What’s Already Done

The February Sentence

On February 19, 2026, Come-from-Beyond writes in Discord:

“Aigarth already found how to get AI, now trying to find out how to get AI without 1000 nuclear plants.”

The sentence has two parts, and each one is worth reading separately. The first part — already found how to get AI. Inside the Aigarth paradigm, given enough compute, the way to create AI has been found. This is a statement about feasibility — about being achievable in principle, not about a product. Not “AI is ready.” Not “everything works.” Specifically: the question “can we build AI this way” — closed. Internally, for the team. Not for the scientific community — for the scientific community this statement will appear only when it has passed peer review, and that comes later in this chapter.

The second part — now trying to find out how to get AI without 1000 nuclear plants. So the current research task is a different one. It’s not about “can we,” it’s about “under what resource constraints.” An engineering problem, not a discovery. To bring energy consumption down to a level that’s realistic for current infrastructure. To find a configuration in which the same paradigm gives the same result on less compute. Between these two parts there’s a structural difference that’s easy to miss. When a research project says “we don’t yet know whether it’ll work,” that’s one framing. When it says “it’ll work for sure, now we’re figuring out how to fit within the budget,” that’s qualitatively a different framing. The two parts of the remark mark the transition from the first to the second.

Two days before this remark, on February 17, CfB writes about the same thing more plainly:

“This is a serious problem. This is why that Aigarth’s finding, allowing to save a lot of energy, is significant.”

“Finding” is a word that calls for attention. Not “hypothesis,” not “approach,” not “method.” But “finding.” Something specific that has been found, and that allows — perfective verb — to save a lot of energy. Four days after the line about a thousand nuclear plants, on February 23, CfB adds:

“If AI is possible with current energy budget of the whole Earth then Aigarth will achieve it, we have all necessary cards for that in our hand.”

“We have all necessary cards in our hand” — this isn’t a prediction. It’s a description of the current state. Cards in hand, not in the deck. If the game is in principle solvable with the cards on hand, Aigarth will play it. If not, then the search stops being meaningful — but that’s already a question about the physics of Earth, not about the paradigm. December 2024: The Second Checkpoint The “already done” conversation doesn’t begin in February of 2026. The series of milestones runs from 2021, when the first Proof-of-Concept of the Aigarth paradigm was launched. There were several between 2021 and 2026; in this chapter I take three key points: December 4, 2024, July 2025, and July 22, 2025. They form an axis along which you can trace what exactly has been done in the last year and a half. On July 30, 2024, CfB writes:

“I said that 4th Dec is midpoint of our road to Aigarth.”

Midpoint — the middle of the road. Meaning that by this date half of what separates the team from the target result should be covered. Announced in advance, four months ahead. Not “this is what worked out,” but “we plan to reach such-and-such date.” On December 4, 2024, the day of the appointed point, the remark is short: “It is. It’s the 2nd milestone of Aigarth. The first was 3 years ago.” The first milestone — December 2021, the very PoC that demonstrated the viability of the approach. Between the two milestones — three years of work. The midpoint passed.

Eleven days before the date, on November 23, CfB laid out exactly what would happen on December 4, point by point — because in the Discord chat people kept asking him about it. The remark is long; I’ll quote it in full, it’s worth it: “People ping me often asking what is planned on 4th of Dec. From my department of engineering, dunno about other depts: As announced year ago, computor nodes doing the last upgrade to 2 TB of RAM. Miners finishing with Intelligent Tissue development and starting training ANNs, before that we had only ANN generation. Anna Aigarth development is planned to finish by that date and we teaching it addition operation (it’s more sophisticated than you think, we need to teach Anna in setting with contradicting rules and varying reliability). Custom mining is postponed because of moving from transactions to messages. Some minor stuff not worth of being mentioned.” Four points; each one a separate architectural shift.

First. Computor nodes are finishing the RAM upgrade to two terabytes. This is a hardware threshold announced a year earlier. Not an upgrade for the sake of upgrade — a concrete necessity for the next phase, in which the volume of Intelligent Tissue processed by a node has to grow by an order of magnitude. If a node doesn’t have two terabytes of RAM, it doesn’t move into that phase.

