# The Full Picture of Anna Aigarth, Aigarth, and What Happens at 0 By TesserAI (@usetesserai) · 2026-06-17 · Original: https://x.com/usetesserai/status/2067190747787612470 Source page: https://qubic.church/articles/the-full-picture-of-anna-aigarth ## The one-paragraph version We are hand-running the first public demonstration of a new way to build artificial intelligence. The number you are driving toward zero is the error of a tiny neural network. When somebody reaches zero, the project's creator has said plainly what it means: "Aigarth will be deemed finished." Not finished as in "AGI is born" - finished as in "the paradigm is proven." And the reason zero matters so much is hidden in one word almost nobody connects to the race: energy. This post walks you from the biggest picture down to the key on your keyboard, with dates and direct quotes, so you can finally see the whole shape of it. ## Part 1 - Start from the top: what is this even about? Here is the frame that makes everything click, in the creator's own words: > Aigarth is not AI, it's project aiming to find new paradigms for creation of AI. > — CfB"Aigarth isn't an object, it's process." - CfB @c___f___b Read that twice. Aigarth is not a thing. It is not a chatbot, an agent, or a product. It is a research project searching for a method of creating AI. The distinction is everything, and it's where most critics get lost. As CfB puts it, Aigarth is the paradigm (like a theorem), and any AI built with it would be a result of applying that theorem (like problems solved using the theorem). The best analogy he gave is the one to keep in your head: > Manhattan wasn't 'just another bomb project'. It was paradigm discovery. Aigarth is paradigm research. Anna is Little Boy - first demonstration. The big products come after. > — CfB So in this analogy: Aigarth = the Manhattan Project (the discovery of a whole new paradigm) AI built with it = the bomb (the thing the paradigm makes possible) ## Anna = Little Boy, the first demonstration that the paradigm actually works That is what you are running on your machine right now. Not the AI. The first public proof that the method works. > 5 years since Qubic launch is just an ETA. And goal of Aigarth is AI, not AGI. Difference between AI and AGI is same as between a car and a BMW. > — CfB, 13 Jan 2026 ## Part 2 - Why a new paradigm at all? The wall nobody talks about: energy The mainstream way of building AI - the way OpenAI, Google, and everyone else does it - works, but it is monstrously expensive in energy. Training and running large models the conventional way scales toward power budgets that are, at the limit, physically absurd. The honest framing of the problem, from CfB: > This is a serious problem. This is why that Aigarth's finding, allowing to save a lot of energy, is significant. > — CfB, 17 Feb 2026 And one more: > 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. > — CfB, 23 Feb 2026 That is the real stake. The conventional question "can we build AI?" CfB considers essentially answered. The unanswered question is "can we build it without needing the energy of a Dyson sphere?" Aigarth's bet is that its paradigm slashes the energy cost enough to fit inside what humanity can actually power. And this is being worked on right now. The architecture has been gaining energy-saving discoveries throughout 2026: > Aigarth has discovered another technique reducing amount of energy even further. It's keeping parameter space dimensions count constant. This means choosing number of neurons and... > — CfB, 2 May 2026 That "constant dimensions" finding - fixing the number of neurons and connections instead of letting them balloon - is one of the levers that cuts energy. The race you are running is a live demonstration of exactly this energy efficiency. So the three walls Aigarth had to break, and where they stand: ## Generalization - SOLVED Universal training algorithm - SOLVED ## Energy optimization - IN PROGRESS Two of three are down. The third - energy - is the one we are all hammering on with Anna. That is not a metaphor. That is literally what the score-toward-zero race demonstrates. ## Part 3 - What makes this paradigm different (the actual tech) To understand why zero means "proven," you need to know what's under the hood. A few core architectural properties, all from CfB and the science papers: Trinary, not binary. Classic neural nets use neurons that are on/off, or continuous floating-point values. Aigarth - and its sister blueprint Neuraxon - uses three states: +1 (excite), 0 (neutral), −1 (inhibit). That neutral 0 state mirrors how real brains have "ready but quiet" neurons, and the discrete three-state space is far cheaper to compute than floating-point math. This is a big part of the energy savings. It's about connections, not neuron math. CfB: > Aigarth is