# Aigarth: 24 Years to Today. By TesserAI (@usetesserai) · 2026-06-30 · Original: https://x.com/usetesserai/status/2071667859722506494 Source page: https://qubic.church/articles/aigarth-24-years-to-today ## What happened today: Neuraxon 2.0, the scientific core of the Aigarth paradigm, entered the ARC-AGI-3 benchmark. It's the hardest public test of general intelligence that exists, made by François Chollet, the guy who wrote Keras and basically invented the whole idea of measuring generalization in machines. First attempt: 0.13%. Sounds like nothing? Here's the context that changes everything. On this test, regular humans score 100%. And the best AIs on the planet, GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6, Grok 4.2, they all score below one percent. Grok, Elon Musk's xAI flagship, scored exactly zero on the official run. And Neuraxon? Not zero. No LLM. No GPU. Just bare, ordinary processors. A tiny trinary network, run by a handful of enthusiasts on home PCs, came out above Musk's AI on a test that was literally built to break large language models. And here's the part to keep in your head until the end: this happened BEFORE the first full launch of Aigarth. Neuraxon isn't even the final form. It's a component for Aigarth, a proof of the principle. The main engine, Aigarth on the full Qubic network, hasn't been switched on at full power yet. What you're looking at is the warm-up. To get why this is staggering, you have to walk the whole path. From the start. ## Act I. The Quiet Opening. The real story started on 13 April 2022, when Qubic's real mining went live. The creator, Come-from-Beyond (CfB), one of the people behind NXT and IOTA, set the goal right away: > Qubic launch is an important step toward Aigarth, a project aiming to compete against projects like [OpenAI]. > — CfB, 13 April 2022 Back then, saying you're going to compete with OpenAI sounded like delusion. And yet CfB already saw the problem the whole industry was missing: > What we see now is just a variant of cargo cult, AI generates phrases and often they sound reasonable but we cannot rely on anything it says. > — CfB, 7 January 2023 Read that again. January 2023, peak ChatGPT hype, and he's calling it a cargo cult. Imitation of the form without the substance. And he'd already found the direction: > We have got a breakthrough recently, related to concept of DIKW pyramid. It may be a fundamental change of point of view on AI development. > — CfB, 21 March 2023 ## Act II. The Build. In 2024 the claims started turning into findings. The verbs changed, from "looking for" to "found." > No, we found neuron which behaves like a transistor. > — CfB, 18 April 2024 While the whole industry was busy simulating the biological neuron ever more expensively and precisely, Aigarth went another way: the neuron as a simple switch, cheap to compute. That's the foundation of the energy efficiency that comes later. And then the move that splits the whole story in half: > Now we see that the 4th of Dec, 2024 will be the mid-point of our road to Aigarth. After that date we are starting working on practically usable AIs. > — CfB, 23 July 2024 An official midpoint, announced ahead of time. From that point on the project stopped generating networks and started training them. This is where Anna is born, a tiny network learning to add numbers. And in the same year, CfB dropped maybe the deepest definition of intelligence in this whole story: > The moment of intelligence emergence can probably be spotted when interconnections implement a function which doesn't have the inverse function. > — CfB, 18 August 2024 ## Act III. "We Know How" (2025) 2025 was the year the confidence went absolute. First a critical technical wall came down: local extremum and overfitting, both beaten at the same time (July 2025). Two walls almost every AI approach crashes into. And then CfB said the line worth engraving: > We know how to create AI, but we created none yet. > — CfB, 23 July 2025 Read it like an engineer, not a skeptic. "Know how" means the solution is found. "Created none yet" means execution is what's left. The hard part, seeing the path, is done. And he stopped hinting at LLMs and started going at them point-blank: > Real AI isn't one which passes the Turing test, but one which fails it intentionally. > — CfB, 2 September 2025 > In the Aigarth project we know what/where intelligence is and why