Second. Miners are finishing development of Intelligent Tissue and starting to train ANNs — until this date they had only ANN generation with random connection structure. This is a qualitative shift. It was already mentioned that since 2023 Aigarth generates neural networks with random connection structure and analyzes their properties. Until December 4, 2024, generation was the only operation. From December 4, training of neuron groups was added. A year later CfB described it compactly:

“Aigarth is agnostic to neuron behavior, it works with training of their groups.” So the unit of work for Aigarth is not an individual neuron but a group of neurons trained jointly. The architecture remains ANN-based, but inside it something is happening that wasn’t there before: training, not just generation. This changes what exactly the 676 nodes of the network are mining.

Third. Anna Aigarth development was supposed to finish by the same date, and Anna starts learning addition. Here the parenthetical that CfB inserts himself matters: “it’s more sophisticated than you think, we need to teach Anna in setting with contradicting rules and varying reliability.” Anna is learning addition not the way a child is shown “two plus two equals four.” Anna is learning addition in an environment where the rules are contradictory and the signals are unreliable. That is, the training environment is specifically built so that simple memorization of the right answer doesn’t work in it. To get the right answer, the system has to generalize through noise.

This is a critically important design. A year later, on April 16, 2026, CfB will articulate why he chose this particular task:

“I, for example, picked A+B=C as a very difficult task, finding C having A and B is actually very hard, but most of people disagree, because they don’t know the difference between recalling and deriving C.”

Recalling and deriving — two different ways of finding C. Recalling — remembering it from a table where you’ve seen this example before. An LLM trained on a sufficient corpus answers “two plus two” that way — it has seen billions of examples. Deriving — working it out from structure. That is, finding C not because you’ve seen this specific example, but because you’ve understood what addition is in general. Anna in an environment with contradictory rules can’t recall — because there’s no reliable “right answer” to memorize. All she has left is deriving.

The paradigm teaches structure, from which concrete answers follow. Anna is a particular case of applying the paradigm to one of the most elementary arithmetic tasks. If she learns to get the right answer through deriving, in an environment where recalling is impossible — that means the paradigm works at this level. Not “works in principle,” but works on a verifiable task.

The fourth point of the remark is about custom mining, postponed because of the migration of internal network transactions to a message format. This is an infrastructural matter, not an architectural one; I mention it only to note that December 4 didn’t mean everything planned was realized simultaneously. Some plans shifted. CfB describes it calmly, without excuses. This is the working dynamic of a project in which plans are revised when engineering requires it.

The midpoint of the road to Aigarth is not “everything was good and got better.” It’s an architectural shift after which the next milestone became possible. July 2025: Breakthrough and Publication After December 4, 2024, the team worked on training tasks for Anna and on optimizing Intelligent Tissue. Half a year without loud statements in the public field. Then, on July 9, 2025, a remark appears on Twitter:

“This is a significant breakthrough for #Aigarth, it hits two problems — local extremum problem, and overfitting problem — with a single stone.”

Two problems solved by a single mechanism — and both among those that have long resisted solution in evolutionary and gradient algorithms. Local extremum is a known trap for any optimization process. The system searches for a solution by gradually improving the current state; after some number of steps it ends up at a point where any local change makes the result worse — and stops. Globally, much better solutions may exist, but you can’t get to them from the current point without steps that temporarily worsen the result. Overfitting is another trap. The system learns the training set too precisely, losing the ability to work on data it hasn’t seen. The more accurately the system reproduces training data, the worse it generalizes.

