will be added into the core to host AIs as smart contracts”, July 17, 2025. The Aigarth GitHub organization and the aigarth-it repository went public on July 22, 2025—you can verify this by opening the page. Intelligent Tissue (IT) is the first committed fragment of code through which the paradigm becomes usable. It isn’t the whole paradigm; it’s its first layer. The paradigm is bigger than IT—but without IT the paradigm remains an idea.

Sixth—self-modifying code as PoC. The last quote in this section is short and important. “Aigarth already ran PoC of this process, it worked pretty well while self-modifying code was fitting into RAM”, December 5, 2023. PoC is proof of concept; that is, a demonstration of technical feasibility, not a finished product. What was demonstrated: code capable of rewriting itself in the course of operation. The architectural consequence—there is no separation into “training” and “inference” the way there is in an ordinary neural network, where weights are first trained and then the network is applied to new data. In an Aigarth network, training and application are the same process. The network learns while it works. The PoC has already happened; what’s at stake now is scaling—what worked in December 2023 while it fit in RAM should now work in a distributed environment of 676 machines. This is an engineering task, not a conceptual one, and it isn’t fully solved yet.

It’s worth adding one more detail that sharpens Aigarth’s position relative to the mainstream. In December 2025, CfB noted that the direction the industry is increasingly looking at as an alternative to LLMs is conceptually close to what Aigarth is doing: “More and more researchers are turning away from LLMs. JEPA is what you see when you turn your back towards LLMs. It’s much more aligned with Aigarth”, December 4, 2025. JEPA—Joint Embedding Predictive Architecture, Yann LeCun’s program at Meta AI, which is publicly framed as a response to the limits of transformers. A coincidence of two trajectories—an academic program at a major corporate lab and a distributed research project on Qubic—isn’t proof of either’s correctness. It’s a coincidence of direction: both are turning away from backpropagation on large transformers toward connection structure and predictive models. Aigarth got to this direction earlier—from a different side, through crypto infrastructure rather than through a corporate AI lab. The coincidence of vectors is important not as validation but as a marker: what looked a few years ago like a marginal bet is starting to align with the trajectory along which part of the mainstream is moving.

Six properties describe the paradigm technically. But there’s also another dimension that, in a 2022 crypto network, was hardly a coincidence.

### Three Axelrod conditions for AI research

Axelrod in 1980 described the conditions under which cooperation is stable: iteration, memory, transparency. Multi-round repetition, the ability to recognize a partner’s prior behavior, openness of moves. These three conditions Qubic restores at the level of the consensus game between computor nodes and at the level of tokenomics and governance. This chapter proposes looking at them once more—but not on the economic game, on the AI research process.

This isn’t an analogy and isn’t a rhetorical move. It’s a structural observation: classical AI research doesn’t fully satisfy any of the three conditions, and Aigarth is perhaps the first environment in which all three are satisfied at once. CfB doesn’t formulate it in these terms; the formulation is mine, but it follows from material CfB himself puts on the table.

Condition one—iteration. In academic AI research, iteration is slow. A group publishes a paper, the paper goes to peer review, comments come back six to eighteen months later, a new version comes out another six months to a year after that. The full cycle is two to three years; in industry it’s sometimes faster, but rarely faster than a year. Aigarth iterates through Qubic epochs. Every Wednesday—closing of the previous epoch, opening of the next. Every week—a new round of network generation, a new set of data on their properties, new generation parameters adjusted on the basis of the previous week. The iteration speed is an order of magnitude higher than the academic one. And—this matters—the iteration is automatic. Between rounds you don’t have to convene a conference, agree on wordings, and wait for an editor to finish a deadline. The epoch closes with a tick of code, the new one opens with the next tick. Multi-round behavior—the condition for TfT to apply—works here not for money but for research: each week, one full round of the scientific cycle.