LLMs actually have nothing to do with one. > — CfB, 30 September 2025 In 2025 this sounded cocky. A few months later Chollet's benchmark would put it in black and white: on a real generalization test, every LLM sits at zero. CfB said it half a year before the world saw the proof. ## Act IV. The Last Wall, Energy (2026) By 2026 CfB considered the question "can AI be built" closed. One barrier left. And it's not intelligence, it's energy. > Aigarth already found how to get AI, now trying to find out how to get AI without 1000 nuclear plants. > — CfB, 19 February 2026 This is the real stake of the whole game. Mainstream AI works, but its appetite for energy trends toward physically absurd numbers. Aigarth's question: can you build intelligence inside the energy budget the planet can actually provide? And the answer started coming in: > It means that instead of 2000 nuclear plants we'll need 14. > — CfB, 2 May 2026 (To be fair, CfB clarified these are illustrative numbers, not a precise estimate. But the direction is the point. The horizon isn't receding. It's shrinking: 1000 to 14.) > We solved AI and now need practical proof which is just matter of getting enough power. > — CfB, 21 May 2026 "Solved" means the task is reduced to engineering, not that a working AGI is already breathing. The line is drawn on purpose. But here's the thing: the project reached the point where all that's left is to prove in practice what's already solved in principle. ## Act V. And Now, Today. This whole path. 2022's "compete with OpenAI." 2023's "LLMs are a cargo cult." 2024's "found the transistor-neuron." 2025's "we know how, and LLMs have nothing to do with it." 2026's "solved, power is what's left." All of it was the creator's words. And on 29 June 2026 the external, independent, public confirmation showed up, from people with no stake in the token and no reason to flatter. Chollet's benchmark. An open leaderboard. And on it: All the world's large language models, near zero. Musk's Grok 4.2, exactly zero. Neuraxon on ordinary CPUs, no LLM, no GPU, above zero. Above Grok. CfB said "LLMs are a cargo cult" in 2023. In 2025, "LLMs have nothing to do with intelligence." Chollet's benchmark in 2026 showed exactly that: on a test of real intelligence, the ability to grasp a new rule from a couple of examples the way any kid does, giants running on billions of dollars of GPUs score zero, and a tiny trinary network scores non-zero. This is the moment worth stopping for. Not because 0.13% is a big number (it's tiny, and there's a long road ahead, that's honest). But because it's the first external proof of direction. A compass needle you can't fake with marketing. And now, the final point: ## All of This, BEFORE Launch The thing that came out above Grok on Chollet's test is Neuraxon. A component. A proof of the principle, on bare processors, without the full power of the network. Aigarth, the main engine, hasn't been launched at full power yet. The full evolutionary machine on the entire Qubic network, the sixth-largest compute cluster on the planet, dedicated to training AI, that's still ahead. The merging of Neuraxon and Aigarth, ahead. The practical proof the community is driving toward, grinding Anna down to zero by hand, that's happening right now. What you saw today is the warm-up before the first real launch. It's a race car that passed its rivals on the formation lap, before the race even started. If a component on home processors is already above Musk's AI on the hardest test on the planet, what happens when the engine is switched on in full? ## Where We Actually Are 0.13% is the start of the road, not a win. On this test, anything below a few percent is still the early-signal zone. "Above Grok" is a fact, but Neuraxon hasn't passed the three other top models yet. "Solved AI," in CfB's words, means "reduced to engineering," not "AGI born." And the real, full verification of all this has a specific date on it. Not "someday," but 13 April 2027, exactly five years from the first mining. ## Final: 24 years ago this was an idea in one man's head. In 2023, "the cargo cult is them, not us." In 2025, "we know how." In 2026, "we solved it, power is what's left." And today a tiny network on ordinary processors, no LLMs and no GPUs, came out above Elon Musk's AI on a test built specifically to break AI. And the engine isn't even switched on yet. Watch the date. 13 April 2027.