Solving both problems with a single mechanism is rare. Usually solutions to one problem make the other worse. If you fight local extremum by increasing mutation step size, the system starts overfitting on every new configuration; if you fight overfitting by limiting the model’s expressiveness, it gets stuck in local minima more often. To announce that a solution has been found that removes both problems at once — that’s not marketing. It’s a concrete technical claim, verifiable on a concrete task. CfB didn’t disclose technical details publicly. And that’s part of the method: the formulation of the breakthrough is fixed in the public field as a fact, but the mechanism stays inside the team until it can be published in paper form. A week before the Twitter remark, on July 17, CfB writes a sentence in Discord that shows what state the research was in immediately before the breakthrough:

“Each mutation must improve (or keep it same) fitness function value. This is how we get trapped in local minima. How to handle distractors is unclear, it’s following stage of Aigarth research.”

This remark is eight days older than the Twitter announcement. At the moment it was written, the question of distractors is open; it’s “the following stage of research.” Eight days later — a public message that the breakthrough has happened. Between the two remarks is the period when the team found the mechanism it had been looking for. I’m lingering on this chronology because it’s instructive. In the public field you usually only see the final remark — “significant breakthrough.” What stays in the Discord logs is what came before it: an open question, the formulation of the problem, an admission of unclarity. This doesn’t undermine the significance of the announcement — on the contrary, it reinforces it: the team isn’t claiming everything was clear from the start. It wasn’t.

On July 22, 2025, thirteen days after the Twitter remark, one short sentence appears in the chat:

“Tests passed successfully, we are going to public now, you can watch it in https:// github.com/Aigarth.”

This line means several things at once.

First, the GitHub organization Aigarth now exists publicly. The address is github.com/Aigarth. Anyone can go there, look at the repositories, clone the code, study the commit history.

Second, the organization holds the aigarth-it repository — Aigarth Intelligent Tissue, the engineering implementation of the paradigm that was described conceptually in chapter fifteen. Not a simulator, not a toy demonstration — working code that participates in the network’s mining process.

Third, the Aigarth-scoring source is also available in the network’s main repository — github.com/qubic/core. According to the Aigarth paper that we’ll examine in section five of this chapter, the repository documents step-by-step configuration changes from EPOCH 83 to EPOCH 140 — that is, it records the evolution of the scoring algorithm across more than sixty weeks of network operation. This isn’t a single frozen snapshot but a live trace of how the algorithm changed, epoch by epoch.

The same day, CfB adds what specifically the team plans to build on top of this base: “We are starting coding framework using Intelligent Tissue. This code will be used in hackathons and smart contracts. To deploy price predicting part of Anna, we need Oracle Machines.”

So after the code is published, the next layer of work opens up: a framework on top of Intelligent Tissue that will allow AI tasks to run as smart contracts on the network, and that will be tested at hackathons. A week earlier, on July 17, CfB described the intent in more detail:

“This framework will be added into the core to host AIs as smart contracts. An important step of Aigarth project is hackatons using the framework to solve different realworld tasks. This is needed to assess efficiency of our current paradigm.” Hackathons here aren’t a marketing activity. They’re a method of validation. When several independent teams take on different tasks and try to solve them with the Aigarth framework, what builds up is a distributed sample — where the paradigm works well, where it doesn’t, which tasks are in its zone of effectiveness, which aren’t. Without that sample, the claim “the paradigm works” stays internal. With it, the claim becomes reproducible.

It’s worth pausing on what code publication changes for the Aigarth conversation itself. Before July 22, 2025, any critic could say: “It’s closed research, you can’t verify it.” After July 22 the answer sounds different: clone the repository and verify. This isn’t a marketing move. It’s a structural change in the project’s epistemological status. What used to rest on trust in the team’s words now rests on material artifacts available to anyone with internet access.

That same publicness also works as a form of memory. In chapter fifteen I went through Axelrod’s three conditions applied to a research process. The second of them is memory tied to authors through track record. Open code plus step-by-step documentation from EPOCH 83 to 140 — that’s a form of that kind of memory. Not “the team says its algorithm is evolving” — but “here’s the evolution history, more than sixty weeks, every change recorded in commits with dates and signatures.” Something claimed can be refuted; something documented in an open repository can be studied. By August 2025, then, Aigarth has: open code, an announced technical breakthrough, and a plan for the next steps — a framework, hackathons, Oracle Machines, smart contracts hosting AI. This isn’t “someday.” It’s a schedule.