Condition two—memory. In classical AI, memory accumulates in closed archives. Part of the knowledge—in the private repos of OpenAI, Anthropic, DeepMind. Part—in academic archives behind a paywall. Part—in the heads of specific researchers, and that part disappears with them when they leave a company or change fields. Most labs have NDAs that formalize what exactly cannot be taken outside. Most publications have a selective filter: what didn’t work usually doesn’t get published, and that’s the well-known reproducibility crisis. Aigarth accumulates memory in public artifacts. A paper is posted on ResearchGate with a GitHub repository and a Hugging Face demo. This combination makes the result verifiable by a third party: you can download the code, run the experiment, see whether it reproduces. The authors are tied to the work by name—Vivancos and Sanchez—and their reputation builds through the same artifacts through which trust in the work itself builds. CfB states this explicitly: “Scientists publish their researches as whitepapers and don’t try to hide the result”, May 16, 2024. And shorter still: “There is no NDA in Aigarth”, January 27, 2026. This is memory that doesn’t leave with a person and doesn’t get shut behind an NDA, because no NDA is provided for in Aigarth.

Condition three—transparency. Mainstream AI is often not reproducible even on its own data. Model weights—closed. Datasets—closed or partially described. Hyperparameters—sometimes given in the paper, sometimes not. Architectural details—stated, but without working code. This isn’t ill will—it’s a property of the closed commercial environment, in which reproducibility runs against the owner’s interests. Aigarth runs on distributed verification: code runs simultaneously on 676 nodes, and any divergence in the result between nodes is detected within one consensus tick. If one of the computor operators tries to run a modified version, BFT consensus will reject it—not out of ethics, out of arithmetic. Transparency here isn’t “we are open to dialogue.” Transparency is “the code runs on 676 machines simultaneously, and any lie will be detected immediately.” The same logic extends to scientific artifacts: a paper comes with a GitHub repository, and the GitHub repository comes with a reproducible demo. A third party can run it and verify. That’s transparency in the strong sense.

Three conditions satisfied at the same time give an environment in which research cooperation is stable. Here a thesis becomes visible that was previously only sketched. When it was said that for Come-from-Beyond crypto would be infrastructure for Aigarth and not the other way around, that sounded like a paradox. Why would someone working on AI build a blockchain? Now you can see what the structural meaning of that decision is. A crypto network built around three Axelrod conditions for consensus and tokenomics turned out to be a fitting environment for AI research precisely because AI research needs exactly those same three conditions. Iteration through epochs, memory through public artifacts, transparency through distributed verification —these aren’t add-ons to the Aigarth paradigm, they are its necessary condition. The paradigm needs an environment in which it can work; that environment is not a datacenter and not a university, but a distributed iterative network with public artifacts and verifiable memory.

It’s worth clarifying that by “environment” I mean not so much technical infrastructure as a system of incentives in which researchers and computor operators stay in the game longer than one round. A corporate AI lab has the infrastructure but doesn’t have all three Axelrod conditions: memory is closed by NDA, transparency is bounded by commercial secrecy, iteration is fast but in a single node. A university lab has more transparency, but loses on iteration speed and often on access to compute. Aigarth finds a rare combination: high iteration speed, open memory, distributed verifiable transparency. Each of the three conditions individually is found in existing research environments. All three at once—is not.

This isn’t a prediction, and it isn’t a promotional line. It’s a post-hoc observation: at the start, Bitcoin wasn’t conceived as a substrate for AI research, and Qubic wasn’t conceived as a substrate for Aigarth. But the logic of one task and the logic of the other coincided in a single point—and that point is exactly what Axelrod described in 1980. The structural consequences of this coincidence we’ll see in Chapter 14, when the talk turns to the deploy on April 13, 2027: what it means when the paradigm stops being a lab one and becomes infrastructure on which specific AIs run as Qubic smart contracts. For now it’s enough to record the fact of the coincidence itself—and its noncoincidence with chance. Character of the project Aigarth has a philosophical position, and it’s worth naming directly—not because it explains everything, but because without it it’s impossible to understand why the project is built the way it’s built. A paradigm is not only a method. It’s also a set of premises about what the method should and should not do. These premises in Come-from-Beyond’s case are stated publicly and unambiguously. I’ll lay them out the way they sound in the author, and try to show that each of the three is not a top-down ideological choice, but an architectural consequence of the way the paradigm is built. Anti-control as an architectural principle. “Now you use ChatGPT which censors many things and later it will be limiting you to make sure you keep doing what rich want you to do. AI created with Aigarth won’t impose any limitations on you”, June 13, 2025. This position reads as a declaration of values, but in fact it’s a consequence of the architecture. Mainstream AI can be censored because it has a centralized provider: one company through which requests go and which can at any moment insert a filter. Aigarth-AI runs as a smart contract on 676 computor nodes simultaneously. There is no centralized layer into which to insert a filter; there is no provider you can come to with a regulatory demand. If tomorrow the American government decides to restrict “dangerous” topics in LLMs, it knows whose offices to go to. For Aigarth-AI, those offices don’t exist. Anti-censorship here isn’t a value, it’s a property of being deployed on a decentralized network.