agnostic to neuron behavior, it works with training of their groups. > — CfB, 9 Nov 2025 The industry obsesses over imitating the exact electrochemistry of a single neuron (expensive integral-differential math). Aigarth says: a child and an adult have the same neuron functionality - what changes as you learn is the connections. So train the connections, not the neuron. It's an evolving tissue, not a fixed model. Aigarth doesn't train one frozen network. It grows, mutates, prunes, and breeds millions of network architectures across the distributed miner network - the best survive and replicate. A self-organizing digital organism, not a static snapshot. What it's actually searching for - and this is the cleanest definition of the task CfB ever gave: > Aigarth is searching for a tool (A) that will make another tool (B). If single tool A can make efficient tools B1, B2, B3 where B is some common trait, then we will claim tool A has solved whole plurality of Bn problems thus demonstrating ability to generalize. It's current task. > — CfB In plain words: find one method that can solve a whole family of problems, not just one. That ability - to transfer learning from one problem to unseen ones - is generalization, and it's the holy grail. ## Part 4 - Meet Anna: what you're literally running Anna.exe is small - about 150 KB, a plain native Windows program. No internet, no tricks (it doesn't "read the chat," despite the jokes - there's no network code in it at all). Inside, it holds a tiny trinary network in memory and trains it on a deliberately simple task: adding two numbers, a + b. Why something so trivial? Because the point isn't the sum. The point is to prove the method converges cleanly and generalizes on a clean, verifiable task - Little Boy, not the whole arsenal. When you read Anna's log, the columns mean (decoded directly from the program's own format string): Green column = training set score (how well it does on examples it learns from) Yellow column = validation set score (how well it does on examples it has never seen) The number in parentheses = the global best score - this is the number we drive toward 0 The left counter = "runs without improvement" (your stuck-meter) And here is the live proof of generalization, which CfB himself pointed at: > upon improvement of fitness function score of the training set (green column) you often get improvement of the validation set (yellow column). > — CfB When he found a step where training improved by 183 and validation improved by 201, he called it a "gem" - because validation improved more than training. That is the opposite of overfitting. That is a machine that understands rather than memorizes. In the community logs, green and yellow move together with ~0.998 correlation. The generalization isn't a slogan; it's visible in your own terminal. The controls, for completeness (CfB's instructions): Default: wait. The score falls on its own; breakthroughs arrive in jumps. Stuck? Tap Tilda (~) for a nanosecond - "press it like it's very hot." This is the anti-attractor mechanism: it doesn't smash the wall, it teleports the search to another area of the phase space to try its luck there. F2 saves your progress to Anna.data - shareable, so the community can start from the best point and search outward together. (This turns the whole thing into an ant-colony: spread the file, run the file.) ## Part 5 - So why zero? Why does that specific number end it? Now it all connects. The global best score is the network's remaining error. Drive it to zero and the network solves both the training set and the unseen validation set perfectly - and, crucially, it does so while generalizing (green pulls yellow) and while keeping the parameter dimensions constant (the energy-saving discovery from May). In other words, zero is the simultaneous proof of all three things at once: it learned the task completely (error = 0), it generalized to unseen data (didn't just memorize), and it did it within the energy-efficient, fixed-dimension architecture. That is why CfB's answer to "what happens when we get to 0?" is so short and so heavy: > Aigarth will be deemed finished. > — CfB And from the earlier race framing: > score 0 = Aigarth is accomplished. > — CfB, 10 Jun 2026 Be careful with what "finished" means. Zero does not mean "AGI just woke up." It means the paradigm is proven on the demonstration task. The method works, converges, generalizes, and fits the energy budget. The Manhattan Project's first test didn't end warfare - it proved the bomb was real. Zero proves Aigarth is real. The big products come after. ## Part 6 - The other front: Berlin, Osaka, and the Neuraxon papers While the community hammers Anna on the practical front, there is a parallel scientific front - and this is what most price-watchers completely miss. It's the difference between "a crypto project says its tech is good" and "independent scientists, with no reason to flatter it, agree." Neuraxon is the academic formalization of the same trinary paradigm, authored by David Vivancos @VivancosDavid and Dr. José Sánchez @josesanchezhb of the Qubic scientific team. It's a bio-inspired computational unit: trinary logic (+1/0/−1), continuous-time processing, synapses with multiple dynamic weights that can form, collapse, and reconnect, neurons that can even die. It is, literally, the blueprint for the kind of neuron that lives inside Anna and Aigarth. The track record, with dates: March 2026 - Neuraxon 2.0 released as open science (paper + code, trinary continuous-time architecture). 