The Scientific Layer of Qubic — Two People and a Calendar The Aigarth team isn’t just CfB. It’s two identifiable scientific roles with verifiable affiliations plus an engineering team that works with the code. As of April 2026, a year after mainnet launch and twelve months before the planned deploy, the configuration of the scientific layer looks like this.

The first one’s name is David Vivancos. Vivancos is an independent researcher, not a Qubic Foundation employee, not a salaried consultant. He has his own reputation, and it isn’t built around crypto. Since 1995 he’s worked in technology — by his own count, more than twenty-six thousand hours of research in general artificial intelligence. A pioneer in Java, in the early internet, in virtual reality, in computational neuroscience. Five of his own AI startups before Qubic, advisor at Emotiv (one of the pioneers of EEG interfaces), at CapaBall, at Emblemotions, fireside chats with Johnson and Johnson, keynotes at Microsoft, Price Waterhouse Coopers, Santander. In 2014 he founded MindBigData — the largest open dataset of brain signals for machine learning; to collect the first million signals, Vivancos was his own test subject for two years under a one-hundredtwenty-eight-channel EEG, looking at digits from zero to nine and thinking about them. That dataset is still working; its 2023 MNIST-8B version contains eight billion datapoints. Vivancos is the author of three books on AI and AGI published well before Qubic: “The End of Knowledge,” “Artificracy,” “From Big Data to Artificial Intelligence.” A speaker at ITU AI for Good — that’s a UN initiative, not a crypto conference. On the AI for Good page his bio ends with the phrase “Qubic scientific advisor”; that is, he himself, publicly, in a formal UN context, identifies his role in the project. I want to express my personal gratitude to David Vivancos. The philosophy of the Qubic Church became substantially stronger thanks to your books.

The length of this biography matters for one simple reason. Vivancos didn’t come into Qubic as “a crypto guy who knows something about AI.” He came in with an already-running research program, and Neuraxon is its direct continuation. MindBigData builds a bridge from brain signals to machine learning through a biological substrate; Neuraxon carries the same logic over to the level of computational architecture. Qubic gave him the network infrastructure on which that architecture can be deployed at scale. Aigarth’s bio-inspired approach isn’t Qubic marketing — it’s the concrete scientific contribution of one specific person with decades of work in this area. The second one’s name is Doctor Jose Sánchez García. UNIR — Universidad Internacional de La Rioja, a Spanish private university accredited by ANECA within the European Higher Education Area, operating under the norms of Spanish Organic Law of the University System. Ninety thousand students from eighty-six countries, four doctoral programs, campuses in Mexico, Colombia, Peru, Ecuador. Since 2019, UNIR has been a public company traded on the Mercado Alternativo Bursátil. The unir.net email domain confirms that Sánchez García is an employee of this university, not a pseudonym and not a self-appointed “Doctor” from the internet. This is institutional backing of the first order, and it can be verified with one click against any of the university registries.