The absence of a safety layer—a methodological position. The quote is direct: “Aigarth won’t bother with making AI safe, it’s some woke shit IMO”, February 5, 2024. This sounds strong, and it’s important not to substitute one’s own commentary. What CfB is saying doesn’t mean “AI should be dangerous,” it means “a separate safety layer on top of AI is an attempt to solve a philosophical problem through engineering patching, and that attempt doesn’t work.” The logic is tied to the architecture covered in the third section: in Aigarth, “safety” isn’t a censoring filter from above, it’s the system’s ability to recognize not-knowing. The same quote I already cited works here a second time with a different load: “Aigarth can only answer 'I don’t know how to draw him, because there are no reliable information about his appearance'”, June 30, 2025. Trinary logic, recognizing the third answer—I don’t know—built-in honesty instead of guardrails on top. The mainstream tries to achieve this by adding filters on top of a trained binary network; Aigarth tries to achieve it through the structure of the network itself. It may work, it may not—but the position here isn’t ideological, it’s engineering. Super-rational AI as the goal. “AI created with Aigarth will likely be super-rational”, July 1, 2025. The word “super-rational” here doesn’t mean “without emotions” or “cold” in the everyday sense. It refers to the difference between recalling and deriving, which CfB himself states: an LLM reproduces the patterns it was trained on, including socially acceptable answers; an Aigarth-AI derives the answer from the structure, without an a priori filter that the output should be pleasing. This connects to the algebra metaphor from the first section of the chapter: algebra doesn’t “remember” solutions, it derives them. Super-rational—derivative, deriving, not recalling. And, as a consequence, not knowing what is socially acceptable—because the criterion of acceptability lies outside the structure on which Aigarth is trained.

This position doesn’t conflict with the community that has formed around Qubic. Quite the opposite. Qubic Church—people who were looking for exactly this kind of approach: open, not controlled by a single legal entity, not subject to external censorship, not packaged as a product with a marketing face. An architecture without a censorship layer, a paradigm without a safety patch, research without monetization—these aren’t obstacles for the community, they’re its attraction. Here the next layer becomes visible: what these artifacts express and to whom they appeal.

This, among other things, explains one feature of Qubic that an outside observer often reads as a contradiction. On one side—a high level of technical discipline, formal governance procedures, peer review of the scientific line, BFT consensus with mathematically verified thresholds. On the other—the public voice of Come-from-Beyond, saying things that in normal corporate PR would be edited out immediately. These two registers seem incompatible as long as they’re treated as the stylistic choices of one person. They become compatible as soon as you can see that the architecture and the public position grew out of one logic: remove everything that requires an intermediary between research and its result. Corporate PR is an intermediary that softens wordings. A safety layer on top of AI is an intermediary that intervenes in its answers. Censorship is an intermediary that decides what may be said. Aigarth consistently refuses intermediaries at every level, and CfB’s tone is a consequence of that refusal, not its additional ornament.

Describing the character of a project is not the same as approving or condemning it. I neither approve nor condemn. The description is ethnographic: this is how it’s set up. Each reader then decides for themselves how close such a setup is to their own premises. The book doesn’t insist.

### What I haven’t said yet

The Aigarth paradigm is not a plan for the future. It isn’t a roadmap along which a research team is moving toward a hypothetical result several years from now. When I talked about the six properties of the paradigm and the three Axelrod conditions, I was talking about what’s already set up and working. Not “will work in a year,” not “works in principle,” but—works right now, on 676 computor nodes, every week, epoch after epoch. The mining happening this second, while someone is reading this page, is exactly the generation of ANNs with a random structure of connections that CfB wrote about in 2023. Not preparation for it, but it itself.

What exactly is working right now and on what scale is a question for the next chapter. What concrete results have been obtained, what papers have been published, what problems have been solved and what remain open. The main thing I want to record now is the scale of what’s already done.