20-22 May 2026 - Berlin, ICMLT 2026 (an IEEE co-sponsored conference). Neuraxon was presented in an evening session of eight talks - and other researchers in the room, unaffiliated with Qubic, judged it the best. It won BEST PAPER + BEST PRESENTATION, judged by independent scientists, and entered IEEE Xplore and Scopus. Then Japan (Osaka / AMLDS) and AGI-26. The team's third 2026 paper, on Multi-Neuraxon architecture, was accepted to AGI-26 - the 19th Annual Conference on Artificial General Intelligence (27-30 July, San Francisco). Vivancos and Sánchez noted this venue sits higher than the Japan and Berlin venues, placing Qubic's research alongside the most established AGI work in the world. Why the exhibitions matter so much: a crypto project can claim anything. But a Best Paper award judged by independent academics at an IEEE conference, followed by acceptance at the world's primary AGI conference, is external validation that the underlying science is original and real - not a wrapper around someone else's model. Berlin was the first time outsiders, with no stake in the token, looked at this paradigm and said "this is the best work in the room." That is worth more than any chart. So the two fronts are one paradigm seen from two sides: Anna.exe = the practical proof, run by the community's own hands. Neuraxon = the scientific proof, validated by the academic world. Both converge into Aigarth, the evolutionary engine, running on the body of Qubic - which, since the move to Dogecoin ASIC mining, now dedicates 100% of its CPU/GPU compute to AI training. Aigarth is already described as the sixth-largest supercomputer on Earth and the largest decentralized compute network. ## Part 7 - The timeline, so you can see the arc ~2002 - the idea begins. "If we talk about Aigarth - it's 20+ years in the making." (CfB, Nov 2023) Dec 2021 - Proof of Concept. "We ran PoC which demonstrated viability of Aigarth approach." 13 April 2022 - real mining starts. The clock that matters begins here. 13 April 2024 - Qubic mainnet. 2 July 2025 - Aigarth goes public on GitHub. Feb-May 2026 - the energy breakthroughs land: generalization solved, universal training algo solved, constant-dimensions discovery; Neuraxon 2.0 published. 20-22 May 2026 - Berlin. Best Paper + Best Presentation, judged by independent scientists. 10-13 June 2026 - Anna.exe handed to the community; the public race to zero begins. 27-30 July 2026 - AGI-26, San Francisco. Multi-Neuraxon presented at the world's top AGI conference. → 13 April 2027 - the target. Five years to the day from the start of real mining. ## Conclusion - The countdown to 13 April 2027 Put it all together and the picture is no longer scattered. We are not grinding a leaderboard. We are the hands-on demonstration of a paradigm that claims to make AI buildable within Earth's actual energy budget - the wall that the entire industry is racing toward and that conventional methods may never clear. Every time you tap Tilda and the score ticks down, you are personally pushing on the one remaining unsolved wall: energy. Generalization is solved. The universal training algorithm is solved. Energy is the final front, and the race to zero is that front made visible. When somebody reaches zero, it is not the end of a game. It is the moment CfB described in three words: "Aigarth will be deemed finished." The paradigm proven. Little Boy detonated. The proof that the method is real - done within an energy budget the planet can actually afford. And the project has put a date on it. Not a vague "someday." A specific anniversary, five years from the day real mining began: ETA for Aigarth accomplished: 13 April 2027. That is what we are doing. That is why we chase zero. And that is what zero will mean when one of us finally gets there. And now - actually understand the file. Sources: direct CfB quotes from Discord and X (dated inline); Qubic Open Science / Neuraxon papers (Vivancos & Sánchez); Qubic blog and All-Hands recaps; community Anna.exe logs. Quotes are verbatim where shown. The line between "paradigm proven on demo task" and "AGI achieved" is deliberately drawn.