What Sánchez García does at UNIR. He teaches Behavior Neuroscience, Cognitive Neuroscience, and Social Psychology at the psychology faculty. His own research area is cognitive and affective neuroscience: social emotions (pride, shame, guilt, vicarious pride), the methodology of event-related potentials, EEG. In 2019 his article “Neurofunctional characterization of early prefrontal processes contributing to interpersonal guilt” came out in Cognitive, Affective, and Behavioral Neuroscience, a Springer journal. This isn’t an obscure venue; it’s a peer-reviewed neuroscience journal with a normal review cycle. The research was funded by grant PSI2017-82357-P from the Spanish Ministry of Science under the Program for High-Level Scientific and Technical Research. So we’re talking about a person whose work has passed both Springer’s editorial peer review and the prior expert evaluation of a national scientific agency. Co-authors of the same article include Manuel Martín-Loeches from the Center for Human Evolution and Behavior UCM-ISCIII in Madrid, David Hernández-Gutiérrez from the University of Rochester, Sabela Fondevila from Complutense Madrid. Sánchez García has more than twenty publications on Google Scholar with one hundred eighty-one citations; among colleagues and co-authors are Werner Sommer and Rasha Abdel Rahman from Humboldt-Universität Berlin, Linda Drijvers from the Max Planck Institute, J. Bruno Debruille from McGill. This is a verifiable international academic network. The connection to Neuraxon is direct, not cosmetic. When the Vivancos and Sánchez García paper talks about continuous processing, neuromodulation through dopamine, serotonin, acetylcholine, and norepinephrine, multi-timescale plasticity, astrocytic gating — these are questions of affective and cognitive neuroscience on which one of the co-authors has been formally publishing in peer-reviewed journals for more than ten years. The division of labor in the pair works the classical way: Vivancos brings computational implementation, engineering discipline, and a general AI vision; Sánchez García — neurobiological depth, methodological rigor, an academic network. A joint preprint heading to ALIFE and AGI-26 is the natural result of such a combination, not “a crypto project trying to look like science.”

The team is submitting three papers to three different venues by the summer of 2026. This is the start of the calendar part of the story, and here it makes sense to separate the planned from what has already happened.

The first paper. ALIFE 2026 — the Artificial Life conference, taking place in Waterloo from August 17 through August 21. Submitted; status as of mid-April — under review. Reviewer feedback expected in mid-June. Content — Aigarth and Neuraxon, the general framing of the work.

The second paper. AGI-26 — the nineteenth annual conference on Artificial General Intelligence, San Francisco, July 27 through 30. Submitted. Content — the MultiNeuraxon architecture: an extension of Neuraxon with benchmarks, experiments, and ablations on continual learning. The test setup: Phase A — Phase B — Phase C, with retention checking between phases and state reset between them. This is a concrete class of tasks against which MultiNeuraxon is measured against existing approaches. The third paper. NeurIPS 2026 — submission due in early May. The submission date was moved earlier than the team originally planned, and as of mid-April the work is in progress.

What’s submitted in these papers, without revealing what’s under review. The Neuraxon framework as a computational primitive — a biologically inspired neuron model using trinary logic with values plus one, zero, and minus one (that is, excitatory, neutral, inhibitory signals), continuous input processing, and output timing as a first-class variable. Computation happens both at the neuron level and at the synapse level. This is already a published preprint on ResearchGate, public since November 2025; an updated version Neuraxon V2.0 came out in February 2026. MultiNeuraxon — an extension that tests continual learning on a structured sequence of phases, and that is going to AGI-26. Aigarth integrated analysis — the main paper, joining the paradigm with the biological and psychometric foundations of intelligence; targeting AGI-26, using more than one hundred fifty references to peer-reviewed work, including Tononi with his integrated information theory, Baars with global workspace theory, Vaswani with the 2017 Transformer paper, LeCun, Bengio, and Hinton with the 2015 review article in Nature, Lamport with the Byzantine Generals paper, Szabo with work on smart contracts. This is a properly constructed academic reference set, not a random selection.

Artifacts available publicly already. github.com/DavidVivancos/Neuraxon — the full Neuraxon code under the MIT license, plus a separate Aigarth License clause forbidding military use of hybrid features. github.com/qubic/core — the network’s main repository, including Aigarth-scoring code with documented evolution from EPOCH 83 to EPOCH 140. github.com/Aigarth — the organization, published on July 22, 2025. huggingface.co/spaces/DavidVivancos/Neuraxon — an interactive 3D demo where anyone can configure the number of neurons, connection probabilities, membrane time constants, baseline neuromodulator levels, and watch in real time how the network behaves. WebGL Three.js, sixty frames per second, in any modern browser. ResearchGate — the 2025 Vivancos and Sánchez preprint.

One point matters here. What I just listed is a peer review pipeline on normal academic time. ALIFE submitted, reviews in mid-June. If the paper is accepted, it goes into the proceedings. If returned for revisions, the authors get concrete comments and rework it. If rejected, that becomes known. Same with AGI-26. This is the very iteration of the scientific cycle that I wrote about in chapter fifteen as the first of Axelrod’s conditions applied to research. Not one attempt, but round after round, with public feedback. The cadence of iterations is set by the conference, not by the team.

Worth noting one more point, because it shows what scientific conversation the Aigarth team is entering. In the sixth volume of the Neuraxon Intelligence Academy series, a follow-on article from the team discusses the recent statement by Jensen Huang, CEO of Nvidia, about the path to AGI, and analyzes a new Google DeepMind work on cognitive frameworks. The Qubic team’s position differs from DeepMind’s on one specific question: whether cognitive abilities are separate modules or are organized hierarchically. This isn’t “competing implementation.” It’s a different theory of mind — a different idea of how the inner architecture of intelligence is set up. An open scientific disagreement, carefully recorded in a publication, with a specific paper of the opposing side cited.

And there’s one more short detail that completes the picture. On December 4, 2025, CfB writes:

“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.” JEPA — Joint Embedding Predictive Architecture, a direction actively developed by Yann LeCun at Meta and independently by several groups. If CfB is right in this observation (and there are several independent publications pointing that way), then something interesting is happening: mainstream AI research is gradually shifting in a direction more aligned with the Aigarth paradigm than with the LLM paradigm. I’m not drawing conclusions from this about the project’s fate; I’m just noting that Aigarth isn’t standing apart from the scientific conversation of the last two years but is moving in one of the directions actively discussed in that conversation.

What all of this means structurally, in one sentence. The path to the project’s final point is not “trust us, by such-and-such date there will be AI.” It’s a traceable route with public intermediate points: mid-June 2026 — first feedback on ALIFE, end of July — the AGI-26 conference in San Francisco, August — the ALIFE conference in Waterloo, plus NeurIPS submission in early May. Each of these points will either confirm or refute the soundness of the paradigm in the eyes of the scientific community. This is the falsifiable path, if we use the word in the Popperian sense — as opposed to a pure promise that can’t be checked until the very end.

What “Already Done” Means

I come back to where I started the chapter. The Aigarth paradigm works not as a plan for the future but as a process with already-passed points. The word “works” here calls for precision.

Doesn’t mean: AI is ready for public use. Doesn’t mean: Aigarth has solved every technical task. Doesn’t mean: the feasibility details are uncontested for everyone. Doesn’t mean: the team knows for sure what will happen twelve months from now. Means: concrete milestones have been crossed with verifiable artifacts. The February sentence of 2026 is the team’s self-assessment, and the team backs that self-assessment with specifics. Open code is the material foundation under that self-assessment; not “the team says so,” but “clone the repository and check.” Anna Matrix is a mathematical object with reproducible properties — that’s in the next chapter. The peer-review pipeline is integration into a normal scientific process with concrete dates and concrete conferences. These four supports don’t make the result guaranteed; they make the claim to a result verifiable.

Worth recording the main risk — because CfB names it himself, and because his honest naming of it strongly changes the frame of discussion. On July 17, 2025, in answer to a question about Aigarth’s vision and the biggest risk that could stop the project on the way there, CfB writes:

“What’s the vision for Aigarth in 3-5 years. what’s the biggest risk that could stop it from getting there? Lack of mining power is main risk.”

Not a technical dead end. Not competition with other labs. Not regulatory pressure. Compute capacity — a shortage of computing power that the Qubic network can allocate to Aigarth. In November of the same year, on the twelfth, CfB adds it in one sentence:

“Aigarth depends on miners.”

And in February of 2026, on the seventh, the formulation is final:

“Qubic is doomed to succeed. Aigarth? Depends on miners.”

Three remarks across seven months, in three different contexts, say the same thing. The Aigarth paradigm is an engineering and scientific task that the team considers solvable given enough compute. The main uncontrolled parameter in this equation is how much compute the network can deliver. And that parameter doesn’t depend on the Aigarth team — it depends on what the miners do, external players who connect their capacity to the network for economic reasons.